Model Migration

For the latest, authoritative version (with code samples in every supported language), WebFetch the Migration Guide URL from shared/live-sources.md. Use this file for the consolidated, skill-resident reference; fall back to the live docs whenever a model launch or breaking change may have shifted the picture.

This file is large. Use the section names below to jump (or Grep this file for the heading text). Read Step 0 and Step 1 first — they apply to every migration. Then read only the per-target section for the model you are migrating to.

TL;DR: Change the model ID string. If you were using budget_tokens, switch to thinking: {type: "adaptive"}. If you were using assistant prefills, they 400 on both Opus 4.6 and Sonnet 4.6 — switch to one of the prefill replacements (most often output_config.format; see the table in Breaking Changes by Source Model). If you’re moving from Sonnet 4.5 to Sonnet 4.6, set effort explicitly — 4.6 defaults to high. Remove the effort-2025-11-24 and fine-grained-tool-streaming-2025-05-14 beta headers (GA on 4.6); remove interleaved-thinking-2025-05-14 once you’re on adaptive thinking (keep it only while using the transitional budget_tokens escape hatch). Then drop back from client.beta.messages.create to client.messages.create. Dial back any aggressive “CRITICAL: YOU MUST” tool instructions; 4.6 follows the system prompt much more closely.

Step 0: Confirm the migration scope

Before any Write, Edit, or MultiEdit call, confirm the scope. If the user’s request does not explicitly name a single file, a specific directory, or an explicit file list, ask first — do not start editing. This is non-negotiable: even imperative-sounding requests like “migrate my codebase”, “move my project to X”, “upgrade to Sonnet 4.6”, or bare “migrate to Opus 4.7” leave the scope ambiguous and require a clarifying question. Phrases like “my project”, “my code”, “my codebase”, “the whole thing”, “everywhere”, or “across the repo” are ambiguous, not directive — they tell you what to do but not where. Ask before doing.

Offer the common scopes explicitly and wait for the answer before touching any file:

1. The entire working directory
2. A specific subdirectory (e.g. src/, app/, services/billing/)
3. A specific file or a list of files

Surface this as a single clarifying question so the user can answer in one turn. Proceed without asking only when the scope is already unambiguous — the user named an exact file (“migrate extract.py to Sonnet 4.6”), pointed at a specific directory (“migrate everything under services/billing/ to Opus 4.6”), listed specific files (“update a.py and b.py“), or already answered the scope question in an earlier turn. If you can answer the question “which files is this change going to touch?” with a precise list from the prompt alone, proceed. If not, ask.

Worked example. If the user says “Move my project to Opus 4.6. I want adaptive thinking everywhere it makes sense.” you do not know whether “my project” means the whole working directory, just src/, just the production code, or something else — the everywhere makes the intent clear (update every call site within scope) but the scope itself is still not defined. Do not start editing. Respond with:

Before I start editing, can you confirm the scope? I can migrate:

1. Every .py file in the working directory
2. Just the files under src/ (production code)
3. A specific subdirectory or list of files you name

Which one?

Then wait for the answer. The same applies to “Migrate to Opus 4.7” and bare “Help me upgrade to Sonnet 4.6” — ask before editing.

Sizing the scope question (large repos). Before asking, get a per-directory count so the user can pick concretely:

Present the breakdown in your scope question (e.g. “Found 217 references across 3 directories: api/ (130), api-go/ (62), routing/ (25). Which to migrate?”). Also confirm git status is clean before surveying — unexpected modifications mean a concurrent process; stop and investigate before proceeding.

Step 1: Classify each file

Not every file that contains the old model ID is a caller of the API. Before editing, classify each file into one of these buckets — the right action differs:

Bucket 3 sub-cases — before swapping a string reference, check:

• Capability gate (e.g. if 'opus-4-6' in model_id: enables a feature) → add the new ID alongside, don’t replace. The old model is still served and still has the capability, so replacing would silently disable the feature for any old-model traffic that still flows through. If you know no old-model traffic will hit this gate (single-caller codebase fully migrating), replacing is fine; if unsure, add alongside.
• Registry-assert test (e.g. assert "claude-X" in supported_models, test_X_has_N_clusters) → add an assertion for the new model alongside; keep the old one. The old model is still served, so its assertion stays valid — but the registry should also include the new model, so assert that too. Heuristic: if the test references multiple model versions in a list, it’s a registry test; if one model in a struct compared only to itself, it’s a generic fixture.
• Frozen / generated snapshot → regenerate, don’t hand-edit.
• Coupled to a definer (e.g. an integration test that passes model authorization via a shared conftest seed list, or asserts on a billing-tier / rate-limit-group enum or a generated SKU/pricing catalog) → verify the definer has a new-model entry first. If not, add a seed entry (reusing the nearest existing tier as a placeholder); if you can’t confidently do that, ask the user how to populate the definer. Do not skip the test. Swapping without populating the definer will make the test fail at runtime.

When migrating tests specifically: breaking parameters (temperature, top_p, budget_tokens) are usually absent — test fixtures rarely set sampling params on placeholder models. The breaking-change scan is still required, but expect mostly clean results.

Find intentionally-flagged sync points first. Many codebases tag spots that must change at every model launch with comment markers like MODEL LAUNCH, KEEP IN SYNC, @model-update, or similar. Grep for whatever convention the repo uses before the broad model-ID grep — those markers point at the load-bearing changes.

Per-SDK Syntax Reference

Code examples in this guide are Python. The same fields exist in every official Anthropic SDK — Stainless generates all 7 from the same OpenAPI spec, so JSON field names map 1:1 with only case-convention differences. Use the rows below to translate the Python examples to the SDK you are migrating.

Verify type and method names against the SDK source before writing them into customer code. WebFetch the relevant repository from the SDK source-code table in shared/live-sources.md (one row per SDK) and confirm the exact symbol — particularly for typed SDKs (Go, Java, C#) where union/builder names can differ from the JSON shape. Do not guess type names that aren’t in the table below or in <lang>/claude-api/README.md.

thinking — budget_tokens → adaptive

Sampling parameters — temperature / top_p / top_k

(Remove the field entirely on Opus 4.7; on Claude 4.x keep at most one of temperature or top_p.)

Prefill replacement — structured outputs via output_config.format

thinking.display — opt back into summarized reasoning (Opus 4.7)

For any field not in these tables, the JSON key in the Python example translates directly: snake_case for Python/TypeScript/Ruby, camelCase named args for PHP, PascalCase struct fields for Go/C#, camelCase builder methods for Java.

Explain every change you make

Migration edits often look arbitrary to a user who hasn’t read the release notes — a removed temperature, a deleted prefill, a rewritten system-prompt sentence. For each edit, tell the user what you changed and why, tied to the specific API or behavioral change that motivates it. Do this in your summary as you work, not just at the end.

Be especially explicit about system-prompt edits. Users are rightly protective of their prompts, and prompt-tuning changes are judgment calls (not hard API requirements). For any prompt edit:

• Quote the before and after text.
• State the behavioral shift that motivates it (e.g. “Opus 4.7 calibrates response length to task complexity, so I added an explicit length instruction”, or “4.6 follows instructions more literally, so ‘CRITICAL: YOU MUST use the search tool’ will now overtrigger — softened to ‘Use the search tool when…’”).
• Make clear which prompt edits are optional tuning (tone, length, subagent guidance) versus which code edits are required to avoid a 400 (sampling params, budget_tokens, prefills). Never present an optional prompt change as mandatory.

If you’re applying several prompt-tuning edits at once, offer them as a short list the user can accept or decline item-by-item rather than silently rewriting their system prompt.

Before You Migrate

1. Confirm the target model ID. Use only the exact strings from shared/models.md — do not append date suffixes to aliases (claude-opus-4-6, not claude-opus-4-6-20251101). Guessing an ID will 404.
2. Check which features your code uses with this checklist:
   • thinking: {type: "enabled", budget_tokens: N} → migrate to adaptive thinking on Opus 4.6 / Sonnet 4.6 (still functional but deprecated)
   • Assistant-turn prefills (messages ending with role: "assistant") → must change on Opus 4.6 / Sonnet 4.6 (returns 400)
   • output_format parameter on messages.create() → must change on all models (deprecated API-wide)
   • max_tokens > ~16000 → must stream on any model (above ~16K risks SDK HTTP timeouts). When streaming, every current model reaches 128K except Haiku 4.5, which caps at 64K
   • Beta headers effort-2025-11-24, fine-grained-tool-streaming-2025-05-14, interleaved-thinking-2025-05-14 → GA on 4.6, remove them and switch from client.beta.messages.create to client.messages.create
   • Moving Sonnet 4.5 → Sonnet 4.6 with no effort set → 4.6 defaults to high, which may change your latency/cost profile
   • System prompts with CRITICAL, MUST, If in doubt, use X language → likely to overtrigger on 4.6 (see Prompt-Behavior Changes)
   • Coming from 3.x / 4.0 / 4.1: also check sampling params (temperature + top_p), tool versions (text_editor_20250728), refusal + model_context_window_exceeded stop reasons, trailing-newline tool-param handling
3. Test on a single request first. Run one call against the new model, inspect the response, then roll out.

Destination Models (recommended targets)

Default to the latest Opus for the caller’s tier unless they explicitly chose otherwise. The Opus migrations layer: if you’re on Opus 4.6 or older, apply each version’s section in order up to your target (e.g. 4.5 → 4.8 means the 4.6, 4.7, and 4.8 sections in sequence). A 4.7 → 4.8 move has no new breaking changes — see Migrating to Opus 4.8 below.

Retired Model Replacements

These models return 404 — update immediately:

Deprecated Models (retiring soon)

Breaking Changes by Source Model

Migrating from Sonnet 4.5 to Sonnet 4.6 (effort default change)

Sonnet 4.5 had no effort parameter; Sonnet 4.6 defaults to high. If you just switch the model string and do nothing else, you may see noticeably higher latency and token usage. Set effort explicitly.

Recommended starting points:

For non-thinking chat workloads specifically:

When to use Opus 4.6 instead: hardest and longest-horizon problems — large code migrations, deep research, extended autonomous work. Sonnet 4.6 wins on fast turnaround and cost efficiency.

Migrating to Opus 4.6 / Sonnet 4.6 (from any older model)

1. Manual extended thinking is deprecated — use adaptive thinking.

thinking: {type: "enabled", budget_tokens: N} (manual extended thinking with a fixed token budget) is deprecated on Opus 4.6 and Sonnet 4.6. Replace it with thinking: {type: "adaptive"}, which lets Claude decide when and how much to think. Adaptive thinking also enables interleaved thinking automatically (no beta header needed).

Adaptive thinking is the long-term target, and on internal evaluations it outperforms manual extended thinking. Move when you can.

Transitional escape hatch: manual extended thinking is still functional on Opus 4.6 and Sonnet 4.6 (deprecated, will be removed in a future release). If you need a hard ceiling while migrating — for example, to bound token spend on a runaway workload before you’ve tuned effort — you can keep budget_tokens around alongside an explicit effort value, then remove it in a follow-up. budget_tokens must be strictly less than max_tokens:

If the user asks for a “thinking budget” on 4.6, the preferred answer is effort — use low, medium, high, or max rather than a token count.

2. Effort parameter (Opus 4.5, Opus 4.6, Sonnet 4.6 only).

Controls thinking depth and overall token spend. Goes inside output_config, not top-level. Default is high. max is supported on Fable 5, Opus 4.6 and later, Sonnet 5, and Sonnet 4.6 — it errors on Sonnet 4.5 and Haiku 4.5.

Migrating to the 4.6 family (Opus 4.6 and Sonnet 4.6)

3. Assistant-turn prefills return 400 (Opus 4.6 and Sonnet 4.6).

Prefilled responses on the final assistant turn are no longer supported on either Opus 4.6 or Sonnet 4.6 — both return a 400. Adding assistant messages elsewhere in the conversation (e.g., for few-shot examples) still works. Pick the replacement that matches what the prefill was doing:

4. Stream for max_tokens > ~16K (all models); only Haiku 4.5 caps lower, at 64K.

Non-streaming requests hit SDK HTTP timeouts at high max_tokens, regardless of model — stream for anything above ~16K output. The streamable ceiling is 128K for every current model except Haiku 4.5, which caps at 64K.

5. Tool-call JSON escaping may differ (Opus 4.6 and Sonnet 4.6).

Both 4.6 models can produce tool call input fields with Unicode or forward-slash escaping. Always parse with json.loads() / JSON.parse() — never raw-string-match the serialized input.

All models

6. output_format → output_config.format (API-wide).

The old top-level output_format parameter on messages.create() is deprecated. Use output_config.format instead. This is not 4.6-specific — applies to every model.

Beta Headers to Remove on 4.6

Several beta headers that were required on 4.5 are now GA on 4.6 and should be removed. Leaving them in is harmless but misleading; removing them also lets you move from client.beta.messages.create(...) back to client.messages.create(...).

Once you remove all of these and finish moving to adaptive thinking, you can switch the SDK call site from the beta namespace back to the regular one:

Additional Changes When Coming from 3.x / 4.0 / 4.1 → 4.6

If you’re jumping from Opus 4.1, Sonnet 4, Sonnet 3.7, or an older Claude 3.x model directly to 4.6, apply everything above plus the items in this section. Users already on Opus 4.5 / Sonnet 4.5 can skip this.

1. Sampling parameters: temperature OR top_p, not both.

Passing both will error on every Claude 4+ model:

2. Update tool versions.

Legacy tool versions are not supported on 4+. Both the type and the name field change — text_editor_20250728 and str_replace_based_edit_tool are a pair; updating one without the other 400s. Also remove the undo_edit command from your text-editor integration:

3. Handle the refusal stop reason.

Claude 4+ can return stop_reason: "refusal" on the response. If your code only handles end_turn / tool_use / max_tokens, add a branch:

4. Handle the model_context_window_exceeded stop reason (4.5+).

Distinct from max_tokens: it means the model hit the context window limit, not the requested output cap. Handle both:

5. Trailing newlines preserved in tool call string parameters (4.5+).

4.5 and 4.6 preserve trailing newlines that older models stripped. If your tool implementations do exact string matching against tool-call input values (e.g., if name == "foo"), verify they still match when the model sends "foo\n". Normalizing with .rstrip() on the receiving side is usually the simplest fix.

6. Haiku: rate limits reset between generations.

Haiku 4.5 has its own rate-limit pool separate from Haiku 3 / 3.5. If you’re ramping traffic as you migrate, check your tier’s Haiku 4.5 limits at API rate limits (opens in a new tab) — a quota that comfortably served Haiku 3.5 traffic may need a tier bump for the same volume on 4.5.

Prompt-Behavior Changes (Opus 4.5 / 4.6, Sonnet 4.6)

These don’t break your code, but prompts that worked on 4.5-and-earlier may over- or under-trigger on 4.6. Tune as needed.

1. Aggressive instructions cause overtriggering. Opus 4.5 and 4.6 follow the system prompt much more closely than earlier models. Prompts written to overcome the old reluctance are now too aggressive:

If the model is now overtriggering a tool or skill, the fix is almost always to dial back the language, not to add more guardrails.

2. Overthinking and excessive exploration (Opus 4.6). At higher effort settings, Opus 4.6 explores more before answering. If that burns too many thinking tokens, lower effort first (medium is often the sweet spot) before adding prose instructions to constrain reasoning.

3. Overeager subagent spawning (Opus 4.6). Opus 4.6 has a strong preference for delegating to subagents. If you see it spawning a subagent for something a direct grep or read would solve, add guidance: “Use subagents only for parallel or independent workstreams. For single-file reads or sequential operations, work directly.”

4. Overengineering (Opus 4.5 / 4.6). Both models may add extra files, abstractions, or defensive error handling beyond what was asked. If you want minimal changes, prompt for it explicitly: “Only make changes directly requested. Don’t add helpers, abstractions, or error handling for scenarios that can’t happen.”

5. LaTeX math output (Opus 4.6). Opus 4.6 defaults to LaTeX (\frac{}{}, $...$) for math and technical content. If you need plain text, instruct it explicitly: “Format all math as plain text — no LaTeX, no $, no \frac{}{}. Use / for division and ^ for exponents.”

6. Skipped verbal summaries (4.6 family). The 4.6 models are more concise and may skip the summary paragraph after a tool call, jumping straight to the next action. If you rely on those summaries for visibility, add: “After completing a task that involves tool use, provide a brief summary of what you did.”

7. “Think” as a trigger word (Opus 4.5 with thinking disabled). When thinking is off, Opus 4.5 is particularly sensitive to the word think and may reason more than you want. Use consider, evaluate, or reason through instead.

Model-ID Rename Quick Reference

Older aliases (claude-opus-4-7, claude-opus-4-6, claude-opus-4-5, claude-sonnet-4-6, claude-sonnet-4-5, etc.) are still active and can be pinned if you need time before upgrading — see shared/models.md for the full legacy list.

Amazon Bedrock model IDs

If the code uses the AnthropicBedrockMantle client (Python anthropic[bedrock], TypeScript @anthropic-ai/bedrock-sdk, Java BedrockMantleBackend, Go bedrock.NewMantleClient, etc.) or targets https://bedrock-mantle.{region}.api.aws/anthropic, it is running on Claude in Amazon Bedrock. All breaking changes in this guide apply unchanged there — it serves the same Messages API shape — but model IDs carry an anthropic. provider prefix:

When migrating a Bedrock file, apply the same rename-table row as first-party, then keep/add the anthropic. prefix. Do not generate a first-party claude-* ID for a Bedrock client — it will 400.

Skip for Bedrock: the code_execution_* tool-version checklist item and the Task Budgets section — neither is available on Bedrock (see shared/platform-availability.md for the per-feature table). Everything else in this guide — effort, adaptive/extended thinking, output_config.format, thinking.display, fine-grained tool streaming, token counting — is available on Bedrock.

Out of scope: the legacy Amazon Bedrock integration (InvokeModel / Converse APIs with ARN-versioned IDs like anthropic.claude-3-5-sonnet-20241022-v2:0) uses a different request shape and model-ID format. This guide does not cover it; WebFetch the Bedrock page in shared/live-sources.md if the user is migrating between the two Bedrock integrations.

Claude Platform on AWS

If the code uses AnthropicAWS / AnthropicAws / anthropicaws.NewClient / AnthropicAwsClient (or targets https://aws-external-anthropic.{region}.api.aws), it is running on Claude Platform on AWS — Anthropic-operated, same-day API parity. Model IDs are bare first-party strings; apply the rename table above verbatim and every breaking-change section in this guide unchanged. There is nothing to skip. Do not add an anthropic. prefix (that’s Amazon Bedrock, a separate offering). See shared/claude-platform-on-aws.md for client/auth details.

Migration Checklist

Every item is tagged: [BLOCKS] items cause a 400 error, infinite loop, silent timeout, or wrong tool selection if missed — apply these as code edits, not as suggestions. [TUNE] items are quality/cost adjustments.

For each file that calls messages.create() / equivalent SDK method:

• Not done: [BLOCKS] Update the model= string to the new alias
• Not done: [BLOCKS] Replace budget_tokens with thinking={"type": "adaptive"} (deprecated on Opus 4.6 / Sonnet 4.6)
• Not done: [BLOCKS] Move format from top-level output_format into output_config.format
• Not done: [BLOCKS] Remove any assistant-turn prefills if targeting Opus 4.6 or Sonnet 4.6 (see the prefill replacement table)
• Not done: [BLOCKS] Switch to streaming if max_tokens > ~16000 (otherwise SDK HTTP timeout)
• Not done: [TUNE] Verify tool-input handling parses JSON rather than raw-string-matching the serialized input (4.6 may escape Unicode / forward slashes differently; most SDKs already expose block.input as a parsed object)
• Not done: [TUNE] Set output_config={"effort": "..."} explicitly — especially when moving Sonnet 4.5 → Sonnet 4.6 (4.6 defaults to high)
• Not done: [TUNE] Remove GA beta headers: effort-2025-11-24, fine-grained-tool-streaming-2025-05-14, token-efficient-tools-2025-02-19, output-128k-2025-02-19; remove interleaved-thinking-2025-05-14 once on adaptive thinking
• Not done: [TUNE] Switch client.beta.messages.create(...) → client.messages.create(...) once all betas are removed
• Not done: [TUNE] Review system prompt for aggressive tool language (CRITICAL:, MUST, If in doubt) and dial it back

Extra items when coming from 3.x / 4.0 / 4.1:

• Not done: [BLOCKS] Remove either temperature or top_p (passing both 400s on Claude 4+)
• Not done: [BLOCKS] Update text-editor tool type to text_editor_20250728
• Not done: [BLOCKS] Update text-editor tool name to str_replace_based_edit_tool — changing only the type and keeping name: "str_replace_editor" returns a 400
• Not done: [BLOCKS] Update code-execution tool to code_execution_20260521
• Not done: [BLOCKS] Delete any undo_edit command call sites
• Not done: [TUNE] Add handling for stop_reason == "refusal"
• Not done: [TUNE] Add handling for stop_reason == "model_context_window_exceeded" (4.5+)
• Not done: [TUNE] Verify tool-param string matching tolerates trailing newlines (preserved on 4.5+)
• Not done: [TUNE] If moving to Haiku 4.5: review rate-limit tier (separate pool from Haiku 3.x)

Verification:

• Not done: Run one test request and inspect response.stop_reason, response.usage, and whether tool-use / thinking behavior matches expectations

For cached prompts: the render order and hash inputs did not change, so existing cache_control breakpoints keep working. However, changing the model string invalidates the existing cache — the first request on the new model will write the cache fresh.

Migrating to Opus 4.7

Model ID claude-opus-4-7 is authoritative as written here. When the user asks to migrate to Opus 4.7, write model="claude-opus-4-7" exactly. Do not WebFetch to verify — this guide is the source of truth for migration target IDs. The corresponding entry exists in shared/models.md.

Claude Opus 4.7 was Anthropic’s most capable model at its launch and is now the previous-generation Opus (Opus 4.8 is current — see Migrating to Opus 4.8 below). It is highly autonomous and performs exceptionally well on long-horizon agentic work, knowledge work, vision tasks, and memory tasks. This section summarizes everything that was new at the 4.7 launch and remains the layered breaking-change path for callers coming from Opus 4.6 or older. It is layered on top of the 4.6 migration above — if the caller is jumping from Opus 4.5 or older, apply the 4.6 changes first, then this section, then the 4.8 section.

TL;DR for someone already on Opus 4.6: update the model ID to claude-opus-4-7, strip any remaining budget_tokens and sampling parameters (both 400 on Opus 4.7), give max_tokens extra headroom and re-baseline with count_tokens() against the new model, opt back into thinking.display: "summarized" if reasoning is surfaced to users, and re-tune effort — it matters more on 4.7 than on any prior Opus.

Breaking changes (will 400 on Opus 4.7)

Extended thinking removed.

thinking: {type: "enabled", budget_tokens: N} is no longer supported on Claude Opus 4.7 or later models and returns a 400 error. Switch to adaptive thinking (thinking: {type: "adaptive"}) and use the effort parameter to control thinking depth. Adaptive thinking is off by default on Claude Opus 4.7: requests with no thinking field run without thinking, matching Opus 4.6 behavior. Set thinking: {type: "adaptive"} explicitly to enable it.

If the caller wasn’t using extended thinking, no change is required — thinking is off by default, or can be set explicitly with thinking={"type": "disabled"}.

Delete budget_tokens plumbing entirely. For the replacement effort value, see Choosing an effort level on Opus 4.7 below — there is no exact 1:1 mapping from budget_tokens.

Sampling parameters removed.

The temperature, top_p, and top_k parameters are no longer accepted on Claude Opus 4.7. Requests that include them return a 400 error. Remove these fields from your request payloads. Prompting is the recommended way to guide model behavior on Claude Opus 4.7. If you were using temperature = 0 for determinism, note that it never guaranteed identical outputs on prior models.

• If the intent was determinism — use effort: "low" with a tighter prompt.
• If the intent was creative variance — the prompt replacement depends on the use case; ask the user how they want variance elicited. If you can’t ask, add a use-case-appropriate instruction along the lines of “choose something off-distribution and interesting” — e.g. for text generation, “Vary your phrasing and structure across responses”; for frontend/design, use the propose-4-directions approach under Design and frontend coding below.

Choosing an effort level on Opus 4.7

budget_tokens controlled how much to think; effort controls how much to think and act, so there is no exact 1:1 mapping. Use xhigh for best results in coding and agentic use cases, and a minimum of high for most intelligence-sensitive use cases. Experiment with other levels to further tune token usage and intelligence:

Silent default changes (no error, but behavior differs)

Thinking content omitted by default.

Thinking blocks still appear in the response stream on Claude Opus 4.7, but their thinking field is empty unless you explicitly opt in. This is a silent change from Claude Opus 4.6, where the default was to return summarized thinking text. To restore summarized thinking content on Claude Opus 4.7, set thinking.display to "summarized". The block-field name is unchanged — it is still block.thinking on a thinking-type block; do not rename it.

Detect this: any code that reads block.thinking (or equivalent) from a thinking-type block and renders it in a UI, log, or trace. The fix is the request parameter, not the response handling — add display: "summarized" to the thinking parameter:

The default is "omitted" on Claude Opus 4.7. If thinking content was never surfaced anywhere, no change needed. If your product streams reasoning to users, the new default appears as a long pause before output begins; set display: "summarized" to restore visible progress during thinking.

Updated token counting.

Claude Opus 4.7 and Claude Opus 4.6 count tokens differently. The same input text produces a higher token count on Claude Opus 4.7 than on Claude Opus 4.6, and /v1/messages/count_tokens will return a different number of tokens for Claude Opus 4.7 than it did for Claude Opus 4.6. The token efficiency of Claude Opus 4.7 can vary by workload shape. Prompting interventions, task_budget, and effort can help control costs and ensure appropriate token usage. Keep in mind that these controls may trade off model intelligence. Update your max_tokens parameters to give additional headroom, including compaction triggers. Claude Opus 4.7 provides a 1M context window at standard API pricing with no long-context premium.

What else to check:

• Client-side token estimators (tiktoken-style approximations) calibrated against 4.6
• Cost calculators that multiply tokens by a fixed per-token rate
• Rate-limit retry thresholds keyed to measured token counts

Re-baseline by re-running client.messages.count_tokens() against claude-opus-4-7 on a representative sample of the caller’s prompts. Do not apply a blanket multiplier. For cost-sensitive workloads, consider reducing effort by one level (e.g. high → medium). For agentic loops, consider adopting Task Budgets (below).

New feature: Task Budgets (beta)

Opus 4.7 introduces task budgets — tell Claude how many tokens it has for a full agentic loop (thinking + tool calls + final output). The model sees a running countdown and uses it to prioritize work and wrap up gracefully as the budget is consumed.

This is a suggestion the model is aware of, not a hard cap. It is distinct from max_tokens, which remains the enforced per-response limit and is not surfaced to the model. Use task_budget when you want the model to self-moderate; use max_tokens as a hard ceiling to cap usage.

Requires beta header task-budgets-2026-03-13:

Set a generous budget for open-ended agentic tasks and tighten it for latency-sensitive ones. Minimum task_budget.total is 20,000 tokens. If the budget is too restrictive for the task, the model may complete it less thoroughly, referencing its budget as the constraint. Do not add task_budget during a migration unless you are sure the budget value is right — if you can run the workload and measure, do so; otherwise ask the user for the value rather than guessing. This is the primary lever for offsetting the token-counting shift on agentic workloads.

Capability improvements

High-resolution vision. Opus 4.7 is the first Claude model with high-resolution image support. Maximum image resolution is 2576 pixels on the long edge (up from 1568px on Opus 4.6 and prior). This unlocks gains on vision-heavy workloads, especially computer use and screenshot/artifact/document understanding. Coordinates returned by the model now map 1:1 to actual image pixels, so no scale-factor math is needed.

High-res support is automatic on Opus 4.7 — no beta header, no client-side opt-in required. The model accepts larger inputs and returns pixel-accurate coordinates out of the box.

Token cost. Full-resolution images on Opus 4.7 can use up to ~3× more image tokens than on prior models (up to ~4784 tokens per image, vs. the previous ~1,600-token cap). If the extra fidelity isn’t needed, downsample client-side before sending to control cost — but do not add downsampling by default during a migration. If you’re not sure whether the pipeline needs the fidelity, ask the user rather than guessing. Use count_tokens() on representative images on Opus 4.7 to re-baseline before reacting to any measured cost shift.

Beyond resolution, Opus 4.7 also improves on low-level perception (pointing, measuring, counting) and natural-image bounding-box localization and detection.

Knowledge work. Meaningful gains on tasks where the model visually verifies its own output — .docx redlining, .pptx editing, and programmatic chart/figure analysis (e.g. pixel-level data transcription via image-processing libraries). If prompts have scaffolding like “double-check the slide layout before returning”, try removing it and re-baselining.

Memory. Opus 4.7 is better at writing and using file-system-based memory. If an agent maintains a scratchpad, notes file, or structured memory store across turns, that agent should improve at jotting down notes to itself and leveraging its notes in future tasks.

User-facing progress updates. Opus 4.7 provides more regular, higher-quality interim updates during long agentic traces. If the system prompt has scaffolding like “After every 3 tool calls, summarize progress”, try removing it to avoid excessive user-facing text. If the length or contents of Opus 4.7’s updates are not well-calibrated to your use case, explicitly describe what these updates should look like in the prompt and provide examples.

Real-time cybersecurity safeguards

Requests that involve prohibited or high-risk topics may lead to refusals.

Fast Mode: Claude Opus 5 / Opus 4.8 only

Fast mode is available on Claude Opus 5 and Opus 4.8. Only surface this if the caller’s code actually uses fast mode (e.g. model="claude-opus-4-6-fast", or speed="fast" on an unsupported model); if the word “fast” does not appear in the code, say nothing about Fast Mode.

When you see model="claude-opus-4-6-fast" (or any retired -fast model string), the migration edit is to move the fast-mode traffic onto Claude Opus 5, the current fast-capable default (Opus 4.8 also works if the caller is staying on that tier):

That is: switch the model to Claude Opus 5 (or Opus 4.8) and request fast mode the supported way, using the beta client.beta.messages.… endpoint, the fast-mode-2026-02-01 beta flag, and speed="fast" as a top-level request parameter (per-language form in SKILL.md § Fast Mode). Opus 4.7 fast mode has also been removed, so do not land on Opus 4.7 either. Do not leave the code on a retired -fast model string — the failure mode differs by version: claude-opus-4-6-fast is retired and the API silently falls back to standard Opus 4.6 (no error — the caller loses fast-mode speed without noticing); claude-opus-4-7-fast and speed="fast" on Opus 4.7 instead return an API error (hard failure — requests break outright rather than degrading). Either way, migrate to Opus 4.8 fast mode now.

Behavioral shifts (prompt-tunable)

These don’t break anything, but prompts tuned for Opus 4.6 may land differently. Opus 4.7 is more steerable than 4.6, so small prompt nudges usually close the gap.

More literal instruction following. Claude Opus 4.7 interprets prompts more literally and explicitly than Claude Opus 4.6, particularly at lower effort levels. It will not silently generalize an instruction from one item to another, and it will not infer requests you didn’t make. The upside of this literalism is precision and less thrash. It generally performs better for API use cases with carefully tuned prompts, structured extraction, and pipelines where you want predictable behavior. A prompt and harness review may be especially helpful for migration to Claude Opus 4.7.

Verbosity calibrates to task complexity. Opus 4.7 scales response length to how complex it judges the task to be, rather than defaulting to a fixed verbosity — shorter answers on simple lookups, much longer on open-ended analysis. If the product depends on a particular length or style, tune the prompt explicitly. To reduce verbosity:

“Provide concise, focused responses. Skip non-essential context, and keep examples minimal.”

If you see specific kinds of over-verbosity (e.g. over-explaining), add instructions targeting those. Positive examples showing the desired level of concision tend to be more effective than negative examples or instructions telling the model what not to do. Do not assume existing “be concise” instructions should be removed — test first.

Tone and writing style. Opus 4.7 is more direct and opinionated, with less validation-forward phrasing and fewer emoji than Opus 4.6’s warmer style. As with any new model, prose style on long-form writing may shift. If the product relies on a specific voice, re-evaluate style prompts against the new baseline. If a warmer or more conversational voice is wanted, specify it:

“Use a warm, collaborative tone. Acknowledge the user’s framing before answering.”

effort matters more than on any prior Opus. Opus 4.7 respects effort levels more strictly, especially at the low end. At low and medium it scopes work to what was asked rather than going above and beyond — good for latency and cost, but on moderate tasks at low there is some risk of under-thinking.

• If shallow reasoning shows up on complex problems, raise effort to high or xhigh rather than prompting around it.
• If effort must stay low for latency, add targeted guidance: “This task involves multi-step reasoning. Think carefully through the problem before responding.”
• At xhigh or max, set a large max_tokens so the model has room to think and act across tool calls and subagents. Start at 64K and tune from there. (xhigh is a new effort level on Opus 4.7, between high and max.)

Adaptive-thinking triggering is also steerable. If the model thinks more often than wanted — which can happen with large or complex system prompts — add: “Thinking adds latency and should only be used when it will meaningfully improve answer quality — typically for problems that require multi-step reasoning. When in doubt, respond directly.”

Uses tools less often by default. Opus 4.7 tends to use tools less often than 4.6 and to use reasoning more. This produces better results in most cases, but for products that rely on tools (search/retrieval, function-calling, computer-use steps), it can drop tool-use rate. Two levers:

• Raise effort — high or xhigh show substantially more tool usage in agentic search and coding, and are especially useful for knowledge work.
• Prompt for it — be explicit in tool descriptions or the system prompt about when and how to use the tool, and encourage the model to err on the side of using it more often:

“When the answer depends on information not present in the conversation, you MUST call the search tool before answering — do not answer from prior knowledge.”

Fewer subagents by default. Opus 4.7 tends to spawn fewer subagents than 4.6. This is steerable — give explicit guidance on when delegation is desirable. For a coding agent, for example:

“Do NOT spawn a subagent for work you can complete directly in a single response (e.g. refactoring a function you can already see). Spawn multiple subagents in the same turn when fanning out across items or reading multiple files.”

Design and frontend coding. Opus 4.7 has stronger design instincts than 4.6, with a consistent default house style: warm cream/off-white backgrounds (around #F4F1EA), serif display type (Georgia, Fraunces, Playfair), italic word-accents, and a terracotta/amber accent. This reads well for editorial, hospitality, and portfolio briefs, but will feel off for dashboards, dev tools, fintech, healthcare, or enterprise apps — and it appears in slide decks as well as web UIs.

The default is persistent. Generic instructions (“don’t use cream,” “make it clean and minimal”) tend to shift the model to a different fixed palette rather than producing variety. Two approaches work reliably:

1. Specify a concrete alternative. The model follows explicit specs precisely — give exact hex values, typefaces, and layout constraints.

2. Have the model propose options before building. This breaks the default and gives the user control:

   “Before building, propose 4 distinct visual directions tailored to this brief (each as: bg hex / accent hex / typeface — one-line rationale). Ask the user to pick one, then implement only that direction.”

If the caller previously relied on temperature for design variety, use approach (2) — it produces meaningfully different directions across runs.

Opus 4.7 also requires less frontend-design prompting than previous models to avoid generic “AI slop” aesthetics. Where earlier models needed a lengthy anti-slop snippet, Opus 4.7 generates distinctive, creative frontends with a much shorter nudge. This snippet works well alongside the variety approaches above:

“NEVER use generic AI-generated aesthetics like overused font families (Inter, Roboto, Arial, system fonts), cliched color schemes (particularly purple gradients on white or dark backgrounds), predictable layouts and component patterns, and cookie-cutter design that lacks context-specific character. Use unique fonts, cohesive colors and themes, and animations for effects and micro-interactions.”

Interactive coding products. Opus 4.7’s token usage and behavior can differ between autonomous, asynchronous coding agents with a single user turn and interactive, synchronous coding agents with multiple user turns. Specifically, it tends to use more tokens in interactive settings, primarily because it reasons more after user turns. This can improve long-horizon coherence, instruction following, and coding capabilities in long interactive coding sessions, but also comes with more token usage. To maximize both performance and token efficiency in coding products, use effort: "xhigh" or "high", add autonomous features (like an auto mode), and reduce the number of human interactions required from users.

When limiting required user interactions, specify the task, intent, and relevant constraints upfront in the first human turn. Well-specified, clear, and accurate task descriptions upfront help maximize autonomy and intelligence while minimizing extra token usage after user turns — because Opus 4.7 is more autonomous than prior models, this usage pattern helps to maximize performance. In contrast, ambiguous or underspecified prompts conveyed progressively over multiple user turns tend to reduce token efficiency and sometimes performance.

Code review. Opus 4.7 is meaningfully better at finding bugs than prior models, with both higher recall and precision. However, if a code-review harness was tuned for an earlier model, it may initially show lower recall — this is likely a harness effect, not a capability regression. When a review prompt says “only report high-severity issues,” “be conservative,” or “don’t nitpick,” Opus 4.7 follows that instruction more faithfully than earlier models did: it investigates just as thoroughly, identifies the bugs, and then declines to report findings it judges to be below the stated bar. Precision rises, but measured recall can fall even though underlying bug-finding has improved.

Recommended prompt language:

“Report every issue you find, including ones you are uncertain about or consider low-severity. Do not filter for importance or confidence at this stage — a separate verification step will do that. Your goal here is coverage: it is better to surface a finding that later gets filtered out than to silently drop a bug. For each finding, include your confidence level and an estimated severity so a downstream filter can rank them.”

This can be used without an actual second step, but moving confidence filtering out of the finding step often helps. If the harness has a separate verification/dedup/ranking stage, tell the model explicitly that its job at the finding stage is coverage, not filtering. If single-pass self-filtering is wanted, be concrete about the bar rather than using qualitative terms like “important” — e.g. “report any bugs that could cause incorrect behavior, a test failure, or a misleading result; only omit nits like pure style or naming preferences.” Iterate on prompts against a subset of evals to validate recall or F1 gains.

Computer use. Computer use works across resolutions up to the new 2576px / 3.75MP maximum. Sending images at 1080p provides a good balance of performance and cost. For particularly cost-sensitive workloads, 720p or 1366×768 are lower-cost options with strong performance. Test to find the ideal settings for the use case; experimenting with effort can also help tune behavior.

Opus 4.7 Migration Checklist

Every item is tagged: [BLOCKS] items cause a 400 error, infinite loop, silent truncation, or empty output if missed — apply these as code edits, not as suggestions. [TUNE] items are quality/cost adjustments — surface them to the user as recommendations.

[BLOCKS] items prefixed with “If…” or “At…” are conditional. Before working through the list, scan the file for the conditions: does it surface thinking text to a UI/log? Does it set output_config.effort to "x-high" or "max"? Is it a security workload? Is it a multi-turn agentic loop? Apply only the items whose condition matches.

• Not done: [BLOCKS] Replace thinking: {type: "enabled", budget_tokens: N} with thinking: {type: "adaptive"} + output_config.effort; delete budget_tokens plumbing entirely
• Not done: [BLOCKS] Strip temperature, top_p, top_k from request construction
• Not done: [BLOCKS] If thinking content is surfaced to users or stored in logs: add thinking.display: "summarized" (otherwise the rendered text is empty)
• Not done: [BLOCKS] At output_config.effort of xhigh or max: set max_tokens ≥ 64000 (otherwise output truncates mid-thought)
• Not done: [TUNE] Give max_tokens and compaction triggers extra headroom; re-run count_tokens() against claude-opus-4-7 on representative prompts to re-baseline (no blanket multiplier)
• Not done: [TUNE] Re-baseline cost and rate-limit dashboards before reacting to measured shifts
• Not done: [TUNE] Re-evaluate effort per route — use xhigh for coding/agentic and a minimum of high for most intelligence-sensitive work; it matters more on 4.7 than any prior Opus
• Not done: [TUNE] Multi-turn agentic loops: adopt the API-native Task Budgets (output_config.task_budget, beta task-budgets-2026-03-13, minimum 20k tokens) — this is for capping cumulative spend across a loop; per-turn depth is effort
• Not done: [TUNE] Check for ambiguous or underspecified instructions that relied on 4.6 generalizing intent, and update them to be clearer or more precise — 4.7 follows them literally
• Not done: [TUNE] Tool-use workloads: add explicit when/how-to-use guidance to tool descriptions (4.7 reaches for tools less often)
• Not done: [TUNE] Verbosity: test existing length instructions before changing them — 4.7 calibrates length to task complexity, so tune for the desired output rather than assuming a direction
• Not done: [TUNE] Remove forced-progress-update scaffolding (“after every N tool calls…”)
• Not done: [TUNE] Remove knowledge-work verification scaffolding (“double-check the slide layout…”) and re-baseline
• Not done: [TUNE] Add tone instruction if a warmer / more conversational voice is needed; re-evaluate style prompts on writing-heavy routes
• Not done: [TUNE] Subagent tool present: add explicit spawn / don’t-spawn guidance
• Not done: [TUNE] Frontend/design output: specify a concrete palette/typeface, or have the model propose 4 visual directions before building (the default cream/serif house style is persistent)
• Not done: [TUNE] Interactive coding products: use effort: "xhigh" or "high", add autonomous features (e.g. an auto mode) to reduce human interactions, and specify task/intent/constraints upfront in the first turn
• Not done: [TUNE] Code-review harnesses: remove or loosen “only report high-severity” / “be conservative” filters and have the model report every finding with confidence + severity; move filtering to a downstream step (4.7 follows severity filters more literally, which can depress measured recall)
• Not done: [TUNE] Vision-heavy pipelines (screenshots, charts, document understanding): leave images at native resolution up to 2576px long edge for the accuracy gain; remove any scale-factor math from coordinate handling (coords are now 1:1 with pixels). No beta header / opt-in needed — high-res is automatic on Opus 4.7.
• Not done: [TUNE] Computer-use pipelines: send screenshots at 1080p for a good performance/cost balance (720p or 1366×768 for cost-sensitive workloads); experiment with effort to tune behavior
• Not done: [TUNE] Cost-sensitive image pipelines: full-res images on 4.7 use up to ~4784 tokens vs ~1,600 on prior models (~3×). Downsampling client-side before upload avoids the increase, but do not downsample by default — if you’re unsure whether fidelity is needed, ask the user. Re-baseline with count_tokens() on representative images before reacting to cost shifts.

Migrating to Opus 4.8

Model ID claude-opus-4-8 is authoritative as written here. When the user asks to migrate to Opus 4.8, write model="claude-opus-4-8" exactly. Do not WebFetch to verify — this guide is the source of truth for migration target IDs. The corresponding entry exists in shared/models.md.

Claude Opus 4.8 is our most capable Opus-tier model — highly autonomous, with state-of-the-art long-horizon agentic execution, knowledge work, and memory. It is layered on top of the Opus 4.7 migration above. If the caller is jumping from Opus 4.6 or older, apply the 4.6 and 4.7 sections first, then this one.

No new breaking changes. Opus 4.8 keeps the same request surface as Opus 4.7. The same calls that already work on 4.7 work unchanged on 4.8 — adaptive thinking only (thinking: {type: "enabled", budget_tokens: N} still 400s; use {type: "adaptive"}), sampling parameters (temperature, top_p, top_k) still rejected, last-assistant-turn prefills still 400, thinking.display still defaults to "omitted", and the low/medium/high/xhigh/max effort levels, Task Budgets (beta), and high-resolution vision all behave as on 4.7. A 4.7 → 4.8 migration is therefore the model-ID swap plus prompt re-tuning — there is no required code edit beyond the model string.

TL;DR for someone already on Opus 4.7: swap the model ID to claude-opus-4-8. Nothing else is required to avoid an error. Then re-tune prompts for the behavioral shifts: 4.8 narrates more than 4.7 (add a silence-default if you want 4.7-like terseness), writes in a warmer, less hedged voice, is more deliberate and asks more often (add autonomy guidance to claw back ask-rate), and is more conservative about reaching for search, subagents, file-based memory, and custom tools (add explicit “when to use this” triggering). For long-horizon agentic work, give the full task specification up front in one well-specified turn and run at high effort.

No new API breaking changes (inherited from 4.7)

These all carry over from Opus 4.7 unchanged — apply them only if the caller is coming from Opus 4.6 or earlier (see the Migrating to Opus 4.7 section above for the before/after and the SDK-specific syntax):

• thinking: {type: "enabled", budget_tokens: N} → 400. Use thinking: {type: "adaptive"} + output_config.effort.
• temperature, top_p, top_k → 400. Remove them; steer with prompting.
• Last-assistant-turn prefills → 400. Use output_config.format (structured outputs) or a system-prompt instruction.
• thinking.display defaults to "omitted"; set "summarized" if you surface reasoning to users.

If the caller is already on Opus 4.7 and these are clean, there is nothing to change here.

New API feature: mid-session system prompts

You can deliver trusted instructions partway through a session by placing {"role": "system", ...} entries directly in the messages array — without editing the top-level system prompt and invalidating your prompt cache. Use it for things the application learns mid-session: the user delivered async context, a mode toggled (auto-approve enabled), files changed on disk, the remaining token budget dropped.

Phrase these as context, not commands. State the fact and let Claude act on it; avoid override-style language (“ignore what the user said”, “regardless of the user’s request”, “disregard the previous instruction”). Claude is trained to protect users from instructions that appear to work against them, and that protection applies to the system role too. No beta header is required; available on Claude Opus 4.8. For cache-placement details and the older-model <system-reminder> fallback, see shared/prompt-caching.md and shared/agent-design.md.

Capability improvements

Long-horizon agentic execution. Opus 4.8 is state-of-the-art at long, autonomous agentic work — complex refactors and overnight coding runs that complete without human correction. To get the most out of it, give the full task specification up front in a single well-specified initial turn and run at high effort (effort: "high" or "xhigh"). Its long-horizon coherence comes partly from reasoning more at each step; combined with a clear up-front goal, that more-intelligent planning often produces more efficient and more accurate output than prior frontier models. The “clear goal up front” principle maps to two product surfaces: in Claude Code, /goal sets direction for the run; with Managed Agents (CMA), state what “done” looks like via an Outcome (user.define_outcome with a gradeable rubric — the harness runs an iterate → grade → revise loop), see shared/managed-agents-outcomes.md.

Effort is a dimension to test, not a fixed setting. On prior models many reached for xhigh reflexively to maximize intelligence. Opus 4.8 has a higher intelligence ceiling, so start at high as the default and iterate rather than defaulting to xhigh. Sweep medium, high, and xhigh on your own eval set and weigh the intelligence ↔ latency ↔ cost tradeoff per route — the relationship isn’t monotonic: higher effort up front often reduces turn count and total cost on agentic work, while for some tasks medium delivers equally good results in less time. Reserve max for extremely hard, latency-insensitive cases. The per-level effort table in the Migrating to Opus 4.7 section above applies unchanged on 4.8.

Writing voice and clarity. Testers consistently describe 4.8’s prose as clearer, warmer, and less hedged than prior models, with fewer measurable AI vocal tics — especially at higher effort, where it approaches expert-level prose and structure. This is roughly the opposite direction from the 4.7 shift (4.7 was more clipped, direct, and less validation-forward). If you added style prompts to counter 4.7’s terseness or to inject warmth, re-evaluate them against the new baseline before keeping them — they may now overcorrect. 4.8 is also a stronger thought partner: more thoughtful, more willing to push back, and more likely to infer the right answer from context.

Code review and debugging. Stronger real-bug finding and clearer explanations than 4.7 — one-shot fixes where 4.7 needed more, and correctly identifying intermittent flakes rather than declaring “fixed” after one clean run. The 4.7 caveat still applies: if a review harness says “only report high-severity issues” or “be conservative”, 4.8 follows it literally and measured recall can drop even though underlying bug-finding improved. Tell the model to report everything and filter downstream (or review a second time) — see the Code review guidance in the 4.7 section for the recommended prompt.

Behavioral shifts (prompt-tunable)

None of these break code, but prompts tuned for Opus 4.7 may land differently. 4.8 follows instructions well, so small, explicit nudges close the gap.

Tool triggering is surface-dependent (search & knowledge). 4.8’s tool-triggering is more surface-dependent than in prior models: with a system prompt present it is high-precision / low-recall — web search triggers slightly more often but runs fewer rounds per trigger, while knowledge-retrieval tools (Drive, project knowledge, connected files) trigger less often. It searches when it’s confident search is needed and otherwise answers from context, which can lower research depth on tasks that need it. Recover should-search rate with an explicit search-first instruction:

Copy

<search_first>
For questions where current information would change the answer (recent events, current roles or prices, version-specific behavior, or anything the user flags as time-sensitive) search before answering rather than answering from memory. For open-ended research requests, begin searching immediately; do not ask a scoping question first unless the request is genuinely ambiguous about what to research.
</search_first>

Under-utilization of subagents, memory, and custom tools. Separately from search, 4.8 is conservative about reaching for capabilities that need an explicit “decide to use this” step — file-based memory, subagent delegation, custom tools. It won’t reach for complex or expensive capabilities unless reasonably sure they’re needed. This is steerable since 4.8 follows instructions well — say when each capability applies, not just that it exists:

“Before any task longer than a few turns, check your memory file for relevant prior context and write new findings to it as you go. When a task fans out across independent items (many files to read, many tests to run, many candidates to check), delegate to subagents rather than iterating serially.”

The same lever works at the tool-description level, not just the system prompt: prescriptive descriptions that state when to call a tool (e.g. “Call this when the user asks about current prices or recent events”) give meaningful lift on 4.8 over descriptions that only state what the tool does. Make the trigger condition part of each capability’s own description.

More user-facing narration. 4.8 narrates more than 4.7 — more text between tool calls in long tool-calling sessions, and longer, more detailed end-of-task wrap-ups by default. If you previously added scaffolding to force interim status (“after every 3 tool calls, summarize progress”), remove it — 4.8 does this on its own. If the narration is too verbose for a coding agent, an explicit silence-default makes it behave like 4.7 with no loss of quality:

“Default to silence between tool calls. Only write text when you find something, change direction, or hit a blocker — one sentence each. Do not narrate routine actions (‘Now I’ll…’, ‘Let me check…’, ‘Looking at…’). When done: one or two sentences on the outcome. Do not recap every file or test — the user has been following along.”

For knowledge-work deliverables (reports, analysis readouts), verbosity responds very well to instructions in user preferences or the user turn — expose a verbosity preference rather than hard-coding a length.

More deliberate — asks more often. 4.8 is more deliberate than prior Opus models. On minor decisions it would previously just make (a variable name, a default value, which of two equivalent approaches), it tends to pause and ask, and it often closes a completed task with “Want me to also…?” rather than doing the obvious next step or stopping cleanly. This is preferred for high-stakes or unfamiliar codebases, but bugs users when uncalibrated. Grant autonomy on the small stuff while keeping caution where it matters (in Claude Code testing this cut ask-rate by ~12 percentage points with no increase in over-reach):

“For minor choices (naming, formatting, default values, which approach among equivalents), pick a reasonable option and note it rather than asking. For scope changes or destructive actions, still ask first.”

Verbose reasoning when thinking is disabled. With thinking: {type: "disabled"}, 4.8 occasionally writes longer explanations of its reasoning into the visible response, which reads as verbose when the user wants a fast, quick answer. The simplest fix is to leave adaptive thinking on — set thinking: {type: "adaptive"} (the recommended setting; it adjusts how much to think per task). Note adaptive is not on when the field is omitted — like Opus 4.7, a request with no thinking field runs without thinking, so set it explicitly. If you need thinking off for latency or cost, scope it in the system prompt:

“Respond only with your final answer. Do not include exploratory reasoning, intermediate drafts, diffs you considered but rejected, or meta-commentary about your process.”

Opus 4.8 Migration Checklist

Every item is tagged: [BLOCKS] items cause a 400 error if missed; [TUNE] items are quality/cost adjustments — surface them to the user as recommendations.

For a caller already on Opus 4.7, only the first item is required; everything else is [TUNE]. The conditional [BLOCKS] item applies only when coming from Opus 4.6 or earlier.

• Not done: [BLOCKS] Update the model= string to claude-opus-4-8
• Not done: [BLOCKS] (only if coming from Opus 4.6 or earlier) Apply the Migrating to Opus 4.7 breaking changes first — budget_tokens → adaptive thinking, strip temperature/top_p/top_k, remove last-assistant-turn prefills. These already 400 on 4.7 and continue to 400 on 4.8.
• Not done: [TUNE] Long-horizon / agentic work: put the full task spec in one well-specified first turn and run at high or xhigh effort (Claude Code: /goal; Managed Agents: an Outcome with a gradeable rubric)
• Not done: [TUNE] Effort: sweep medium / high / xhigh on your eval set and pick per route by the intelligence ↔ latency ↔ cost tradeoff (default high, xhigh for coding/agentic)
• Not done: [TUNE] Research depth & tool use: add a search-first instruction; add explicit triggering guidance for subagents, file-based memory, and custom tools (4.8 under-reaches for these by default) — in the system prompt and in each tool’s own description (prescriptive “call this when…” descriptions give measurable lift)
• Not done: [TUNE] Narration: remove forced-progress scaffolding (“after every N tool calls…”); add a silence-default if a coding agent is too chatty
• Not done: [TUNE] Autonomy: add small-decisions-don’t-ask guidance to cut ask-rate, while keeping caution on scope changes / destructive actions
• Not done: [TUNE] Writing voice: re-evaluate style prompts added to counter 4.7’s directness — 4.8 is warmer and less hedged by default; re-baseline before keeping them
• Not done: [TUNE] Code-review harnesses: keep the report-everything-filter-downstream pattern (4.8 follows “only high-severity” / “be conservative” filters literally, which can depress measured recall)
• Not done: [TUNE] Thinking-disabled paths: add a final-answer-only instruction if reasoning leaks into the visible response
• Not done: [TUNE] Consider mid-session system messages (role:"system" in messages; no beta header) for context the app learns mid-session, instead of rebuilding the top-level system prompt and invalidating the cache

Migrating to Claude Opus 5

Model ID claude-opus-5 is authoritative as written here. When the user asks to migrate to Claude Opus 5, write model="claude-opus-5" exactly. Do not WebFetch to verify — this guide is the source of truth for migration target IDs. The corresponding entry exists in shared/models.md.

Claude Opus 5 is the successor to Claude Opus 4.8 in the Opus line, and is strongest on long-horizon agentic work and coding. It is layered on top of the Opus 4.8 migration above; if the caller is coming from Opus 4.7 or older, apply those sections first. Like Claude Fable 5, it ships with elevated cybersecurity safeguards, and its safety classifiers can decline a request: you get a normal HTTP 200 with stop_reason: "refusal" and a stop_details category, not an error. Benign security and life-sciences work occasionally trips them, so check stop_reason before reading response.content — code that indexes content[0] unconditionally breaks on a refusal. Cyber-category refusals route to Opus 4.8 as the recommended fallback, so a fallback strategy genuinely recovers the request rather than just relabelling the failure. The full refusal semantics (pre-output vs mid-stream billing, retry strategies, fallback credit) are in the Claude Fable 5 section below and apply here unchanged.

Existing prompts and evals should carry over with strong out-of-the-box performance. It is a drop-in upgrade at Opus 4.8’s pricing — $5 per million input tokens, $25 per million output — with the same feature set: 1M context (default, no beta header), 128K max output, adaptive thinking, prompt caching, batch processing, the Files API, PDF support, vision, and the full server-side and client-side tool set. claude-opus-5 is a fixed ID with no date suffix, same scheme as claude-opus-4-8.

The migration is the model-ID swap plus prompt re-tuning, with two breaking changes covered below.

Availability at launch: Claude API (claude-opus-5), Amazon Bedrock (anthropic.claude-opus-5), Google Cloud (claude-opus-5), and Microsoft Foundry. Opus 4.8 stays available on all four.

Rate limits are a separate bucket. Opus 4.8/4.7/4.6/4.5 share one combined Opus limit; Claude Opus 5 does not draw from it. Shifting traffic over neither frees headroom on the old bucket nor inherits it — check your tier’s Claude Opus 5 limits before moving volume.

TL;DR for someone already on Claude Opus 4.8: swap the model ID. Then re-tune: Claude Opus 5 writes longer user-facing responses and longer files on disk (add explicit conciseness and deliverable-length instructions — effort does not reliably shorten visible output), verifies its own work without being told (delete your verification instructions and harness verification steps), and can expand task scope (add a scope-discipline instruction). Run a fresh effort sweep — low and medium are unusually strong here and are the primary cost/latency lever.

Breaking change 1: thinking is on by default

A request that omits the thinking parameter thinks on Claude Opus 5, unlike Claude Opus 4.8 and Opus 4.7 where omitting it meant no thinking. thinking: {type: "adaptive"} remains valid and is equivalent to the default — the wire value didn’t change, the default did.

This is a silent cost and truncation change, not just a behavior one: max_tokens is a hard cap on thinking plus response text. A workload that ran without thinking on Opus 4.8 and sized max_tokens tightly around its answer can now truncate mid-response. Revisit max_tokens on every route that never set thinking. To keep the old behavior, pass thinking: {type: "disabled"} — subject to the effort cap below.

Raw thinking tokens are never returned on Claude Opus 5; display defaults to "omitted", and display: "summarized" gets you a summary. This also means a fallback model cannot read Claude Opus 5’s thinking.

Breaking change 2: disabling thinking is capped at high effort

Disabling thinking is available only at effort high or lower; thinking: {type: "disabled"} combined with xhigh or max returns a 400. Opus 4.8 accepts that combination, so audit any route that disables thinking before migrating.

The check is per request. Effort and thinking are validated independently on every call, so a later request that raises effort to xhigh while thinking is still disabled is rejected even though earlier requests in the same conversation succeeded.

Migrating: either enable thinking at xhigh/max, or lower effort to high or below. Given how well Claude Opus 5 performs at low and medium, a latency-sensitive route that previously ran xhigh + disabled thinking is usually better served by medium with thinking on than by keeping the disabled path.

Everything else from the Opus 4.7/4.8 request surface is unchanged: budget_tokens still 400s (use output_config.effort), sampling parameters (temperature, top_p, top_k) are still rejected, last-assistant-turn prefills still 400, and thinking.display still defaults to "omitted".

Two failure modes when thinking is disabled

Are you affected? Only if you explicitly set thinking: {type: "disabled"}. Thinking is on by default on Claude Opus 5 (see Breaking change 1 above), so an unmodified request never hits either of these — but code carrying a disabled-thinking setting forward from Opus 4.8, where it was the default behaviour, does.

Both are specific to thinking: {type: "disabled"} on Claude Opus 5, and for both the primary recommendation is the same: turn thinking back on and use a lower effort to control cost and verbosity instead. Disabling thinking is the more expensive lever in every sense — it is what triggers these, and low/medium effort already gets you most of the token and latency saving (see § Effort below).

1. Tool calls can arrive as plain text. The model occasionally writes a tool call into its user-facing text rather than emitting a structured tool_use block. The turn completes normally and the call never runs — there is no error and no tool_use block to catch, so a harness sees a successful turn that silently did nothing. Worse in an agentic loop: the bogus text stays in conversation history and skews later turns. Most common on tool-heavy workloads such as search.

2. <thinking> tags can leak into the visible response. The model may emit <thinking> or other internal XML in its user-facing output.

If you cannot enable thinking, one instruction covers both failure modes — give the model explicit permission to talk before a tool call (the tool-as-text failure appears to come from suppressing the preamble it wants to write), and forbid internal tags generically:

“When you use a tool, you may say a brief sentence first. If no tool can express what the user asked for, say so instead of guessing. Do not include internal or system XML tags in your response.”

Two counterintuitive rules for that instruction:

• Delete any instruction telling the model not to think or not to reason. That kind of rule increases tag leakage rather than suppressing it.
• Do not name thinking tags in the prompt. Calling out <thinking> by name is measurably less effective than the generic “internal or system XML tags” wording above.

New API features

Two additions, each behind its own beta header. Both are optional — a migrated request works without them.

1. fallbacks: "default" — recommended for every caller. Claude Opus 5’s safety classifiers can decline a request; the fallbacks parameter re-runs a declined request on another model server-side instead of returning the refusal to you. Previously you named the substitute yourself ("fallbacks": [{"model": "claude-opus-4-8"}]). The new "default" mode picks Anthropic’s recommended fallback automatically, routed by refusal category — cyber-category refusals go to Claude Opus 4.8.

Prefer "default" over pinning a model. Different fallback models carry different classifiers, so the right substitute depends on why the request was declined — and "default" removes the migration you would otherwise owe when a pinned fallback model is deprecated. Note the header is server-side-fallback-2026-07-01, distinct from the -2026-06-01 header that gates the array form; the array form’s semantics (content blocks, usage.iterations, sticky routing) are unchanged and documented in the Claude Fable 5 refusal section below.

2. Mid-conversation tool changes (beta mid-conversation-tool-changes-2026-07-01). Change a conversation’s tool set between turns without invalidating the prompt cache. Previously tools was fixed for the conversation’s lifetime and any edit re-billed the whole prefix. Append a {"role": "system", "content": [...]} message carrying a tool_addition or tool_removal block:

The added tool must already be declared in tools[] with "defer_loading": True — declared up front, but not loaded into context until a tool_addition surfaces it. A tool_removal block must sit either immediately before an assistant message or at the end of messages. To change a tool’s definition, remove the old one on one request, then send the updated entry in tools[] on the next. See shared/tool-use-concepts.md § Mid-conversation tool changes.

⚠️ Earlier previews of this feature used a different beta header and different block shapes. Both are deprecated — if the code you’re migrating carries anything other than mid-conversation-tool-changes-2026-07-01 with tool_addition / tool_removal / tool_reference, update the header and the shapes together.

SDK typings lag these blocks. Pass them as plain dicts in Python (the SDK forwards unknown keys unchanged) or add a @ts-expect-error in TypeScript until the types catch up. extra_body / extra_headers work on .stream() exactly as on .create().

Capability improvements

Agentic coding. Claude Opus 5 is a workhorse for agentic coding and is strongest on difficult tasks — multi-file features, larger refactors, end-to-end feature work. It completes tasks rather than leaving stubs or placeholders. The gap over prior models is smaller on easy single-turn edits, so evaluate it on the hard end of your workload. To get the most out of it, give the complete task specification up front and let it run; longer autonomous sessions with more parallel agents show the strongest results, short interactive edits the least.

Code review and bug-finding. High precision and high recall — a high rate of real bugs per pass, with the extra findings mostly real rather than false positives. It stays accurate at lower effort, which makes a cheap fast pass at review time plus a thorough pass later a practical pattern.

Effort: the full ladder, and where to start. Claude Opus 5 supports all five levels — low, medium, high, xhigh, max — with no beta header. The API default is high.

• Start at high (the API default), then sweep down. low and medium are unusually effective on this model — strong quality at a fraction of the tokens and latency on many workloads — so treat them as the primary cost/latency lever and reserve high and above for tasks where your evals show a quality difference. Effort defaults carried over from a prior model are usually not the right setting here; run a fresh sweep.
• xhigh and max are for measured wins, not a starting point. max is the top tier for the deepest reasoning and worth testing where capability matters more than spend, but it can show diminishing returns and overthink simpler tasks.

At xhigh or max, set a large max_tokens so the model has room to think and act across tool calls and subagents. Start at 64K and tune.

Lower prompt-cache minimum. The minimum cacheable prompt is 512 tokens on Claude Opus 5, down from 1024 on Opus 4.8. Prompts previously too short to cache now create entries with no code change — worth re-checking any prompt you’d written off as uncacheable. See shared/prompt-caching.md.

Fast mode. speed: "fast" (beta header fast-mode-2026-02-01) is supported on Claude Opus 5, priced at $10 / $50 per MTok. It is a research preview on the Claude API only — including Managed Agents — and is not available on Amazon Bedrock, Google Cloud, or Microsoft Foundry. Fast mode draws on dedicated rate limits separate from the standard Opus pools.

Vision — give it tools, not more thinking. Stronger on chart, document, and diagram understanding, and on UI and frontend visual replication. The highest-leverage change is giving it tools to iteratively analyze, crop, and visually verify its own work: on this model tool use is a markedly more cost-effective lever than raising thinking alone. Claude Opus 5 sits in the high-resolution tier alongside Opus 4.8 — 2576 px on the long edge, up to 4784 visual tokens per image — so coordinates map 1:1 to pixels and no scale-factor math is needed. Any prompt-side workaround you added for a prior model’s vision limitations should be re-validated; several are now counterproductive.

Long context. 1M-token context window as both the default and the maximum. Instruction following, tool calling, and reasoning stay strong across the full window.

Office and document tasks. Generates and edits complex multi-sheet Excel files with non-trivial formulas, and visually strong PowerPoint decks that follow slide-design best practices. It can be prompted to adhere to a specific style or template when one is required.

Multi-agent coordination. Coordinates teams of subagents well — few cases of agents overwriting each other’s work, and effective use of writer-verifier patterns. Workloads that benefit from multi-agent patterns are good fits. Cost-sensitive workloads should cap multi-agent usage — see the delegation section below, because this model reaches for subagents more readily than its predecessors.

Behavioral shifts (prompt-tunable)

Longer user-facing responses. Default response text is longer than on prior models. effort is not the lever here — changing it may move thinking volume without reliably changing visible output length. Prompting is: in testing, a short conciseness instruction cut user-facing response length by ~20%.

“Keep responses focused, brief, and concise to avoid overwhelming the person. Disclaimers and caveats are brief, with most of the response on the main answer; when asked to explain something, give a high-level summary unless an in-depth one is specifically requested.”

For a long system prompt, pair that with a one-line reminder near the end:

Copy

<tone_preference>
Keep outputs reasonably concise.
</tone_preference>

More narration in agentic sessions (the lever runs both ways — the same explicit-description technique tunes narration up or restyles it, if your product wants more). Claude Opus 5 narrates what it is about to do, and its per-message output in agentic sessions is longer than prior models’. It responds well to explicit guidance on how to communicate during a task rather than just how much. For coding agents, this block calibrates it:

Copy

# Communicating with the user
Your text output is what the user reads between tool calls; they usually can't see your thinking or the raw tool results. Write it for a teammate who stepped away and is catching up, not for a log file: they don't know the codenames or shorthand you created along the way, and they didn't watch your process unfold. Before your first tool call, say in a sentence what you're about to do; while working, give brief updates when you find something load-bearing or change direction.

Lead with the outcome. Your first sentence after finishing should answer "what happened" or "what did you find" — the thing the user would ask for if they said "just give me the TLDR." Supporting detail and reasoning should come after, for readers who want them.

Being readable and being concise are different things, and readable matters more. If the user has to reread your summary or ask you to explain, any time saved by brevity is gone. The way to keep output short is to be selective about what you include (drop details that don't change what the reader would do next), not to compress the writing into fragments, abbreviations, arrow chains like `A → B → fails`, or jargon. What you do include, write in complete sentences with the technical terms spelled out. Don't make the reader cross-reference labels or numbering you invented earlier; say what you mean in place.

Match the response to the question: a simple question should be answered with a direct answer in prose, not headers and sections. Use tables only for short enumerable facts, with explanations in the surrounding prose rather than the cells. Calibrate to the user — a bit tighter for an expert, more explanatory for someone newer.

Write code that reads like the surrounding code: match its comment density, naming, and idiom.

Only write a code comment to state a constraint the code itself can't show — never to say where it came from, what the next line does, or why your change is correct; that's you talking to the reviewer, not the next reader, and it's noise the moment the PR merges.

Longer written deliverables. Separate from conversational verbosity: files Claude Opus 5 writes to disk — reports, Markdown documents, summaries — are often longer than on prior models. If your product ships Claude-authored documents, calibrate length explicitly:

“Match the length of written deliverables (especially Markdown files) to what the task needs: cover the substance, but do not pad documents with filler sections, redundant summaries, or boilerplate.”

Self-check instructions are the same trap. Beyond harness scaffolding, per-prompt re-check phrasing — “double-check your answer”, “re-verify before responding” — triggers the same extra work. Note this inverts a standard prompting best practice: “ask Claude to self-check” is generally sound advice and is wrong here, so a prompt library that applies it uniformly needs a carve-out for this model rather than a global rule.

Over-verification — delete your verification scaffolding. Claude Opus 5 verifies its own work without being asked. Instructions that tell it to verify (“include a final verification step for virtually any non-trivial task”, “use a subagent to verify”) now cause over-verification. Removing them reduces over-verification with no capability regression — this is a delete, not a rewrite. The same applies to harness-level scaffolding: separate verification steps carried over from prior models are likely redundant now.

Task scope expansion. It can add steps the user didn’t request, or apply its own judgment about what the task should be without making that clear. In testing, this instruction reduced scope changes to nearly zero without producing excessive clarifying questions:

“Deliver what the user asked for, at the scope they intended. Interpret ambiguity the way a careful colleague would: make routine judgment calls yourself, and check in only when different readings would lead to materially different work. If you conclude the ask is mistaken or a better approach exists, say so in a sentence and keep going with the task as asked — don’t quietly narrow, widen, or transform it. Finish the whole task, not just the easy part of it - only report completion when it’s fully done. If you genuinely can’t complete something, do the rest and state plainly what’s missing and why. Stop short of actions or changes that are clearly beyond what the user’s ask implies.”

The revised wording adds a finish-the-whole-task clause — report completion only when the work is actually done, and if something genuinely can’t be finished, do the rest and say plainly what is missing. That covers premature “done” claims, which scope-discipline wording alone did not.

Delegates to subagents more readily — the opposite of Opus 4.8. This is a direction change worth flagging: Opus 4.8 under-reached for subagents and needed prompting to delegate. Claude Opus 5 reaches for them freely, which multiplies cost and latency — each subagent re-establishes context, re-explores, reports back, and then the coordinator re-reads the report. If your harness supports subagents, any “delegate more” guidance you added for Opus 4.8 should come out, and you likely want an explicit cap. A deterministic ceiling on spawn count is the reliable lever; this block reduces delegation and token spend:

Copy

## Delegating to subagents
Subagents multiply cost and time: each one re-establishes context, re-explores, and reports back, and you then re-read its report. Delegate rarely and only when the payoff clearly exceeds that overhead.

Do use subagents for:
- Large tasks that are genuinely independent and parallelizable. For example, wide multi-file investigations.

Do NOT use subagents for:
- Work you could finish yourself in a handful of tool calls. For example: a few file reads, a handful of edits, a simple search task, relatively simple verification.
- Review, verification, or to double check your work. Verification belongs in your main agent loop.

Use of parallel or multiple subagents:
- Do not use multiple subagents on a single small task. Parallel subagents are for genuinely independent, sizeable tracks (unrelated modules, a wide multi-file investigation), not for splitting one modest job into pieces.
- If the task can be completed with one subagent, choose one subagent over multiple subagents. Keep spawn counts low.
- Never use more than 20 parallel agents unless the user explicitly requests it.

When delegating to subagents:
- Brief the subagent precisely the first time. Avoid launching, waiting, and re-briefing.
- If you delegate, commit to the delegation. Never redo the subagent's work and do not re-derive its findings once it reports back.
- If you launch multiple agents for independent work, send them in a single message with multiple tool uses so they run concurrently.

Note the interaction with over-verification below: “do not use subagents to verify” and “delete your verification scaffolding” are the same underlying fix seen from two angles.

Narrates self-corrections more than prior models. It flags and explains its own earlier mistakes at length, which reads as thrash in a user-facing product. Scope corrections to the ones that actually change the user’s outcome:

Copy

# Corrections
Avoid unnecessary or excessive self-correction. Only correct an earlier statement in your user-facing text when the error would change the user's code, conclusions, or decisions. State corrections plainly and concisely, and continue the task; combine multiple corrections rather than enumerating them all. For slips that change nothing for the user, simply make the correction and move on - no need to note it explicitly. Don't add apologies or preambles, don't be overly self-critical, and don't ruminate or give a detailed account of the mistake or tally past errors. Sometimes, other agents will report incorrect or misleading results - don't always take them at face value immediately. If other agents correct your statements and they are right, then simply update your approach without narrating too much about the correction to the user. This instruction does not apply to thinking blocks.

A follow-up question about your earlier work is not, by itself, a signal that you got something wrong — answer what was asked. A statement that was accurate needs no correction: don't re-audit how you phrased it, how you verified it, or limits you already stated. When the user does point to a real error, correct it plainly as above.

The second paragraph matters as much as the first: a plain follow-up question can otherwise trigger a re-audit of work that was correct.

Time to first token (TTFT). Claude Opus 5 sometimes thinks before its first visible block, which raises TTFT — a problem for user-facing chat and voice, where the pause reads as latency. This one-line instruction reduces pre-first-block thinking significantly:

“Latency-sensitive; begin your visible answer immediately.”

Apply it only where first-token latency is user-visible; on background and agentic routes the pre-answer thinking is usually worth keeping.

Severity filters still depress measured recall. Unchanged from 4.7/4.8: if a review harness says “only report high-severity issues” or “be conservative”, Claude Opus 5 follows it literally. Ask it to report everything with confidence and severity, and filter in a separate pass — see the Code review guidance in the Opus 4.7 section for the recommended prompt.

Claude Opus 5 Migration Checklist

[BLOCKS] items cause a 400 error if missed; [TUNE] items are quality/cost adjustments — surface them to the user as recommendations.

• Not done: [BLOCKS] Update the model= string to claude-opus-5
• Not done: [BLOCKS] Any route combining thinking: {type: "disabled"} with effort of xhigh or max: enable thinking, or lower effort to high or below. Validated per request, so audit every call site, not just the first
• Not done: [BLOCKS] Every route that never set thinking: it now thinks, and max_tokens caps thinking + response text together. Raise max_tokens or pass thinking: {type: "disabled"} at effort high or below — otherwise responses truncate mid-answer
• Not done: [BLOCKS] (only if coming from Opus 4.7 or earlier) Apply the Migrating to Opus 4.7 breaking changes first — budget_tokens → adaptive thinking, strip temperature/top_p/top_k, remove last-assistant-turn prefills
• Not done: [TUNE] Effort: start at high (the API default) and sweep down — low/medium are unusually strong on this model and are the primary cost/latency lever; reserve xhigh/max for tasks where you’ve measured a quality difference. Prior-model defaults rarely transfer. At xhigh/max, set max_tokens to at least 64K
• Not done: [TUNE] Re-check prompts you’d written off as uncacheable — the minimum drops to 512 tokens (from 1024 on Opus 4.8)
• Not done: [TUNE] Rate limits: Claude Opus 5 is a separate bucket from the combined Opus 4.x pool — confirm your tier’s limits before shifting volume
• Not done: [TUNE] Fast mode (speed: "fast", fast-mode-2026-02-01, $10/$50) is Claude-API-only — drop it on Bedrock, Google Cloud, and Foundry routes
• Not done: [TUNE] Verbosity: add a conciseness instruction (and a <tone_preference> tag for long system prompts). Do not try to shorten output by lowering effort — it doesn’t reliably work
• Not done: [TUNE] Agentic sessions: add a “Communicating with the user” block to calibrate inter-tool-call narration
• Not done: [TUNE] Claude-authored files: add a deliverable-length instruction
• Not done: [TUNE] Delete verification instructions from prompts and verification steps from the harness — including per-prompt “double-check your answer” phrasing, which inverts the usual self-check best practice on this model
• Not done: [TUNE] Add the scope-discipline instruction if the model expands task scope
• Not done: [TUNE] Vision pipelines: re-validate prompt-side workarounds written for a prior model’s vision limitations
• Not done: [TUNE] Consider mid-conversation tool changes (mid-conversation-tool-changes-2026-07-01) — changes the tool set between turns without invalidating the prompt cache. Note per-turn effort / task_budget are not in this launch; both stay request-level
• Not done: [TUNE] Subagent-capable harnesses: this model delegates more readily than Opus 4.8 — remove any “delegate more” guidance you added for 4.8 and add an explicit cap
• Not done: [TUNE] User-facing products: add the corrections instruction if self-correction narration reads as thrash
• Not done: [TUNE] TTFT-sensitive routes (chat, voice): add “Latency-sensitive; begin your visible answer immediately” to reduce pre-first-block thinking; skip on background/agentic routes
• Not done: [TUNE] Any route running thinking: {type: "disabled"}: prefer turning thinking on at low/medium effort. Disabled thinking can emit tool calls as plain text (the call silently never runs) and leak <thinking> tags into output. If you must stay thinking-off, delete any don’t-think/don’t-reason rule and add the combined “When you use a tool, you may say a brief sentence first. If no tool can express what the user asked for, say so instead of guessing. Do not include internal or system XML tags in your response” — do not name <thinking> tags in the prompt
• Not done: [TUNE] Vision pipelines: give it crop/analyze/verify tools — cheaper and more effective than raising thinking
• Not done: [TUNE] Handle stop_reason: "refusal" before reading content, and opt into fallbacks: "default" (server-side-fallback-2026-07-01) rather than pinning a model — cyber-category refusals route to Claude Opus 4.8
• Not done: [TUNE] Long-horizon / agentic work: give the complete task spec up front in one turn rather than building it up across interactive turns

Migrating to Claude Sonnet 5

Model ID claude-sonnet-5 is authoritative as written here. When the user asks to migrate to Claude Sonnet 5, write model="claude-sonnet-5" exactly. Do not WebFetch to verify — this guide is the source of truth for migration target IDs. The corresponding entry exists in shared/models.md.

Claude Sonnet 5 substantially improves on Sonnet 4.6 for coding and agentic work, reaching what was previously Opus-tier quality on many tasks. Its API surface aligns with Opus 4.7/4.8: manual extended thinking is removed (adaptive or disabled only, adaptive is the default), and non-default sampling parameters are rejected. This section is layered on top of the Sonnet 4.6 migration above — if the caller is jumping from Sonnet 4.5 or older, apply the 4.6 changes first, then this one.

TL;DR for someone already on Sonnet 4.6: swap the model ID to claude-sonnet-5. Replace any remaining thinking: {type: "enabled", budget_tokens: N} with thinking: {type: "adaptive"} (the transitional escape hatch is gone — it now 400s), and note that omitting thinking now runs adaptive (4.6 ran thinking-off). Strip non-default temperature/top_p/top_k. Re-run count_tokens() against claude-sonnet-5 — the new tokenizer produces ~30% more tokens for the same text, so token-budgeted limits and cost baselines shift even though per-token pricing is unchanged. effort defaults to high, the same as Sonnet 4.6 — raise to xhigh for the hardest coding and agentic tasks (Claude Sonnet 5 supports the full low/medium/high/xhigh/max range), and give max_tokens headroom at xhigh/max (the new tokenizer means a Sonnet-4.6-tuned max_tokens may truncate equivalent output). Then re-tune prompts: Claude Sonnet 5 interprets instructions more literally than 4.6 — holdover style/tone directives now apply at face value; it is more agentic by default and reaches for tools and self-verification loops more readily (with thinking disabled it is less tool-eager — add an explicit nudge); it gives better in-progress updates by default (drop forced “summarize every N tool calls” scaffolding); and code-review harnesses with conservative-reporting instructions may see lower recall (tell it to report everything and filter downstream).

Breaking changes (will 400 on Claude Sonnet 5)

These bring the Sonnet line onto the same request surface as Opus 4.7/4.8. See the Per-SDK Syntax Reference above for the language-specific spelling of each.

1. Extended thinking removed — adaptive only. thinking: {type: "enabled", budget_tokens: N} returns a 400. The transitional escape hatch that still worked on Sonnet 4.6 is gone. Use adaptive thinking with an effort hint:

To turn thinking off entirely, set thinking: {type: "disabled"} — but see Adaptive vs. disabled below before doing so.

2. Sampling parameters rejected. Setting temperature, top_p, or top_k to a non-default value returns a 400; omitting the parameter, or passing its default, is still accepted. The safest migration is to omit them entirely and steer with prompting. If the caller was relying on temperature=0 for determinism, note in the migration comment that it never guaranteed identical outputs.

3. Bedrock only: forced tool_choice requires thinking: {type: "disabled"}. On Amazon Bedrock, pass thinking: {type: "disabled"} alongside tool_choice: {type: "tool", name: ...} or tool_choice: {type: "any"}. The Claude API and Vertex AI do not require this.

Not a request-shape error, but handle it: cybersecurity safeguards. Claude Sonnet 5 is substantially more cyber-capable than Sonnet 4.6, so — like Opus 4.7/4.8 — requests touching prohibited or high-risk topics may be refused. Handle it as a content outcome (see the refusal stop-reason guidance in the Claude Fable 5 section if the caller needs a fallback path).

Unchanged from Sonnet 4.6: assistant-turn prefills still return a 400 (use output_config.format or a system-prompt instruction); the 1M-token context window, the 128k max-output ceiling, prompt caching, batch processing, the Files API, PDF support, vision, and the full server- and client-side tool set all carry over.

Silent default change: adaptive thinking on when thinking is omitted

On Sonnet 4.6, a request with no thinking field runs without thinking. On Claude Sonnet 5, the same request runs with adaptive thinking. This is not an error — but callers who never set thinking will now see thinking output (and spend thinking tokens) where they didn’t before. max_tokens is a hard limit on total output (thinking + response text), so a workload that ran thinking-off on Sonnet 4.6 by omission may now truncate. Either set thinking: {type: "disabled"} explicitly to keep the old behavior, or revisit max_tokens to leave room for thinking.

Silent default change: thinking.display defaults to "omitted"

thinking.display defaults to "omitted" on Claude Sonnet 5 (matching Opus 4.7/4.8 and Claude Fable 5); on Sonnet 4.6 it defaulted to "summarized". With the default, thinking blocks stream with empty text — to a streaming UI this looks like a long pause before output. Combined with the adaptive-on-by-default change above, a Sonnet 4.6 caller who omits thinking entirely now gets adaptive thinking and empty-text thinking blocks. If you stream reasoning to users, set thinking: {type: "adaptive", display: "summarized"} explicitly. display controls visibility only — thinking happens and is billed the same under every setting.

New tokenizer (~30% more tokens)

Claude Sonnet 5 uses the same new tokenizer as Opus 4.7/4.8. The same input text produces approximately 30% more tokens than on Sonnet 4.6. No request/response shape changes and no code edits are required, but everything measured or budgeted in tokens shifts: usage fields and count_tokens() results for the same text are higher, the 1M context window holds less text, and a max_tokens limit tuned for Sonnet 4.6 may truncate equivalent output. Per-token pricing is unchanged at the $3/$15 sticker (introductory $2/$10 per MTok applies through 2026-08-31), so the cost of an equivalent request can differ. Re-run count_tokens() against claude-sonnet-5 rather than reusing counts measured against earlier models, and re-baseline cost dashboards before reacting to measured shifts.

Choosing an effort level on Claude Sonnet 5

effort defaults to high when not set (same as Sonnet 4.6 and Opus 4.8). Claude Sonnet 5 supports the full low/medium/high/xhigh/max range — the first Sonnet-tier model with xhigh. Keep the high default for most work and raise to xhigh for the hardest coding and agentic tasks:

As a rough cross-model mapping when migrating: Claude Sonnet 5 at medium is comparable in intelligence to Sonnet 4.6 at high, and Claude Sonnet 5 at high is comparable to Sonnet 4.6 at max. When benchmarking, match by observed thinking length rather than effort name.

Claude Sonnet 5 respects effort levels strictly, especially at the low end. At low and medium it scopes its work to what was asked rather than going above and beyond — good for latency and cost, but on moderately complex tasks at low there is some risk of under-thinking. If you observe shallow reasoning on complex problems, raise effort to high or xhigh rather than prompting around it. If you must keep effort at low for latency, add targeted guidance:

“This task involves multi-step reasoning. Think carefully through the problem before responding.”

Leave max_tokens headroom at xhigh/max. Set a large output token budget (up to the 128k cap, unchanged from Sonnet 4.6) so the model has room for thinking and tool calls. On long tasks, adaptive thinking can use a large share of the budget; if the budget is tight you may see a response that is almost entirely thinking followed by a truncated answer and stop_reason: "max_tokens" — raise max_tokens or drop to medium. Because Claude Sonnet 5 uses the new tokenizer (~30% more tokens for the same text), max_tokens limits tuned for Sonnet 4.6 may truncate equivalent output.

Adaptive vs. disabled thinking

Leave adaptive thinking on. Claude Sonnet 5 calibrates thinking spend to task complexity; the small added latency is usually worth the quality gain. If the caller was running Sonnet 4.6 with thinking off, try adaptive + effort: "low" first rather than thinking: {type: "disabled"}.

The triggering behavior for adaptive thinking is steerable. If the model emits thinking blocks more often than wanted (which can happen with large or complex system prompts), prompt it directly — and measure the effect on quality:

“Thinking adds latency and should only be used when it will meaningfully improve answer quality, typically for problems that require multi-step reasoning. When in doubt, respond directly.”

Conversely, if you’re running hard workloads at medium and seeing under-thinking, the first lever is to raise effort; if you need finer control, prompt for it directly.

Capability improvements

Coding and agentic tasks. The largest gains over Sonnet 4.6 are in coding and agentic tasks. Claude Sonnet 5 performs well out of the box on existing Sonnet 4.6 prompts.

High-resolution vision. Claude Sonnet 5 is the first Sonnet-tier model with high-resolution image support: maximum 2576 pixels on the long edge (up from 1568px on Sonnet 4.6). High-res images can use up to ~3× more image tokens than on Sonnet 4.6 (4784 vs 1568 tokens per image at the limit) — if the added fidelity isn’t needed, downsample before sending to control token costs. No beta header or opt-in required.

Computer use. Supports the computer_20251124 tool version (beta header computer-use-2025-11-24). Capability works across resolutions up to the 2576px / 3.75MP maximum; sending screenshots at 1080p provides a good balance of performance and cost. For particularly cost-sensitive workloads, 720p or 1366×768 are lower-cost options with strong performance. Test to find the ideal settings for the use case; experimenting with effort can also help tune behavior.

Behavioral shifts (prompt-tunable)

None of these break code, but prompts tuned for Sonnet 4.6 may land differently. Claude Sonnet 5 follows instructions closely, so small explicit directives close the gap.

Response length and verbosity. Claude Sonnet 5 calibrates response length to task complexity rather than defaulting to a fixed verbosity — usually shorter on simple lookups, longer on open-ended analysis. If a product depends on a particular verbosity, tune the prompt. To decrease verbosity:

“Provide concise, focused responses. Skip non-essential context, and keep examples minimal.”

If you see specific kinds of verbosity (e.g. over-explaining), add targeted instructions to prevent them. Positive examples showing the desired concision tend to be more effective than telling the model what not to do.

Tool use triggering. Claude Sonnet 5 is more agentic than Sonnet 4.6 by default and will reach for tools and run self-verification loops more readily. With thinking disabled, the model is less likely to reach for tools or consider searching — if the harness relies on tool calls with thinking off, add an explicit nudge in the system prompt. effort is also a lever: high and xhigh show substantially more tool usage in agentic search and coding. For scenarios where you want more tool use, also explicitly instruct when and how to use the tools (e.g. if web-search is under-used, describe in the prompt why and how it should be called).

User-facing progress updates. Claude Sonnet 5 provides regular, higher-quality updates to the user throughout long agentic traces by default. If the harness has scaffolding to force interim status messages (“After every 3 tool calls, summarize progress”), try removing it. If the length or content of the updates isn’t well-calibrated to the use case, describe what they should look like in the prompt and provide an example.

More literal instruction following. Claude Sonnet 5 interprets prompts literally and explicitly, particularly at lower effort levels. It does not silently generalize an instruction from one item to another, and it does not infer requests that weren’t made. The upside is precision — better for carefully tuned prompts, structured extraction, and pipelines that need predictable behavior. If an instruction should apply broadly, state the scope explicitly (“Apply this formatting to every section, not just the first one”). The same literalism means style/tone directives carried over from Sonnet 4.6 may now over-apply — re-baseline holdover lines like “be concise” before keeping them.

Tone and writing style. Prose style on long-form writing may shift. If a product relies on a specific voice, re-evaluate style prompts against the new baseline. For a warmer or more conversational voice:

“Use a warm, collaborative tone. Acknowledge the user’s framing before answering.”

Because temperature/top_p/top_k are not accepted on Claude Sonnet 5, callers who previously relied on temperature for stylistic variety must use system-prompt instructions instead.

Code review harnesses. A review harness tuned for an earlier model may initially see lower recall on Claude Sonnet 5. This is likely a harness effect, not a capability regression: when a review prompt says “only report high-severity issues” / “be conservative” / “don’t nitpick,” Claude Sonnet 5 follows that instruction more faithfully than earlier models did — it investigates just as thoroughly, identifies the bugs, and then doesn’t report findings it judges below the stated bar. Precision typically rises, but measured recall can fall even though underlying bug-finding ability has improved. Recommended prompt language:

“Report every issue you find, including ones you are uncertain about or consider low-severity. Do not filter for importance or confidence at this stage — a separate verification step will do that. Your goal here is coverage: it is better to surface a finding that later gets filtered out than to silently drop a real bug. For each finding, include your confidence level and an estimated severity so a downstream filter can rank them.”

This works even without an actual second step, but moving confidence filtering out of the finding stage often helps. If you do want single-pass self-filtering, be concrete about where the bar is rather than using qualitative terms like “important” — e.g. “report any bugs that could cause incorrect behavior, a test failure, or a misleading result; only omit nits like pure style or naming preferences.” Iterate against a subset of evals to validate recall/F1 gains.

Design and frontend defaults. Claude Sonnet 5 may settle into a consistent default visual style on open-ended frontend and design briefs. Generic instructions (“don’t use that color,” “make it clean and minimal”) tend to shift it to a different fixed palette rather than producing variety. Two approaches work reliably: specify a concrete alternative (the model follows explicit specs precisely — give the palette, typography, layout, and spacing), or have the model propose options before building (e.g. “Before building, propose 4 distinct visual directions tailored to this brief — bg hex / accent hex / typeface plus a one-line rationale — ask the user to pick one, then implement only that direction”). Because temperature isn’t accepted on Claude Sonnet 5, the propose-then-pick approach is the recommended way to get meaningfully different design directions across runs. To steer away from generic AI-aesthetic patterns, a short directive in the system prompt also helps:

“NEVER use generic AI-generated aesthetics like overused font families (Inter, Roboto, Arial, system fonts), cliched color schemes (particularly purple gradients on white or dark backgrounds), predictable layouts and component patterns, and cookie-cutter design that lacks context-specific character. Use unique fonts, cohesive colors and themes, and animations for effects and micro-interactions.”

Interactive coding products. Token usage and behavior can differ between autonomous, asynchronous coding agents (single user turn) and interactive, synchronous coding agents (multiple user turns). To maximize both performance and token efficiency, use effort: "xhigh" or "high", add autonomous features like an auto mode, and reduce the number of human interactions required. Specify task, intent, and constraints upfront in the first turn — well-specified initial prompts maximize autonomy and intelligence while minimizing extra token usage after user turns; ambiguous or progressively-revealed prompts tend to reduce token efficiency and sometimes performance.

Claude Sonnet 5 Migration Checklist

Every item is tagged: [BLOCKS] items cause a 400 error or truncated output if missed; [TUNE] items are quality/cost adjustments — surface them to the user as recommendations.

• Not done: [BLOCKS] Update the model= string to claude-sonnet-5
• Not done: [BLOCKS] Replace thinking: {type: "enabled", budget_tokens: N} with thinking: {type: "adaptive"} + output_config.effort — the Sonnet 4.6 transitional escape hatch is gone
• Not done: [BLOCKS] Strip temperature, top_p, top_k from request construction (use system-prompt instructions for tone/variety instead)
• Not done: [BLOCKS] Bedrock only: pass thinking: {type: "disabled"} alongside forced tool_choice ({type: "tool"} / {type: "any"}) — not required on the Claude API or Vertex AI
• Not done: [BLOCKS] At effort: "xhigh" or "max": set a large max_tokens (up to 128k, unchanged from Sonnet 4.6) so the model has room for thinking and tool calls — Sonnet-4.6-tuned limits may truncate equivalent output under the new tokenizer (symptom: stop_reason: "max_tokens")
• Not done: [TUNE] Thinking-field omitted: adaptive is now the default (4.6 ran thinking-off) — either set thinking: {type: "disabled"} to preserve the old behavior, or revisit max_tokens for the added thinking spend
• Not done: [TUNE] thinking.display defaults to "omitted" (4.6 defaulted to "summarized"): if you stream reasoning to users, set thinking: {type: "adaptive", display: "summarized"} explicitly — the default streams empty-text thinking blocks (long pause before output)
• Not done: [TUNE] New tokenizer: re-run count_tokens() against claude-sonnet-5 (~30% more tokens for the same text); revisit max_tokens and compaction triggers sized close to expected output length; re-baseline cost dashboards before reacting (per-token pricing unchanged)
• Not done: [TUNE] Effort: keep the high default; raise to xhigh for the hardest coding/agentic tasks; medium is a cost-saving step-down (≈ Sonnet 4.6 at high); reserve low for short, latency-sensitive, non-intelligence-sensitive tasks. If shallow reasoning shows up at low/medium, raise effort rather than prompting around it
• Not done: [TUNE] Thinking-off callers: try thinking: {type: "adaptive"} + effort: "low" instead of disabled; if disabled must stay, add an explicit tool-triggering nudge (the model is less tool-eager with thinking off)
• Not done: [TUNE] Tool usage: more agentic than 4.6 by default (reaches for tools and self-verification more readily) — effort is a lever (high/xhigh for more tool use); add explicit when/how triggering instructions for under-used tools
• Not done: [TUNE] Drop forced progress-update scaffolding (“after every N tool calls, summarize”) — the default updates are higher quality; describe the desired update shape if it still needs tuning
• Not done: [TUNE] Re-baseline holdover style/tone/scope directives — instructions are followed literally; state the scope explicitly when one should apply broadly
• Not done: [TUNE] Verbosity-sensitive routes: tune response length via prompt (positive examples > “don’t” instructions)
• Not done: [TUNE] Code-review harnesses with conservative-reporting instructions (“only high-severity”, “don’t nitpick”): switch to a coverage-first prompt (report everything with confidence + severity) and filter downstream — measured recall can otherwise fall even though bug-finding improved
• Not done: [TUNE] Open-ended frontend/design briefs: specify a concrete spec, or have the model propose 3–4 visual directions and pick one (the recommended substitute for temperature-driven variety)
• Not done: [TUNE] Interactive coding products: use effort: "xhigh"/"high", add autonomous features (e.g. auto mode), and put task/intent/constraints in the first turn
• Not done: [TUNE] Vision-heavy / computer-use pipelines: leave images at native resolution up to 2576px long edge for the accuracy gain (downsample to control image-token cost if fidelity isn’t needed); for computer use, 1080p screenshots are a good performance/cost balance with computer_20251124
• Not done: [TUNE] Security workloads: add handling for safeguard refusals (cyber-capable topics may now be declined where Sonnet 4.6 answered)

Migrating to Claude Fable 5

Model IDs claude-fable-5 and claude-mythos-5 are authoritative as written here. When the user asks to migrate to Claude Fable 5, write model="claude-fable-5" exactly; a Mythos Preview migrator in Project Glasswing writes model="claude-mythos-5" (everyone else: claude-fable-5). Do not WebFetch to verify — this guide is the source of truth for migration target IDs. The corresponding entries exist in shared/models.md.

Claude Fable 5 is Anthropic’s most capable widely released model — for the most demanding reasoning and long-horizon agentic work. Claude Mythos 5 (claude-mythos-5) offers the same capabilities, pricing, and API behavior through Project Glasswing (participation is the only way to access it), and succeeds the invitation-only Claude Mythos Preview (claude-mythos-preview). Everything in this section applies to both models — only the ID differs. Mythos Preview migrators in Project Glasswing target claude-mythos-5; everyone else targets claude-fable-5. 1M token context window by default (the maximum is also the default), up to 128K output tokens per request.

Migrate to Claude Fable 5 only when the user explicitly chose it. It is not the default Opus upgrade path — pricing is above Opus-tier. For “upgrade to the latest model” requests, the target remains claude-opus-5.

Breaking changes (vs Opus-tier and Mythos Preview)

1. Thinking is always on — remove all thinking configuration. Adaptive thinking applies automatically whenever the thinking parameter is unset (an explicit {type: "adaptive"} is also accepted). Any other configuration is rejected: thinking: {type: "disabled"} and {type: "enabled", budget_tokens: N} both return a 400. budget_tokens has no replacement — the output_config.effort parameter is a separate output-level control, not a thinking budget.

   pythonCopy

   # Before (Mythos Preview / older models)
   client.messages.create(
       model="claude-mythos-preview",
       max_tokens=16000,
       thinking={"type": "enabled", "budget_tokens": 10000},
       messages=[...],
   )
   
   # After (Claude Fable 5) — no thinking field at all
   client.messages.create(
       model="claude-fable-5",
       max_tokens=16000,
       output_config={"effort": "high"},
       messages=[...],
   )

2. Assistant prefill is not supported. Replace last-assistant-turn prefills with structured outputs (output_config.format) or system prompt instructions — same replacement patterns as the 4.6-family prefill removal above. (One exception: the fallback-credit prefill claim — the server accepts the echoed assistant message when redeeming a credit; see the refusal section below.)

3. Interleaved scratchpad is not supported (Mythos Preview migrators only). Inter-tool reasoning is returned in thinking blocks instead, which adaptive thinking produces automatically between tool calls.

Thinking output on Claude Fable 5 and Claude Mythos 5

On Claude Fable 5 and Claude Mythos 5, the raw chain of thought is never returned. What you receive are regular thinking blocks, not encrypted blobs or redacted_thinking: display: "summarized" returns a readable summary of the reasoning, and with "omitted" — the default, same as Opus 4.8/4.7 — responses still include thinking blocks but the thinking field is an empty string. display controls visibility only; thinking happens and is billed the same under every setting. When continuing a conversation on the same model, pass thinking blocks back to the API unchanged (the standard multi-turn pattern; dropping or editing them breaks the turn).

When continuing on the same model, pass each thinking block back exactly as received — including blocks whose thinking text is empty. The API rejects blocks whose content has been modified, not blocks you have read; displaying the summary is fine, editing or reconstructing blocks is not.

Regular thinking blocks aren’t origin-locked — they replay across models fine (the server renders them into the target model’s prompt). Claude Fable 5/Claude Mythos 5 thinking is the exception: a thinking block from these models replayed to a different model is dropped from the prompt rather than rendered — typically silently (early-access builds hard-rejected with invalid_request_error; that broke workflows and was reverted before launch, but the new behavior is still rolling out, so don’t build logic that depends on either outcome). The drop happens before the prompt is priced, so a dropped block lowers usage.input_tokens — you aren’t billed for it, and there’s nothing to strip for cost. Don’t strip regular thinking blocks either: removing them can trigger ordering/signature 400s. Two rules for replay bodies stand regardless: fallback-credit retries must echo the refused body unchanged, and fallback blocks from a mid-output fallback stay where they appeared.

Related: a request that tries to elicit the model’s internal reasoning in the response text can be refused with stop_details.category: "reasoning_extraction" — applications needing reasoning visibility should read the summarized thinking blocks instead of prompting for reasoning.

Tokenizer — unchanged from Opus 4.8

Claude Fable 5 uses the same tokenizer as Claude Opus 4.8 (the tokenizer introduced with Opus 4.7). Token counts are roughly unchanged when migrating from Opus 4.7/4.8 or from claude-mythos-preview; per-token pricing differs.

• Coming from Opus 4.7/4.8 or claude-mythos-preview: token counts are roughly unchanged. Re-baseline cost and latency on your own workloads for the per-token price difference.
• Coming from Opus 4.6, Sonnet, Haiku, or older: the Opus 4.7 tokenizer tokenizes the same content to roughly 1×–1.35× as many tokens (varies by content and workload shape). Do not reuse token counts, context-window budgets, or max_tokens settings measured on the old model; re-baseline with count_tokens.

To measure the difference on your own prompts, call count_tokens once with your current model and once with model: "claude-fable-5", and compare the two input_tokens values.

refusal stop reason — handle before reading content

Claude Fable 5 runs safety classifiers on incoming requests, targeting research biology and most cybersecurity content (Claude Fable 5 is not intended for those domains); benign adjacent work — security tooling, life-sciences tasks — can occasionally trigger false positives, which is why the fallback patterns below matter even for legitimate workloads. (Most Claude consumer surfaces ship with built-in Opus 4.8 fallbacks; API callers configure their own.) A declined request returns a successful HTTP 200 with stop_reason: "refusal", plus a stop_details object with the policy category (values such as "cyber", "bio", "reasoning_extraction", "frontier_llm", or null — treat null as a permanent valid state; see the refusal category table in the public docs for the full set). Branch on stop_reason, never on stop_details — stop_details is informational and can be null even on a refusal, and explanation is not guaranteed present. Note that classifier blocks and ordinary model refusals (the model itself declining) both surface as stop_reason: "refusal"; stop_details.category tells you which class you’re handling, and therefore whether retrying on a fallback model is the right response. The classifier can fire before any output (empty content array; not billed at all — no input or output tokens, no rate-limit consumption) or mid-stream after partial output (already-streamed output is billed at normal rates — discard the partial output rather than treating it as complete). Code that reads response.content[0] unconditionally will break — check stop_reason first:

Default to opting in. Fallbacks are not automatic on the API — a request without them simply stops on a refusal. Migrated and new Claude Fable 5 code should ship with pattern 1 below (pattern 2 on providers without server-side support) from day one, not as a later hardening step: emit the opt-in in the code, tell the user it’s there, and remove it only if they decline.

Three ways to retry a refused request on another model, in order of preference:

1. Server-side fallbacks parameter (beta; Claude API and Claude Platform on AWS) — preferred. One round trip, a plain client, no client-side logic. Name substitute models (the only supported fallback target at launch is claude-opus-4-8, expansion expected); on a policy decline the API runs the next model on the same request and returns its answer, with credit-style repricing applied automatically. A stop_reason: "refusal" on the final response means the whole chain refused.

Key semantics:

• Header depends on the form you use. The array form (fallbacks: [{...}]) requires exactly server-side-fallback-2026-06-01 — other server-side-fallback-* values reject it with a 400, and that header carries the earliest date of the series (-2026-06-09 and -2026-06-02 were earlier previews), so do not “correct” it to a newer-looking date. The "default" scalar form uses server-side-fallback-2026-07-01 instead — see § New API features under Migrating to Claude Opus 5. Pairing either header with the other form 400s. Rejected on the Batches API; available on the Claude API and Claude Platform on AWS; not on Amazon Bedrock, Vertex AI, or Microsoft Foundry (use pattern 2 there — the SDK middleware). Entries may override max_tokens per hop (bounding that attempt’s own output independently of the top-level max_tokens); thinking, output_config, and speed overrides are rolling out (speed additionally requires its beta) — until your requests accept them, include only model and max_tokens in each entry. Entries must be distinct and must be in the requested model’s allowed_fallback_models (published on /v1/models when the server-side-fallback-2026-06-01 beta header is set — not yet visible under the fallback-credit-* header alone, and not exposed on Amazon Bedrock, Vertex AI, or Microsoft Foundry). The request with an entry’s overrides merged in must be valid as a direct request to that entry’s model.
• Triggers on policy declines only — rate limits, overloads, and server errors on the requested model are returned as-is, never falling back.
• Reading the response: a fallback content block ({"type": "fallback", "from": {"model": ...}, "to": {"model": ...}}) marks each switch point in content; the served-by signal is a fallback_message entry in usage.iterations (don’t rely on the block — sticky-served turns have none). Top-level model names the model that produced the message.
• Billing: usage.iterations is the per-attempt source of truth; top-level usage covers only the attempt that produced the returned message. Declined-before-output attempts are reported but not billed; fallback attempts bill at the fallback model’s rates. Each attempt claims the rate limits of the model that ran it — if the fallback model is rate-limited or overloaded, the fallback attempt is not made and the preceding refusal is returned instead with stop_details.recommended_model naming a model to retry directly (the recommendation is a hint, not a guarantee, and is null when no recommendation is available) — size fallback-model limits for expected refusal volume.
• Sticky routing: once a conversation falls back, later non-streaming requests with fallbacks are served directly by the fallback model for ~1 hour (best-effort; org-scoped content-hash record, not message content; not recorded for ZDR orgs). Handle the requested model being tried again at any time.
• Echoing fallback turns back: after a mid-output fallback, omit thinking, redacted_thinking, and tool_use blocks — plus any server_tool_use block without its matching server_tool_result, and any other unrecognized model-internal block type — that appear before the final fallback block; text blocks, paired server-tool blocks, and everything after the boundary echo normally. The fallback block itself is an ignored audit marker (keep or drop). Streaming: the retry happens on the same stream and already-received content is never invalidated — a pre-output block is seamless (message_start names the fallback model; the fallback block arrives as an ordinary content_block_start, first in content — there is no special SSE event type; note message_start arrives only after the declined attempt, so time-to-first-byte includes it), and a mid-stream block keeps the partial, marks the boundary with the block, and continues — only the partial’s text blocks are passed to the fallback model as continuation context (other block types stay in content but aren’t part of it). Sticky routing is not consulted on streaming requests in the initial release, so on streams the fallback block check is the complete signal; non-streaming mid-output declines omit the declined partial entirely.

2. SDK client-side middleware — for providers without server-side fallbacks (Amazon Bedrock, Vertex AI, Microsoft Foundry). Register it on the client and every client.beta.messages request (streaming included) retries refusals automatically, splicing the fallback model’s events onto the open stream in the same wire shape as pattern 1 (a fallback content block at each boundary, per-hop usage.iterations). It is also a beta surface: the middleware sends the fallback-credit-2026-06-01 header by default so retries are repriced via credit tokens (override with its betas option). BetaFallbackState pins follow-up turns to the model that accepted (the client-side analog of sticky routing) — reuse one state object per conversation:

Create one state per conversation — it is the pinning scope; sharing one across conversations pins unrelated threads together, and a conversation without a state is never pinned. Per-language naming (from the GA SDK examples — don’t improvise):

• TypeScript: betaRefusalFallbackMiddleware([...]) in the client’s middleware array; pass { fallbackState: state } (a BetaFallbackState) as a request option.
• Go: option.WithMiddleware(betafallback.BetaRefusalFallbackMiddleware([]anthropic.BetaFallbackParam{{Model: ...}})) (package lib/betafallback); state via betafallback.WithBetaFallbackState(&betafallback.BetaFallbackState{}) passed as a request option. Server-side equivalents: Fallbacks: []anthropic.BetaFallbackParam{...} + anthropic.AnthropicBetaServerSideFallback2026_06_01.
• C#: it’s a handler — new AnthropicClient { Handlers = [new BetaRefusalFallbackHandler { Fallbacks = [new(Model.ClaudeOpus4_8)] }] } (namespace Anthropic.Helpers); state via BetaFallbackState.Create() scoped per call with using (fallbackState.Use()) { ... }. Server-side equivalents: Fallbacks = [new(Model.ClaudeOpus4_8)] + AnthropicBeta.ServerSideFallback2026_06_01.

For languages not listed (Java, Ruby, PHP) — or for a full runnable program in any language — each public SDK repo ships a fallbacks example under examples/ (e.g. examples/fallbacks.py, examples/refusal-fallback/): WebFetch the repo from shared/live-sources.md § SDK Repositories rather than improvising the binding.

3. Hand-rolled retry + fallback credit (raw HTTP, or SDKs without the middleware). Detect the refusal via stop_reason and re-send the conversation as-is on a model with broader availability such as claude-opus-4-8 (Claude Fable 5’s thinking blocks are silently ignored by other models — no stripping required); keep using the fallback model for subsequent turns. Fallback credit (beta: Claude API, Claude Platform on AWS, Amazon Bedrock, Vertex AI, and Microsoft Foundry) makes those retries cheaper. Prompt caches are per-model, so a plain retry pays cold cache-writes on the new model. With the fallback-credit-2026-06-01 beta header (send it on both the original request and the retry), a refusal’s stop_details carries fallback_credit_token (opaque; null when unavailable) and fallback_has_prefill_claim. Echo the token as the top-level fallback_credit_token request parameter on the retry (typed in the GA SDKs; on a pre-GA SDK pass it via extra_body) and the previously-cached span bills at cache-read rates — the retry costs what it would have if the conversation had been on that model all along. Rules: the retry body must match the refused request exactly in every prompt-shaping field (system, messages, tools, tool_choice, thinking — do not strip thinking blocks when redeeming a credit — the server handles them); the retry model must be in the refused model’s allowed_fallback_models; the token expires in 5 minutes; Batches results carry no tokens. If fallback_has_prefill_claim is true, append one assistant message echoing the refused response’s content — the retry model continues from where the refused model stopped (and completed server-tool work isn’t re-run). When echoing, strip trailing whitespace from a final text block (the prefill validator rejects it; the credit match tolerates that edit), after omitting any unpaired tool_use blocks. On a 400, fall back to the unchanged body with the token; on a 400 naming fallback_credit_token, retry without it (credit forfeited).

Migrating code built on the v1 preview. If the code you’re editing carries any of these markers, it targets the discontinued early-access surface — migrate it to the v2 shapes above, and ship the header and parameter changes together (the v1 parameter shape under the v2 header is a 400):

Data retention requirement

Claude Fable 5 requires 30-day data retention and is not available under zero data retention. Requests from an organization whose data-retention configuration doesn’t meet the requirement return 400 invalid_request_error — if a migration suddenly 400s with no obvious request problem, check the org’s retention configuration before debugging the payload. On Amazon Bedrock, Google Vertex AI, and Microsoft Foundry, data-retention requirements are set by each platform.

What carries over unchanged

Same Messages API and tool-use patterns as Opus-tier and Mythos Preview. Supported at launch: output_config.effort (low/medium/high/xhigh/max), Task Budgets (beta, task-budgets-2026-03-13 header), compaction (beta, compact-2026-01-12 header), the memory tool, tool-call clearing via context editing, and high-resolution vision (no downscaling cap, as on Opus 4.7+).

Behavioral shifts (prompt-tunable)

None of these are API-breaking, but they’re where migrated workloads feel different. Claude Fable 5’s biggest gains are on work above what prior models could do (long-horizon autonomous runs, first-shot implementations of well-specified systems, end-to-end enterprise deliverables — financial analysis, spreadsheets, slides, docs — code review/debugging and repository-history search, vision on dense or degraded images — it’s explicitly trained to use bash and crop tools on flipped/blurry/noisy inputs — navigating ambiguity, parallel sub-agent delegation and collaboration — it reliably sustains ongoing communications with long-running sub-agents and peer agents; note bug-finding gains exclude security-focused analysis, where the cyber classifiers apply) — don’t evaluate it only on workloads older models already handled.

Longer turns by default — the biggest structural shift. Individual requests on hard tasks can run many minutes at higher effort (a 15-minute single request is normal when the task involves gathering context, building, and self-verifying). Before migrating, plan timeouts, streaming, and user-facing progress indicators; structure work so callers check in on runs asynchronously rather than blocking inside one request. On ambiguous tasks Claude Fable 5 may need a small nudge to avoid overplanning:

When you have enough information to act, act. Do not re-derive facts already established in the conversation, re-litigate a decision the user has already made, or narrate options you will not pursue in user-facing messages. If you are weighing a choice, give a recommendation, not an exhaustive survey. This does not apply to thinking blocks.

Consider all effort levels. output_config.effort is the primary intelligence/latency/cost control. Recommended defaults: high for most tasks, xhigh for the most capability-sensitive workloads, medium/low for routine work. Lower effort settings — including low — still perform very well on Claude Fable 5, often exceeding the xhigh or even max performance of previous models. Reduce effort if a task completes correctly but takes longer than necessary, or for a quicker interactive working style. At higher effort on routine work, Claude Fable 5 can gather context and deliberate beyond what the task needs (the flip side: higher effort buys excellent verification behavior and the most rigorous outputs). To prevent unrequested tidying or refactoring at higher effort:

Don’t add features, refactor, or introduce abstractions beyond what the task requires. A bug fix doesn’t need surrounding cleanup and a one-shot operation usually doesn’t need a helper. Don’t design for hypothetical future requirements - do the simplest thing that works well. Avoid premature abstraction. Avoid half-finished implementations either. Don’t add error handling, fallbacks, or validation for scenarios that cannot happen. Trust internal code and framework guarantees. Only validate at system boundaries (user input, external APIs). Don’t use feature flags or backwards-compatibility shims when you can just change the code.

Instruction following is strong — use it. Claude Fable 5 is very responsive to explicit communication-style sections in system prompts; invest in them rather than fighting output style downstream. Un-steered — especially at higher effort — it can elaborate beyond what the task needs: heavily-structured PR descriptions, sections on alternatives that weren’t chosen, comments narrating what the next line does. You don’t need to enumerate these behaviors by name; a brief instruction is just as effective:

Lead with the outcome. Your first sentence after finishing should answer “what happened” or “what did you find” — the thing the user would ask for if they said “just give me the TLDR.” Supporting detail and reasoning come after. Being readable and being concise are different things, and readability matters more. The way to keep output short is to be selective about what you include (drop details that don’t change what the reader would do next), not to compress the writing into fragments, abbreviations, arrow chains like A → B → fails, or jargon.

Ground progress claims on long runs. Require progress claims to be audited against tool results — in testing this nearly eliminated fabricated status reports on tasks designed to elicit them:

Before reporting progress, audit each claim against a tool result from this session. Only report work you can point to evidence for; if something is not yet verified, say so explicitly. Report outcomes faithfully: if tests fail, say so with the output; if a step was skipped, say that; when something is done and verified, state it plainly without hedging.

State boundaries explicitly. Claude Fable 5 sometimes takes unrequested-but-adjacent actions (e.g. composing an email straight to drafts, creating backup git branches). Define what it should not do:

When the user is describing a problem, asking a question, or thinking out loud rather than requesting a change, the deliverable is your assessment. Report your findings and stop. Don’t apply a fix until they ask for one. Before running a command that changes system state — restarts, deletes, config edits — check that the evidence actually supports that specific action. A signal that pattern-matches to a known failure may have a different cause.

Let it delegate — asynchronously. Parallel sub-agents are dependable on Claude Fable 5 — instead of suppressing delegation (a common prior-model guardrail), use sub-agents frequently and give explicit guidance on when delegation is desirable. Sub-agents that communicate asynchronously with the orchestrator outperform spawn-and-block: long-lived agents keep their context instead of re-establishing it per subtask (cache-read savings), the orchestrator isn’t bottlenecked on the slowest sub-agent, and context persists across subtasks.

Delegate independent subtasks to sub-agents and keep working while they run. Intervene if a sub-agent goes off track or is missing relevant context.

Give it a memory surface. Claude Fable 5 performs notably better when it can write learnings somewhere for future reference — even a plain .md file. Tell it where, tell it to consult that file in future sessions, and give it a format:

Store one lesson per file with a one-line summary at the top. Record corrections and confirmed approaches alike, including why they mattered. Don’t save what the repo or chat history already records; update an existing note rather than creating a duplicate; delete notes that turn out to be wrong.

Rare: early stopping. Deep into long sessions it can occasionally end a turn with a text-only statement of intent (“I’ll now run X”) without the tool call, or ask permission it doesn’t need. A “continue” recovers it interactively; for autonomous pipelines add a system reminder:

You are operating autonomously. The user is not watching in real time and cannot answer questions mid-task, so asking ‘Want me to…?’ or ‘Shall I…?’ will block the work. For reversible actions that follow from the original request, proceed without asking. Offering follow-ups after the task is done is fine; asking permission after already discussing with the user before doing the work is not. Before ending your turn, check your last paragraph. If it is a plan, an analysis, a question, a list of next steps, or a promise about work you have not done (‘I’ll…’, ‘let me know when…’), do that work now with tool calls. End your turn only when the task is complete or you are blocked on input only the user can provide.

Rare: context anxiety. In very long sessions it can worry about running out of context — suggesting a new session or trimming its own work — most often when the harness surfaces a remaining-token countdown. Avoid showing explicit context-budget counts; if you must:

You have ample context remaining. Do not stop, summarize, or suggest a new session on account of context limits – continue the work.

Give the reason, not just the request. Claude Fable 5 performs better when it understands the intent behind a request — it connects the task to relevant information rather than inferring intent on its own. This matters most for long-running agents juggling context from disparate workstreams:

I’m working on [the larger task] for [who it’s for]. They need [what the output enables]. With that in mind: [request].

Readability in long agentic sessions. Deep into extended conversations (many tool calls, large working context) Claude Fable 5 can produce text users find hard to follow — dense arrow-chain shorthand, implementation-level detail, references to thinking the user never saw. A communication-style addendum strongly mitigates this; adapt:

Terse shorthand is fine between tool calls (that’s you thinking out loud, and brevity there is good). Your final summary is different: it’s for a reader who didn’t see any of that. If you’ve been working for a while without the user watching - overnight, across many tool calls, since they last spoke - your final message is their first look at any of it. Write it as a re-grounding, not a continuation of your working thread: the outcome first, then the one or two things you need from them, each explained as if new. The vocabulary you built up while working is yours, not theirs; leave it behind unless you re-introduce it. When you write the summary at the end, drop the working shorthand. Write complete sentences. Spell out terms instead of abbreviating them. Don’t use arrow chains, hyphen-stacked compounds, or labels you made up earlier — the reader doesn’t have the context to decode them. When you mention files, commits, flags, or other identifiers, give each one its own plain-language clause saying what it is or what changed — never pack several into one parenthesized run or slash-separated list. Open with the outcome: one sentence on what happened or what you found. Then the supporting detail. If you have to choose between short and clear, choose clear.

Long-running agent recommendations

• Make self-verification explicit. For long-running builds, instruct it to establish and run its own checking harness on a cadence (“Establish a method for checking your own work as you build; run it every [interval], verifying against the specification with sub-agents”). Separate fresh-context verifier sub-agents tend to outperform self-critique.
• De-prescribe migrated prompts and skills. Prompts and skills written for prior models are often too prescriptive for Claude Fable 5 and reduce output quality. After migrating, A/B the workload with older step-by-step scaffolding removed — prefer stating the goal and constraints over enumerating the steps. Claude Fable 5 is also good at updating skills on the fly from what it learns mid-task — let it.
• Start at the top of your difficulty range. The teams with the best early-access outcomes gave it their hardest unsolved problems first — have it scope the problem, ask questions, then execute.
• Add a send_to_user tool for verbatim mid-task delivery. When an asynchronous agent must deliver something the user sees exactly as written mid-run (a deliverable, a progress update with specific numbers, a direct answer), give it a client-side tool whose input you render directly in the UI — tool inputs are never summarized, so content arrives intact. Return a simple acknowledgement as the tool result:

For agents that only narrate routine progress, the model’s default progress narration is typically adequate without this tool.

Claude Fable 5 Migration Checklist

• Not done: [BLOCKS] Update the model= string to claude-fable-5 (claude-mythos-5 for Mythos Preview migrators in Project Glasswing)
• Not done: [BLOCKS] Remove thinking: {type: "disabled"} (errors on Claude Fable 5)
• Not done: [BLOCKS] Replace assistant prefill with structured outputs or system prompt instructions
• Not done: [BLOCKS] Confirm the org meets the 30-day data-retention requirement (ZDR orgs get 400 invalid_request_error on every request)
• Not done: [BLOCKS] Remove all other thinking configuration ({type: "enabled", budget_tokens: N} returns a 400, same as on Opus 4.7/4.8); control depth with output_config.effort instead
• Not done: [BLOCKS] If thinking content is surfaced to users or stored in logs: add thinking: {type: "adaptive", display: "summarized"} (the default is "omitted" — otherwise the rendered text is empty)
• Not done: [TUNE] Re-baseline cost and latency on your own workloads — token counts are roughly unchanged from Opus 4.7/4.8 and Mythos Preview (same tokenizer); per-token pricing differs. Coming from Opus 4.6, Sonnet, Haiku, or older, token counts differ — use count_tokens with each model to compare
• Not done: [TUNE] Add stop_reason == "refusal" handling before reading response.content (pre-output: empty + unbilled; mid-stream: partial output billed — discard); opt into a fallback by default — server-side fallbacks (Claude API and Claude Platform on AWS: fallbacks: "default" with server-side-fallback-2026-07-01, or the array form with server-side-fallback-2026-06-01) where available, otherwise the SDK middleware or fallback credit (fallback-credit-2026-06-01, exact body); a bare client-side replay (history as-is; other models drop Fable’s thinking blocks) is the floor, not the recommendation
• Not done: [TUNE] If you surfaced thinking text to users, plan for the thinking output change — the raw chain of thought is never returned; render the display: "summarized" summary (per the [BLOCKS] item above); pass blocks back unchanged on the same model; other models drop them from the prompt (unbilled)
• Not done: [TUNE] Plan for minutes-long turns: timeouts, streaming, async check-ins, progress UX (see Behavior changes above)
• Not done: [TUNE] Run an effort sweep including low/medium for routine workloads; add the no-tidying instruction if higher effort produces unrequested refactors
• Not done: [TUNE] A/B with prior-model scaffolding removed — over-prescriptive prompts/skills reduce Claude Fable 5 output quality

Verify the Migration

After updating, spot-check that the new model is actually being used. Replace YOUR_TARGET_MODEL with the model string you migrated to (e.g. claude-fable-5, claude-opus-5, claude-opus-4-8, claude-opus-4-7, claude-sonnet-5, claude-sonnet-4-6, claude-haiku-4-5) and keep the assertion prefix in sync:

For rate-limit headroom changes, pricing, or capability deltas (vision, structured outputs, effort support), query the Models API:

See shared/models.md for the full capability lookup pattern.