AI OpenAI To Bedrock

Applies to: OpenAI SDK usage detected in applications hosted on any source cloud → Amazon Bedrock

This file is loaded by design-ai.md when ai-workload-profile.json has summary.ai_source = "openai" or "both". It provides the selection policy and mapping tables for OpenAI → Bedrock migration decisions.

Facts live in references/shared/openai-on-bedrock.md — model IDs, endpoint paths, region matrix, quotas, prompt caching rules, and pricing provenance. Read it before applying this file. Do not restate those facts here.

Model lifecycle: before recommending any Bedrock model, check references/vendored/ai/ai-model-lifecycle.md. Do not recommend Legacy models as primary selections for new migrations.

Recommend defaults (Sep 2026): Claude Sonnet 5 (anthropic.claude-sonnet-5) for balanced/flagship; Claude Opus 4.8 for hardest reasoning; Claude Haiku 4.5 for cost/speed. Sonnet 5 is $2/$10 — the launch rate became the standard price on Sep 1, 2026 (the scheduled increase to $3/$15 was cancelled); comparison tables below use $2/$10. Do not default to any Claude Fable / Mythos frontier model.

Key Insight: OpenAI’s Own Models Run on Bedrock

The old premise of this guide — “OpenAI models are unavailable on AWS, so migrating means switching model families” — is obsolete. GPT-5.6 Sol / Terra / Luna, GPT-5.5, and GPT-5.4 are generally available on Bedrock, at OpenAI’s data-residency-tier rates, counting toward existing AWS commitments.

Two consequences that invert the previous decision logic:

1. When the source model is on Bedrock, keep it. A same-model move has no behavior delta to validate, no prompt re-engineering, and no eval regression risk — the migration is an endpoint and credential change, not a model change. This is the default recommendation.
2. The cost of a same-model move depends on the inference option — state it conditionally. In-Region and Geo CRIS are priced at OpenAI’s data residency tier, exactly 1.10x the standard list price, so those moves cost ~10% more. GPT-5.6 on Global CRIS is priced at the standard list price — cost parity — available only when the workload has no data-residency constraint (GPT-5.5 / GPT-5.4 have no CRIS at all). See shared/openai-on-bedrock.md. The non-cost case — AWS commitments, IAM/VPC/CloudTrail governance, residency, prompt caching, one vendor relationship — carries the in-region path; never present that path as free or neutral.

Cross-family mapping (to Claude / Nova / DeepSeek) is still the right answer in two situations: the source model has no Bedrock equivalent, or the user’s priority is cost and is willing to accept a model change to get it. Both are covered below — but neither is the default.

Selection Policy

Apply in order. Stop at the first tier that resolves.

Tier 0 — Source model is on Bedrock (default path)

If the detected model is GPT-5.6 Sol / Terra / Luna, GPT-5.5, or GPT-5.4, the target is the same model on Bedrock.

Then apply the region gate below. Record model_change: false and the chosen path in aws-design-ai.json → ai_architecture.code_migration: migration_path: "mantle_openai_responses" for the in-region mantle endpoint, or migration_path: "runtime_openai_cris" when a GPT-5.6 target is served via bedrock-runtime with a CRIS id (the model cards recommend runtime for new applications).

Report the assessment honestly: strong_migrate here means “low-risk, well-supported move”, not “cheaper”.

Region gate (applies to every Tier 0 selection — endpoint-aware)

The gate differs by family (see shared/openai-on-bedrock.md § Regional Availability):

GPT-5.6 Sol / Terra / Luna: rarely blocked. If the target region is in the mantle in-region matrix, either endpoint works. If not, the bedrock-runtime CRIS path (Geo us./in., Global global.) covers most commercial regions — offer it, stating the data-residency implication of cross-region routing (Geo stays within the geography; Global routes anywhere and is the cost-parity option). Only a workload that requires strictly in-region processing in a region outside the mantle matrix falls through to Tier 1.

GPT-5.5 / GPT-5.4: hard gate. Mantle in-region only, no CRIS, no fallback. If the target region is not in the matrix:

1. Prefer offering a region change to a supported region, if no data-residency constraint pins the workload. This preserves the zero-behavior-delta benefit, which is the whole point of Tier 0.
2. If the region is fixed, also offer upgrading within the vendor to GPT-5.6 (which reaches the region via CRIS) alongside the cross-family option — a generation change, but same vendor and prompt idioms.
3. Otherwise fall through to Tier 1 and say why explicitly — a region constraint, not a model judgment.

Record the outcome in regional_warnings[].

Tier 1 — Source model is not on Bedrock (present two options, do not pre-pick)

Applies to GPT-4o, GPT-4.1, GPT-4 / GPT-4 Turbo, GPT-3.5 Turbo, the o-series, GPT-5 / 5.1 / 5.2, and all *-Pro and *-Mini / *-Nano variants — none of which are on Bedrock.

There is no same-model target, so a model change is unavoidable. Present both options with their trade-offs and let the user choose. Do not silently default to either.

Option A tier mapping:

Option A is subject to the same region gate as Tier 0.

Tier 2 — Cost priority override

If ai_constraints.ai_priority == "cost", present the cheapest adequate Bedrock target as an explicit alternative alongside the Tier 0 / Tier 1 recommendation — not as a replacement for it. The user asked for a migration, and swapping model families is a bigger change than they may have signed up for; make the trade visible rather than deciding for them.

See “Cost-optimized cross-family alternatives” below for the numbers.

Cost-Optimized Cross-Family Alternatives

All figures per 1M tokens. Percentages are blended savings at a 2:1 input-to-output ratio. Both sides of these comparisons are now Bedrock prices where the OpenAI model is on Bedrock — this is a Bedrock-vs-Bedrock model choice, not a provider comparison. Verify rates via shared/pricing-cache.md — pricing is cache-only, so a GPT-5.x rate that is absent or _unverified_ in the cache has no live lookup to fall back to and must be reported as unavailable rather than priced (see shared/openai-on-bedrock.md).

Against Tier 0 models (same-model baseline vs cross-family)

Claude Sonnet 5 rows use $2/$10 — the launch rate, made the standard price on Sep 1, 2026 (the scheduled increase to $3/$15 was cancelled). Earlier revisions priced Sonnet 5 at $3/$15, which overstated it by a third and put several comparisons below on the wrong side.

GPT-5.6 Sol rows use $4.40/$22.00 in-region — 1.10x the $4/$20 standard rate set by the Aug 21, 2026 reduction, which AWS lists as promotional through at least Nov 21, 2026. The pre-reduction in-region rate was $5.50/$33.00; GPT-5.5 keeps that rate.

Findings worth surfacing to users:

• The Aug 21, 2026 Sol price cut flips the Sol↔Opus comparison. GPT-5.6 Sol is now ~12% cheaper blended than Claude Opus 4.8, where Opus was previously 20% cheaper — but the cut is promotional (through at least Nov 21, 2026), so do not present the saving as durable.
• GPT-5.6 Luna undercuts Claude Haiku 4.5 by ~75%. For the fast/cheap tier, the OpenAI model is the cheaper Bedrock option. Do not reflexively map a cheap OpenAI model to Haiku on cost grounds.
• Sonnet 5’s now-permanent $2/$10 undercuts every same-family GPT tier except Luna. It is ~20% cheaper than GPT-5.6 Terra and ~36% cheaper than GPT-5.4 blended; earlier revisions of this guide called Terra 19% cheaper and GPT-5.4 near parity on the cancelled $3/$15 rate. Cost now favors the cross-family move for Terra/GPT-5.4-class workloads — but it is still a model change; keep the same-model path as the default and present this as the cost alternative.
• Every row above is short-context (272K). For a >272K workload, re-run the comparison at the GPT-5.6 long-context tier (2.0x input, 1.5x output); Claude and Nova do not have an equivalent step, so the cross-family option becomes markedly cheaper.

Sources with no Bedrock equivalent (cross-family reference)

Used for Tier 1 Option B. Source prices are OpenAI’s own; Bedrock prices are the target.

For the rows where the source is still cheaper (GPT-5.1/5, GPT-4.1), note that these models are on a vendor deprecation path anyway; Option A (GPT-5.6 on Bedrock) is usually the better framing than defending a stale model on price.

Open-weight gpt-oss

openai.gpt-oss-120b (0.15 / 0.60) and openai.gpt-oss-20b (0.07 / 0.30) remain available and, unlike the proprietary GPT models, do support bedrock-runtime / Converse. They sit a capability class below the GPT-5.x frontier tier. Recommend them when the user wants OpenAI-architecture models on the Bedrock-native runtime surface (for Guardrails, invocation logging, or Converse-based tooling) rather than through mantle.

Migration Decision Framework

Migrate to Bedrock, same model (Tier 0), if: the source is GPT-5.6 / 5.5 / 5.4 and the target region carries it. This is the common case and needs no cost justification.

Migrate to Bedrock, model change (Tier 1/2), if:

• The source model is not on Bedrock (see the region and catalog lists in shared/openai-on-bedrock.md)
• The user’s priority is cost and they accept a model change — note that the cheapest option is often Nova, and that GPT-5.6 Luna already beats Claude Haiku 4.5 on price
• Bedrock-native features are required that mantle does not expose (Guardrails, Knowledge Bases, invocation logging, Converse-based tooling) — these need a Bedrock-native model or gpt-oss

Consider staying on OpenAI’s own API only if:

• The target region is fixed and carries no suitable model, and a model change is unacceptable
• The workload needs the Realtime API — no Bedrock equivalent
• The workload uses GPT-5.5’s native audio/video (omnimodal) input in ways the Bedrock deployment cannot serve — Claude on Bedrock is text+image only, and the mantle GPT path should be probed for the specific modality before committing
• The workload depends on gpt-image, Whisper, TTS, or OpenAI embeddings — these map to other AWS services (Stability AI, Transcribe, Polly, Titan Embeddings), not to an OpenAI model on Bedrock
• The workload uses Assistants API with file search, vector stores, or code interpreter — see the decision tree below
• Chat Completions cannot be reshaped to Responses and probing confirms Chat Completions is unsupported

Cost belongs on that list, modestly. Bedrock in-region runs ~10% above OpenAI’s standard tier, so a cost-only decision genuinely favours staying. Say so; then weigh it against commitments, governance, and residency rather than dismissing it. For a long-context (>272K) workload the gap is much wider — the 1M tier is 2.0x input and 1.5x output — so profile the context distribution before concluding.

Feature Migration

Common Migration Paths

Same model on Bedrock (Tier 0) — smallest possible change

The application keeps the OpenAI SDK and the same model. Changes are limited to:

1. Base URL → https://bedrock-mantle.{region}.api.aws/openai/v1 (note the openai/v1 segment)
2. Credentials → Bedrock API key or the auto-refreshing BedrockOpenAI token provider, not an OpenAI key
3. Model ID → the openai.gpt-* form
4. IAM → bedrock-mantle:* actions, e.g. via AmazonBedrockMantleInferenceAccess

Full snippet and prerequisites in shared/openai-on-bedrock.md. No prompt changes, no eval regression expected — but still run the eval harness to confirm, because infrastructure and tokenizer-adjacent behavior can differ.

Assistants API → migration decision tree

Assistants API and Responses API are different surfaces. Do not treat all Assistants usage as a config-only change.

1. App already uses the OpenAI Responses API (responses.create) → Cleanest path. Base URL + credential + model ID change.

2. App uses Assistants API only for stateful multi-turn conversation (no hosted tools, no file search, no code interpreter, no persistent Assistant objects, no complex run lifecycle) → Mantle Responses is viable. Requires migrating threads/runs to responses.create — a small API migration (days), not a redesign. Not a zero-code change.

3. App uses Assistants API with simple hosted tools (function calling only) → Mantle Responses with tool use is viable. Moderate migration (1–2 weeks) to adapt tool definitions and the run lifecycle.

4. App uses Assistants API with file search, vector stores, code interpreter, persistent Assistant objects, or complex run lifecycle management → Do not recommend mantle. Evaluate Bedrock AgentCore (Harness sessions/memory, action groups as MCP tools via Gateway, gateway-fronted knowledge bases) for a full agentic replacement (2–4 weeks), or app-managed orchestration if the team prefers to own state. Never target classic Bedrock Agents (bedrock-agent) — it is in maintenance mode and closed to new customers as of July 30, 2026.

When to prefer a Bedrock-native model over the mantle GPT path

Choose Claude / Nova via Converse when the workload needs Knowledge Bases, intelligent prompt routing, structured outputs, or application inference profiles — none of which the GPT path offers. Note what no longer forces a family switch for GPT-5.6: its bedrock-runtime CRIS path provides cross-region inference, invocation logs, and Guardrails (Converse API only; prompt caching is Responses-only on runtime). For GPT-5.5 / GPT-5.4, Guardrails, logging, and cross-region reach still require a different model. The mantle GPT path is the smallest change; the Bedrock-native path remains the most feature-complete.

High-spend tiering

Tier by task complexity: simple (60%) → Nova Micro / GPT-5.6 Luna, moderate (30%) → Nova Pro / GPT-5.6 Terra, complex (10%) → Claude Sonnet 5 / GPT-5.6 Sol. Mixing families across tiers is fine, but note that mantle and bedrock-runtime are different surfaces — a tiered router spanning both needs two client paths.

Volume-Based Recommendations

Low (<1M tokens/day): use the best model for quality. Cost difference is immaterial.

Medium (1–10M tokens/day): present the cost comparison. For Tier 0 the comparison is against the cross-family alternative, not against the source provider.

High (10–100M tokens/day): consider tiering. Enable GPT-5.6 prompt caching first — cached input is both 90% cheaper and exempt from the input-TPM quota, which relieves the binding constraint on mantle.

Very high (>100M tokens/day): tiering plus caching. Check TPM headroom per model per region early; there is no RPM quota and no bedrock-runtime fallback for GPT models, so quota increases are the scaling lever. See shared/ai-migration-guardrails.md.

OpenAI Pricing Tiers

OpenAI’s own API offers Batch (50% off, 24hr), Flex (30–50% off, higher latency), Standard (baseline), and Priority (2x, lowest latency). Source-side figures in this file use Standard. Bedrock exposes Standard / Priority / Flex / Reserved service tiers, but per-model tier support renders as empty cells in the AWS docs — verify before recommending Flex or Reserved as a cost lever.