Setting the file. One moment.
Subchapter 23.23
references/model-migration.mdMarkdown4 KBView on GitHub
Migration checklist for upgrading between Claude model generations on Bedrock. Each generation introduces breaking changes that fail silently or with unclear errors.
Assets
Kb Shim Py| Change | 4.5 Behavior | 4.6 Behavior | Impact |
|---|---|---|---|
| Prefill | Supported | Hard 400 error | MUST remove all prefill before switching. Use structured outputs or system prompt instructions instead. |
| Structured outputs | output_format param | output_config.format param (old name deprecated) | Update param name, or use tool_use for structured output (works on both). On Bedrock Converse API: outputConfig.textFormat. |
| Thinking config | thinking: {type: "enabled", budget_tokens: N} | thinking: {type: "adaptive"} | Failover logic MUST swap the config (not just strip it) to maintain thinking on both sides. |
| Effort parameter | Works on Opus 4.5 only. Errors on Sonnet 4.5 and Haiku 4.5. | GA on all 4.6 models (Opus, Sonnet, Haiku) | Failover to 4.5 Sonnet/Haiku MUST strip the effort parameter. |
| Context window | 200K tokens (Sonnet 4.5 1M deprecated April 30, 2026) | 1M tokens (GA) | Prompts sized for 1M WILL fail on 4.5 failover. This is the biggest silent risk. |
| Cache thresholds | Sonnet 4.5: 1,024 tokens. Opus 4.5: 4,096. | Sonnet 4.6: 2,048 tokens. Opus 4.6: 4,096. | Content cached on 4.5 (1,024–2,047 tokens) will NOT cache on Sonnet 4.6. |
output_config.format or use tool_use for cross-generation compatibility.{type: "enabled", budget_tokens: N} to {type: "adaptive"}.us.anthropic.claude-sonnet-4-5-20250929-v1:0 to us.anthropic.claude-sonnet-4-6.Opus 4.7 is available. Key changes:
bedrock-runtime (same as 4.6). Model ID: us.anthropic.claude-opus-4-7 or global.anthropic.claude-opus-4-7.{type: "adaptive"} config as 4.6. Effort parameter works.This migration is lower-risk than 4.5 → 4.6 since the API contract is consistent. Primary concern is testing output quality and verifying quota/pricing changes.
When running multi-model routing (LiteLLM, custom AI gateways), failover between Claude generations requires config translation:
Primary: Claude Sonnet 4.6
thinking: {type: "adaptive"}
effort: "high"
output_config: {format: ...}
context_window: 1M
Fallback: Claude Sonnet 4.5
thinking: {type: "enabled", budget_tokens: 10000}
effort: STRIP (errors on Sonnet 4.5)
output_format: ... (not output_config)
context_window: 200K (truncate if needed)
prefill: must already be removedMost AI gateways (LiteLLM, custom routers) handle param translation automatically. Verify your gateway supports Claude generation-specific config mapping.
Cache keys are model-specific. Cross-generation failover ALWAYS results in a cache miss on the fallback model. This impacts both latency (cold cache on failover) and cost (cache write charges on both models).
If using failover with prompt caching, account for: