Claude 5.5 Migration: API Changes and Sonnet 4.5 Retirement

Four request patterns need attention

Anthropic’s September 22 release notes document four compatibility boundaries for Opus 5.5. These are not theoretical quality differences. They are request-shape changes that can stop a call before the model answers.

There is one platform qualification: Anthropic says the older computer-use tool continues to work on Amazon Bedrock. That makes provider part of the migration record, not a detail to omit from the test log.

Adaptive thinking is now the operating model

Opus 5.5 uses always-on adaptive thinking. The model decides when deeper reasoning is useful, while the caller sets an effort level. Anthropic lists medium as the default effort. A team that previously used token budgets should not translate the old number into a guessed effort value and call the job complete.

Replay representative work at low, medium and high effort. Compare task completion, tool errors, latency, output length and cost. Save the complete request envelope with each result so a later regression can be traced to the model, effort, tool set or provider.

The release also expands the working envelope

Anthropic documents a one-million-token context window by default and a maximum output of 128,000 tokens. Pricing is listed at $4 per million input tokens and $20 per million output tokens, below the $5 and $25 rates shown for Opus 5.

Availability spans the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry. The surrounding feature set is not identical everywhere. Fast mode is a Claude API preview, while computer-tool compatibility already differs by provider.

Two September 15 beta headers are also relevant to tool-heavy systems. inline-tools-2026-09-15 covers inline tools, and mcp-client-2026-09-15 covers MCP toolsets. A response can include mcp_tool_listing, so parsers that assume every content block is prose or a familiar tool event deserve a fixture test.

A migration sequence that preserves a rollback

1. Inventory every production request field, beta header and tool definition sent to the current model.
2. Split test cases by provider because compatibility is not fully uniform.
3. Remove manual thinking controls and replace them with an explicit effort policy.
4. Replace forced tool selection with automatic selection plus strict input schemas.
5. Update the computer tool where the provider requires it.
6. Teach the response parser to accept new tool-listing and thinking content blocks.
7. Replay read-only traffic, then low-risk writes, before enabling consequential tool actions.
8. Keep the old model route available until error rate, task success and latency are stable.

Do not use one chat prompt as the acceptance test. The breaking changes sit in the API contract, so the test set must include the requests that exercise tools, long context, streaming, retries and provider-specific adapters.

Check where progress text appears

Anthropic notes that progress text may appear inside thinking blocks unless the request uses the relevant thinking.display behavior. Applications that show status messages to users, hide private reasoning content, or transform streamed blocks should verify the display path deliberately.

A robust renderer distinguishes user-visible progress, tool events, final answer text and content that should not be exposed. Treating every incoming text fragment as final prose can create duplicated or misplaced interface copy.

What teams should do today

Create a branch that changes the model and request contract together. Run it against production-shaped fixtures, not a generic benchmark, and record failures by exact request field. The new price and larger context window can be useful, but they do not compensate for a migration that silently drops tools or fails with a 400 response.

For a broader view of search-capable model testing, use our citation-quality test design. It applies the same principle: a headline model result does not replace an evaluation of the behavior your product actually depends on.