Claude Sonnet 4.5 retires on November 30, and API version 2026-10-21 rejects temperature: a regression plan

Two Anthropic changes land close together, and they interact.

First, Sonnet 4.5 retires. Anthropic retires claude-sonnet-4-5-20250929 on November 30, 2026 (announced September 30). The replacement is claude-sonnet-5-5. Azure AI Foundry retires claude-sonnet-4-5 the same day. AWS Bedrock is different: Sonnet 4.5 moves to extended access at higher prices on January 8, 2027, and reaches end of life on April 8, 2027.

Second, sampling parameters are gone. API version 2026-10-21 removes temperature, top_p and top_k. Requests that set them get a 400 error. The new SDK major versions (Python 2.0.0 and TypeScript 1.0.0, both with breaking changes) default to this API version.

Why they collide

To move to claude-sonnet-5-5, many teams will upgrade the SDK. The upgrade switches the API version, and every request that still sets temperature fails. A model migration turns into an outage, and the error message points at a parameter, not the upgrade.

Claude Opus 4.5 and Haiku 4.5 have a similar split. On Azure they retire on November 24 and November 15. On Anthropic’s own API both are still active, with “not sooner than” dates and no formal deprecation yet.

A regression plan

1. Find every Claude call and every sampling parameter. Search your code, and also check the wrappers, gateway configs and framework settings that might add temperature, top_p or top_k without your code mentioning them.
2. Pin the SDK major version. Upgrade it on purpose, in the same change as the parameter removal, not as a side effect of a routine dependency update.
3. Capture your replay set before November 30. Take 50 real prompts with Sonnet 4.5’s current outputs. After the retirement you can’t generate the baseline.
4. Replace what temperature was doing. If you used temperature: 0 for repeatable output, get repeatability another way: tighter instructions, output schemas, and tests that check the parts that must not vary.
5. Swap the model and replay. Run the 50 prompts on claude-sonnet-5-5 and compare quality, latency and cost per call. Look hardest at tool calls and structured outputs.
6. Bedrock users: price the wait. Decide whether to pay extended-access prices from January 8 or migrate before then. Doing nothing is a price increase.

The exposure check flags Claude model IDs and any sampling parameters in pasted code, entirely in your browser. The migration sprint does the inventory, the change and the 50-prompt replay in five business days.