Claude Sonnet 4.5 retires on 30 November 2026: what to switch to and how to test it
Claude Sonnet 4.5 (claude-sonnet-4-5-20250929) retires on the Claude API on 30 November 2026. Anthropic's suggested replacement is Claude Sonnet 5.5 (claude-sonnet-5-5). On our 27-task test, Sonnet 5.5 solved as many tasks as Opus 5.5 at $0.152 per solved task. Switching is not a one-line change: re-test your own tasks before the date.
The dates, by platform
| Platform | Model ID | Retires | Suggested replacement | Source |
|---|---|---|---|---|
| Claude API (Anthropic)claude-sonnet-4-5-20250929 | claude-sonnet-4-5-20250929 | 30 Nov 2026 | claude-sonnet-5-5 | Anthropic |
| AWS Bedrockanthropic.claude-sonnet-4-5-20250929-v1:0 | anthropic.claude-sonnet-4-5-20250929-v1:0 | 08 Apr 2027 | not named | AWS |
Read from each vendor's deprecation page on 11 November 2026. Every other retiring model, with its source, is in our model retirement calendar.
What changes when you switch
- Sampling parameters can break requests. Claude Sonnet 5.5 rejects a non-default temperature, top_p or top_k with an HTTP 400 error. Code that sets temperature and only swaps the model ID will fail on the first call (agno issue #11140). Remove these parameters or leave them at the default.
- Framework defaults point at the retiring model. LangChain.js ChatAnthropic and Agno Claude() default to claude-sonnet-4-5-20250929. If you never set a model, you are on Sonnet 4.5 and those requests stop working on 30 November (langchainjs #11892, agno #11140). Set the model ID explicitly.
- Prompts tuned to the old model drift. A replacement is a different model. Answers that look fine can still drop a tool call, a field or an edge case. Only a run on your own tasks with known answers shows which.
- Cost per task changes. Token prices, answer length and retries all move. Measure cost per finished task, not price per token.
The candidates, measured
We ran the open-source coding agent Aider on 27 programming tasks with hidden tests, changing only the model. Same tasks, same agent, one retry with the test output. Cost is tokens × list price.
| Model | Model ID | Solved (of 27) | Solved first try | Cost per solved task | vs Opus |
|---|---|---|---|---|---|
| Claude Opus 5.5claude-opus-5-5 | claude-opus-5-5 | 26 | 26 | $0.297 | — |
| Claude Sonnet 5.5claude-sonnet-5-5 | claude-sonnet-5-5 | 26 | 25 | $0.152 | −49% |
| Claude Haiku 5.5claude-haiku-5-5 | claude-haiku-5-5 | 22 | 15 | $0.009 | −97% |
- Sonnet 5.5 is the default switch. It is the named replacement, and it matched Opus 5.5 task for task at 49% lower cost per solved task.
- Haiku 5.5 is the cheap challenger. 16 times cheaper per solved task than Sonnet 5.5, but it failed 4 more tasks, all on edge cases and error handling. Test it where a checker catches its misses.
- Opus 5.5 is for the cases Sonnet fails. It costs about twice as much per solved task. Pay for it only where your own tests show Sonnet 5.5 missing a critical case.
One public benchmark; your tasks decide. That is what the pilot measures. Full method and raw data: Aider cost vs quality report and Haiku vs Sonnet vs Opus 5.5.
Re-test before 30 November: 6 steps
- Freeze 20–30 real tasks with expected answers. Take them from production, include the 3–5 cases that must never go wrong, and write the answers down before any run.
- Run the old model and the candidates on the same set. Sonnet 4.5 today, then Sonnet 5.5, Haiku 5.5 and, if you use it, Opus 5.5. Same prompts, same tools, same grading.
- Compare cost per finished task and the named failures. Count retries and failed tasks in the cost. List exactly which tasks each candidate gets wrong.
- Fix prompts and parameters. Remove non-default temperature, top_p and top_k for Sonnet 5.5, set the model ID explicitly, and adjust prompts where a candidate drifts.
- Re-run the frozen set. Confirm the fixes hold and nothing else broke. Keep the set for the next retirement.
- Switch before the date. Move production to the winner before 30 November 2026, not on the day the old ID stops answering.
Re-test before 30 November — $750 per event
Send us 20–30 real tasks with expected answers. We run Sonnet 4.5 against Sonnet 5.5, Haiku 5.5 and Opus 5.5, and report cost per finished task and every task each setup gets wrong. First time? The pilot is free.
Questions people ask
When does Claude Sonnet 4.5 retire?
Claude Sonnet 4.5 (claude-sonnet-4-5-20250929) retires on the Claude API on 30 November 2026, per Anthropic's model deprecations page. On AWS Bedrock, anthropic.claude-sonnet-4-5-20250929-v1:0 retires on 8 April 2027, per the AWS Bedrock legacy models page.
Claude Sonnet 4.5 retires on 30 November. What should I switch to?
Anthropic's suggested replacement is Claude Sonnet 5.5 (claude-sonnet-5-5). On our 27-task Aider test it solved 26 tasks, the same as Claude Opus 5.5, at $0.152 per solved task versus $0.297. Claude Haiku 5.5 cost $0.009 per solved task but solved 22. Run your own tasks on Sonnet 5.5 and one cheaper option before you switch.
Is Sonnet 5.5 a drop-in replacement for Sonnet 4.5?
No. The model ID changes, Sonnet 5.5 returns an HTTP 400 error for a non-default temperature, top_p or top_k, and prompts tuned to Sonnet 4.5 can behave differently. Change the ID, remove those parameters, and re-test your tasks with known answers.
Why does Sonnet 5.5 return a 400 error when I set temperature?
Claude Sonnet 5.5 rejects non-default temperature, top_p and top_k values with an HTTP 400 error (agno issue #11140). Remove these parameters from requests to Sonnet 5.5 or leave them at the default.
Can I keep using Sonnet 4.5 on AWS Bedrock after 30 November?
Yes, until 8 April 2027: that is the retirement date for anthropic.claude-sonnet-4-5-20250929-v1:0 on the AWS Bedrock legacy models page. It buys time, not a fix. The same re-test is due before the Bedrock date.
How do I test a model switch before the deadline?
Freeze 20–30 real tasks with expected answers, run the old model and the candidates on the same set, compare cost per finished task and the named failures, fix prompts and parameters, re-run, and switch before the date. We run this for you as a free pilot, then $750 per re-test event.
Questions or corrections: hello@agenttrustapi.com. Every public number here comes from our Aider report and its raw data.