The Claude Sonnet 4.5 retirement is set for November 30, 2026 on the Claude API, and the replacement Anthropic names is claude-sonnet-5-5. Swapping the model ID is the easy part. Sonnet 4.5 accepts several request shapes that Sonnet 5.5 rejects with a 400, so a one-line change will fail in production if your code relies on any of them. This is the 4.5-specific checklist, pulled together from Anthropic's pages; it sits alongside our wider coverage in AI model releases and pricing: what changes for engineering teams.
What is retiring, and when
One model ID: claude-sonnet-4-5-20250929. Anthropic's model deprecations page lists it as Deprecated since September 30, 2026, with retirement on November 30, 2026. The same page promises "at least 60 days' notice before model retirement for publicly released models," and this one lands right at that line.
The date doesn't cover every place you can call the model. Per the deprecations page, "the dates on this page apply to Anthropic-operated platforms: the Claude API, Claude Platform on AWS, and Microsoft Foundry." Amazon Bedrock and Google Cloud "set their own retirement schedules, so a model's lifecycle status and dates can differ."
The first three rows below come from Anthropic's models overview page.
| Platform | Retirement date stated by Anthropic |
|---|---|
| Claude API | November 30, 2026 |
| Claude Platform on AWS | November 30, 2026 |
| Microsoft Foundry | November 30, 2026 |
| Amazon Bedrock | Set by the partner, not stated on Anthropic's pages |
| Google Cloud (Vertex AI) | Set by the partner, not stated on Anthropic's pages |
Anthropic's Bedrock and Vertex tables both mark Sonnet 4.5 as deprecated (anthropic.claude-sonnet-4-5-20250929-v1:0 on Bedrock, claude-sonnet-4-5@20250929 on Vertex). Neither gives a retirement date. For Bedrock, Anthropic points to Amazon's Bedrock model lifecycle (legacy) page, which covers models launched before September 7, 2026; for Vertex, to Google's Claude models page. Check those, not a blog post, for your real deadline.
What fails after the retirement date
The deprecations page puts it in five words: "Requests to retired models will fail." Anything pinned to claude-sonnet-4-5-20250929 stops working on the Claude API after November 30. Anthropic's pages don't describe an alias or fallback for a retired ID, so don't plan around one.
While you're auditing pins, look at your other 4.5-generation models too. The same table lists Haiku 4.5 as "Not sooner than October 15, 2026" and Opus 4.5 as "Not sooner than November 24, 2026." Those are floors, not announced retirement dates, so treat them as tentative.
Retirement deadlines have a habit of arriving in clusters. The Perplexity Sonar API retirement cut over the same week Anthropic announced the 4.5 deprecation, and the lesson was the same: the date matters less than finding every caller before it passes.
How to migrate from Sonnet 4.5 to Sonnet 5.5
Anthropic's Sonnet 5.5 migration guide spreads the changes across sections written mostly for people coming from newer models. These seven steps are the ones that apply when you start from 4.5.
1. Find every caller
Open the Usage page in Claude Console, click Export, and review the CSV. Anthropic describes it as "usage broken down by API key and model," so filter for claude-sonnet-4-5-20250929 and you've got your list of keys.
That export covers Claude Console usage. If you also run 4.5 through Bedrock or Vertex, check those partner consoles separately. Then grep your code, config and infrastructure-as-code for the model ID, because the CSV won't tell you which repo or service owns a key.
2. Swap the model ID
Change claude-sonnet-4-5-20250929 to claude-sonnet-5-5. Anthropic launched it on September 28, 2026, and its release notes recommend the move: "We recommend migrating to Claude Sonnet 5.5." In Claude Code, /claude-api migrate runs Anthropic's migration skill over your project. Review what it changes rather than merging blind.
3. Replace assistant prefill
If your request ends with an assistant message to steer the output, Sonnet 5.5 rejects it: "This model does not support assistant message prefill. The conversation must end with a user message." The guide notes that "Claude Sonnet 4.5, Claude Haiku 4.5, and older models accept one," which is why this code has probably worked untouched for a year.
Move the intent elsewhere. If you prefilled to force a format, use structured outputs. If you prefilled to set tone or a starting phrase, put that instruction in the system prompt or the final user message.
4. Replace thinking budgets and set effort
Requests with budget_tokens return a 400: "thinking.type.enabled" is not supported for this model. Use "thinking.type.adaptive" and "output_config.effort" to control thinking behavior. The guide says Sonnet 4.5 and Haiku 4.5 use budget_tokens for all thinking, so every extended-thinking call you have needs to change.
Sonnet 4.5 has no effort parameter, and the guide's instruction is to "set an effort level explicitly" through output_config.effort. Pick a value on purpose instead of inheriting a default.
There's a quieter change too. A request with no thinking configuration "ran without thinking" on Sonnet 4.6 and earlier, and on 5.5 the default is adaptive thinking on. Thinking output counts toward your output tokens, so re-check max_tokens for requests that used to run bare.
5. Remove sampling parameters
Non-default temperature, top_p or top_k return a 400 on Sonnet 5.5: "Remove them." Search for all three. They hide in shared client wrappers and per-tenant config more often than in the code you're looking at.
6. Clean up beta headers and formats
Six small edits, all in the migration guide:
- Remove the
interleaved-thinking-2025-05-14header, since adaptive thinking interleaves automatically. - Replace
fine-grained-tool-streaming-2025-05-14witheager_input_streaming: trueon each tool that needs it. - Replace
output_formatwithoutput_config.format. Without thestructured-outputs-2025-11-13beta header, the old parameter returns a 400. - Remove the context-window beta header, and parse tool
inputwith a standard JSON parser. - On Amazon Bedrock, move computer use from
computer_20250124tocomputer_20251124. - If you show thinking text to users, set
display: "summarized", because thinking text is omitted by default on 5.5.
7. Recount tokens and re-test
The tokenizer changed. Against Sonnet 4.5, the guide says the same text produces "about 30% more tokens, depending on the content." Anything sized to a token limit (chunking, truncation, context budgets) needs new numbers.
Images change more. Sonnet 4.5 stops at 1568 pixels and 1,568 tokens. Sonnet 5.5 takes images up to 2576 pixels and 4,784 tokens, and a 2000x1500 image costs about 2.5 times as many tokens.
Then rerun your evals at two or three effort levels, not one. Output quality and token use both move with the setting, and you're choosing it for the first time.
The migration guide also lists changes every starting model faces: forced tool_choice returning 400s, append-only thinking blocks, advisors and text between tool calls. Check each against your code. Our breakdown of the five Sonnet 5.5 breaking changes for code already on Sonnet 5 covers the overlapping ones in more detail.
What it costs to move
Don't assume the bill stays flat because the model name changed. Two of the changes above push token counts up: the tokenizer produces about 30% more tokens for the same text, and adaptive thinking is on by default. Higher per-token rates aren't required for cost per request to rise. Check current rates on Anthropic's pricing pages, then measure real requests with the new tokenizer before you forecast spend.
The image change is the sharpest example. A pipeline that sends 2000x1500 screenshots pays roughly 2.5 times the image tokens on 5.5. That's a documented consequence of the higher resolution, and it's worth deciding whether you want the detail or should downscale first.
FAQ
Can I keep calling Sonnet 4.5 on Bedrock or Vertex after 30 November?
Possibly, but Anthropic doesn't say. Its pages state that Amazon Bedrock and Google Cloud set their own retirement schedules. Both platforms mark Sonnet 4.5 as deprecated in Anthropic's tables, with no retirement date. Check Amazon's model lifecycle page and Google's Claude models page for the dates that apply to you.
What happens to requests after the retirement date?
They fail. Anthropic's deprecations page says so directly: "Requests to retired models will fail." That applies on the Anthropic-operated platforms covered by the November 30, 2026 date.
Do I need to migrate to Sonnet 5.5, or can I go to 4.6?
Anthropic recommends Sonnet 5.5. Sonnet 4.6 is listed as Active, "Not sooner than February 17, 2027," so it works as a stopgap, but that's the table's information, not Anthropic's recommendation. Sonnet 5.5 is listed "Not sooner than September 28, 2027." Going to 4.6 means doing a second migration later, and that floor is only a floor, not a promised date. 4.6 also already rejects prefill and accepts budget_tokens only as deprecated, so you'd still fix those two items now.
How do I find which keys still use Sonnet 4.5?
In Claude Console, go to the Usage page and click Export. The CSV breaks usage down by API key and model, so filter on claude-sonnet-4-5-20250929. It covers Console usage, so check partner consoles separately for Bedrock or Vertex traffic.