Anthropic Deprecates Claude Sonnet 4.5: Retirement Set for November 30, 2026
If your website, app or internal tool calls Claude Sonnet 4.5 through Anthropic's API, it now has an end date. On September 30, 2026, Anthropic notified developers that the model is deprecated and gave it a retirement date of November 30, 2026, according to the company's model deprecations page.
That matters to anyone who has paid for custom software development with an AI feature inside it: a support assistant, a document summarizer, a lead-qualification flow. Once the retirement date passes, requests that still name the old model will fail rather than quietly switch to a newer one.
This article explains what Anthropic has announced, which platforms the dates apply to, what changes when you move to the recommended replacement, and how to plan the work in the roughly eight weeks available.
What Anthropic announced
The deprecations page lists one new entry dated September 30, 2026. The model with the API identifier claude-sonnet-4-5-20250929 is now in the Deprecated state, its retirement date is November 30, 2026, and the recommended replacement is claude-sonnet-5-5.
Anthropic defines its lifecycle terms on the same page. A deprecated model still works but is no longer recommended, and a retired model is no longer available, so requests to it fail. The page also warns that deprecated models are likely to be less reliable than active ones. Anthropic states that it gives at least 60 days' notice before retiring a publicly released model; the gap between September 30 and November 30 is 61 days, so this notice sits close to that minimum.
Which platforms the November 30 date covers
The date does not apply everywhere Claude is sold. Anthropic's deprecations page says its dates cover Anthropic-operated platforms: the Claude API, Claude Platform on AWS and Microsoft Foundry. Partner-operated platforms, which the page names as Amazon Bedrock and Google Cloud, set their own retirement schedules, so the same model can have a different status and date there.
If your application reaches Claude through Bedrock or Google Cloud, check that provider's model table rather than assuming November 30. If you are not sure which route your application uses, that is the first question to put to whoever built it.
What changes when you move to Claude Sonnet 5.5
Moving to the replacement is more than editing one model name. Anthropic's migration guide for Claude Sonnet 5.5 lists several settings that Sonnet 4.5 accepts and Sonnet 5.5 rejects with an error.
Settings that now return errors
According to the guide, the following all return a 400 error on Claude Sonnet 5.5:
- Thinking budgets set with the older "enabled" thinking type and a token budget. The guide says to replace these with an effort level.
- Non-default values for the sampling parameters temperature, top_p and top_k. The guide says to remove them.
- Assistant prefill, where the developer writes the first words of the model's reply to steer its format. The guide suggests structured outputs or clearer instructions instead.
Prefill deserves attention because it was a common way to force a model to answer in JSON or to skip a preamble. An application that depends on it will stop working until that part is rewritten.
Behavior that changes without an error
The guide also says that thinking runs by default on Sonnet 5.5, whereas it is off by default on Sonnet 4.5, and it advises teams to revisit their maximum output token setting as a result. Code that reads the response should read content blocks by type rather than assuming the first block is the answer text. The guide adds that Sonnet 4.5 has no effort parameter, so teams should set an effort level explicitly on the new model and re-test.
What happens to cost
The price per token goes down, but the number of tokens goes up, so the net effect depends on your workload. Anthropic's pricing page lists Claude Sonnet 4.5 at USD 3 per million input tokens and USD 15 per million output tokens, and Claude Sonnet 5.5 at USD 2 and USD 10.
The same page explains that newer Claude models use a different tokenizer that produces approximately 30 percent more tokens for the same text, with the exact increase depending on content. The migration guide repeats that figure for a move from Sonnet 4.5 to Sonnet 5.5 and tells teams to recount tokens and revisit cost. It also notes that Sonnet 5.5 processes images at a higher resolution, which uses more tokens per image; its example is a 2000 by 1500 pixel image costing about 2.5 times as many tokens.
Add default thinking to that and a simple "it is cheaper" conclusion is not safe. Measure a representative sample of your own requests on both models before you forecast a monthly bill.
What this means for your business
The practical lesson is that an AI feature is not a one-time build. The model behind it is a dependency with a published shelf life, much like a PHP version or a mobile operating system. Anthropic's own table shows the pattern: Claude Sonnet 4 and Claude Opus 4 were deprecated on April 14, 2026 and retired on June 15, 2026, and Claude Opus 4.1 was retired on August 5, 2026.
The same table lists dates before which currently active models will not be retired. For Claude Haiku 4.5 that date is October 15, 2026, and for Claude Opus 4.5 it is November 24, 2026. Neither model has been deprecated, and those dates are floors rather than announcements. Still, a business running either model would be sensible to treat a deprecation notice as likely at some point and to budget for it.
There is also a quality question. A newer model will not behave identically. Answers may be longer or shorter, formatted differently, or more cautious in some areas; the migration guide tells teams to handle refusals and configure a fallback. For a customer-facing assistant, that means someone needs to review real outputs before the switch goes live, not after.
What to do next
A sensible plan for the next eight weeks looks like this:
- Find out where Claude Sonnet 4.5 is used. Anthropic's deprecations page describes exporting a CSV from the Usage page in the Claude Console that breaks usage down by API key and model.
- Confirm the route: Claude API, Claude Platform on AWS, Microsoft Foundry, Amazon Bedrock or Google Cloud. The deadline depends on it.
- Ask your developers to work through the Sonnet 4.5 section of the migration checklist and list which rejected settings your code uses.
- Run your existing test prompts against Sonnet 5.5, compare output quality and token counts, and adjust prompts and limits.
- Release the change well before November 30, with a way to roll back while the old model still answers.
If you have no test prompts, build a small set now from real, anonymized requests. It is the cheapest insurance against the next retirement as well as this one. Anthropic's guide also mentions a migration command in its Claude Code tool that applies the model change and produces a list of items to verify by hand, which can shorten the developer work but does not replace testing.
Questions to ask your development partner
If an agency or freelancer built your AI feature, three questions will tell you how exposed you are. Is the model name stored in configuration or scattered through the code? Is there an automated test set that scores outputs, or does someone check by eye? And who is responsible for watching provider deprecation pages, under what support agreement?
Clear answers mean the migration is routine maintenance. Vague answers suggest the feature was built as a demo and never hardened for production, which is worth fixing while there is still time.
Conclusion
Anthropic deprecated Claude Sonnet 4.5 on September 30, 2026 and will retire it on its own platforms on November 30, 2026, with Claude Sonnet 5.5 as the recommended replacement. The move involves real code changes, a different token count and behavior that needs testing, so it should be scheduled as a small project rather than left to the final week.
If you need help auditing an existing AI integration or planning the migration, you can contact Entrant Technologies to talk through your setup.