DeprecatedAI Dev Tools

Claude Sonnet 4.5 (claude-sonnet-4-5-20250929)

by Anthropic

Announced

2026-09-30

Complete

2026-11-30

Refund window closes

—

Refund status

No refund applicable

On September 30, 2026, Anthropic notified developers using Claude Sonnet 4.5 that `claude-sonnet-4-5-20250929` retires on the Claude API on November 30, 2026, and named `claude-sonnet-5-5` as the recommended replacement. After that date, requests to the model fail. The date applies 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 check those separately. The model ID swap is the easy part. Code written for Sonnet 4.5 often relies on things Sonnet 5.5 rejects with a 400 error: a prefilled assistant turn, `thinking.type: "enabled"` with `budget_tokens`, and non-default `temperature`, `top_p` or `top_k`. Sonnet 5.5 also thinks by default on requests that send no `thinking` field, which Sonnet 4.5 never did.

Refund flow

1
Not applicable. Claude API usage is billed per token in arrears, with no balance tied to a model ID, so there is nothing to refund when Sonnet 4.5 retires.
2
If you see charges for claude-sonnet-4-5-20250929 requests dated after November 30, 2026 on the Claude API, treat it as a billing error and contact Anthropic support with the request IDs.
3
Re-budget before you switch, using Anthropic's pricing page on October 7, 2026. Sonnet 4.5 is $3 per million input tokens and $15 per million output tokens. Sonnet 5.5 is $2 and $10, which is about 33% less per token. But the migration guide says Sonnet 5.5 uses a newer tokenizer that produces about 30% more tokens for the same text, and that thinking tokens are billed as output. With the price cut and the token increase alone, the same text costs roughly 13% less. Default thinking on Sonnet 5.5 can add output tokens on top of that. Measure a sample of real requests on both models before you assume a saving.

Migration path

Recommended successor

Claude Sonnet 5.5 (claude-sonnet-5-5)

Why this is the closest fit

It is the replacement Anthropic names for claude-sonnet-4-5-20250929 on its deprecations page, and Anthropic publishes a migration guide with a checklist written for each starting model, including Sonnet 4.5. Per the pricing page it costs less per token than Sonnet 4.5, and per the migration guide it has a larger context window with no beta header and a higher output limit.

What differs from the original

Per Anthropic's Sonnet 5.5 migration guide, these return a 400 error or change behavior coming from Sonnet 4.5. A prefilled last assistant turn is rejected; use structured outputs, the system prompt or the user turn instead. thinking.type "enabled" with budget_tokens is rejected; use adaptive thinking with output_config.effort. Non-default temperature, top_p and top_k must be removed. A request with no thinking field now runs adaptive thinking, so responses can start with thinking blocks and code that reads content[0].text breaks; to turn up-front thinking off send thinking type "between_tools", because "disabled" returns a 400 on Sonnet 5.5. Forced tool use is not supported; use tool_choice auto. Sonnet 4.5 has no effort parameter, so set one explicitly. Remove the interleaved-thinking-2025-05-14 and context-window beta headers, replace fine-grained-tool-streaming-2025-05-14 with eager_input_streaming, and move output_format to output_config.format. The same text is about 30% more tokens, and images can be up to 2576px on the long edge, so a 2000x1500 image costs about 2.5 times as many tokens.

Alternatives by use case

You want the replacement Anthropic recommends

claude-sonnet-5-5 at $2 / $10 per million tokens. Work through the Sonnet 4.5 section of the migration checklist and re-run your evals at two or three effort levels

You want the fewest behavior changes for now

claude-sonnet-4-6 keeps the older tokenizer and $3 / $15 pricing, runs without thinking by default, and is listed as retiring not sooner than February 17, 2027. It still rejects prefill and drops thinking budgets as deprecated, so it buys time, not a free swap

You call Sonnet 4.5 through Amazon Bedrock or Google Cloud

The November 30, 2026 date is for Anthropic-operated platforms. Check your cloud provider's own model table for its retirement date before planning around it

Your pipeline relies on assistant prefill for JSON

Move to structured outputs with output_config.format. On Amazon Bedrock structured outputs are not available for Sonnet 5.5, so describe the format in the prompt and validate it in code

What this tool meant

Sonnet 4.5 was released in late September 2025 and gets about 14 months on the Claude API before retirement, with the 61 days of notice Anthropic's 60-day minimum promises. For anyone running it in a video or content pipeline, for example to write shot lists, prompts or captions, the risk is not the model ID. It is the quiet behavior change: default thinking, a new tokenizer and rejected prefills can change output length, cost and format without any code error you would notice in testing a single call. Keep a small fixed set of real inputs with saved outputs, run them on the replacement before the cutoff, and compare token counts as well as text.