Claude Sonnet 5.5 Ships With Breaking API Changes, and Sonnet 4.5 Now Retires November 30

Claude Sonnet 5.5 shipped on September 28, 2026, and Anthropic says code written for Sonnet 5 can break on it in five ways. Two days later, Anthropic set a retirement date for Claude Sonnet 4.5: November 30, 2026.

Those two announcements belong together. Anyone still running Sonnet 4.5 now has two months to move, and the model Anthropic recommends moving to is Sonnet 5.5. Coming from 4.5, the list of things that break is longer than five.

The short version: Sonnet 5.5 costs the same as Sonnet 5, $2 per million input tokens and $10 output, which is a third less per token than Sonnet 4.5. But it rejects several request patterns that older code relies on, turns thinking on by default, and uses about 30% more tokens than Sonnet 4.5 for the same text. Test before you switch, and start now if you are on 4.5.

This post covers what shipped, the five breaks from Sonnet 5, the extra breaks from Sonnet 4.5, and a checklist. It follows the same pattern as our Claude Opus 5.5 migration guide from last week, and many of the changes are the same.

What shipped with Claude Sonnet 5.5

Anthropic’s Sonnet 5.5 announcement says the model is more than 30% faster than Sonnet 5 and up to 30% cheaper per task. It is available in the Claude apps and through the API.

Prices come from Anthropic’s pricing page, and they match Sonnet 5 exactly. Anthropic’s benchmark figures are its own, so treat them as a starting point. We covered the model this replaces in our Claude Sonnet 5 launch post.

Two migrations, not one

Anthropic’s platform release notes are written for teams moving from Sonnet 5. If that is you, five changes matter.

The model deprecations page adds a second group. It lists claude-sonnet-4-5-20250929 as deprecated on September 30, retiring November 30, with Sonnet 5.5 as the replacement. Teams on Sonnet 4.5 skip a generation and inherit every change since.

For agencies, this second group is the bigger risk. Sonnet 4.5 sits inside a lot of quietly working tools: WordPress AI plugins, support chatbots, content tagging scripts and form processors. Many of those were set up once and never touched.

Break 1: thinking cannot be turned off the old way

On Sonnet 5 you could send thinking: {"type": "disabled"}. On Sonnet 5.5, that returns a 400 error. The replacement is a new setting:

thinking: {"type": "between_tools"}

With between_tools, the model does not think before responding. The short notes it writes between tool calls come back as thinking blocks.

There are limits. Anthropic’s migration guide says between_tools works only at low, medium and high effort. At xhigh or max, it returns a 400. You also cannot change effort partway through a conversation in this mode.

Who hits this: anyone who turned thinking off to save cost on simple, high-volume work, such as classifying tickets or writing alt text.

Break 2: forced tool use is gone

tool_choice: {"type": "any"} and tool_choice: {"type": "tool", "name": "..."} now return a 400 error. Only auto and none are supported.

Anthropic’s fix is to send tool_choice: {"type": "auto"}, mark your tools strict: true, and say in the prompt which tool to use. On Amazon Bedrock, where strict mode is not available, use auto and check the tool input in your own code.

Who hits this: form processors and data-extraction scripts that forced a single tool to get structured output every time. Those are common in CMS integrations.

Break 3: thinking blocks are tied to the account

Thinking blocks produced by Sonnet 5.5 now work only in the account that produced them, or an account linked to it. The release notes say that if another account sends those blocks, the API drops them before the model sees them, and the request still succeeds.

The migration guide adds a stricter rule for accounts created after August 31, 2026. Editing earlier history and replaying a block returns a 400 error.

Who hits this: agencies that store conversations under one API key and replay them under another. Keep conversation history append-only, and use mid-conversation system messages to change instructions. We explained that pattern in our post on mid-conversation system messages.

Break 4: the old computer use tool is rejected

On the Claude API and Google Cloud, Sonnet 5.5 rejects computer_20251124. You need the new computer_toolset_20260801 toolset.

Amazon Bedrock is different. There, Sonnet 5.5 still uses computer_20251124. If you run the same agent on two platforms, you now need two configurations.

Anthropic also says to drop the fine-grained-tool-streaming-2025-05-14 beta header and set eager_input_streaming: true on each tool instead.

Break 5: the advisor tool needs a newer advisor

The advisor tool lets one model ask another for help. On Sonnet 5.5, it rejects Opus 4.8, Opus 4.7 and Sonnet 5 as advisors. The migration guide also lists Opus 4.6 and Sonnet 4.6.

Supported advisors include Opus 5, Opus 5.5, Sonnet 5.5, and the Fable and Mythos 5 models. Advice now comes back encrypted, in an advisor_redacted_result block you cannot read.

Coming from Sonnet 4.5: three more breaks

If you are moving from Sonnet 4.5 because of the November 30 retirement, the migration guide lists three more changes that return 400 errors.

• Thinking budgets. `thinking: {“type”: “enabled”, “budget_tokens”: N}` is rejected. Use `thinking: {“type”: “adaptive”}` with an effort level instead.
• Sampling parameters. Non-default `temperature`, `top_p` and `top_k` are rejected. Remove them.
• Prefill. A conversation can no longer end with an assistant message. The API says the conversation must end with a user message.

Prefill is the one most likely to surprise a web team. Older code often started Claude’s reply with an opening brace to force JSON. On Sonnet 5.5, use structured outputs or a tool with fixed fields instead.

Thinking also changes by default. On Sonnet 4.5, no thinking field meant no thinking. On Sonnet 5.5, a request with no thinking field runs with adaptive thinking on, and that thinking costs output tokens.

The quieter changes in Claude Sonnet 5.5

Some Claude Sonnet 5.5 changes will not throw an error but will change your results or your bill.

Effort is recalibrated. Anthropic says an effort level does not produce the same amount of thinking as it did on Sonnet 5. The default on the API is high. Re-test your effort setting instead of carrying it over.

Text between tool calls moves. Notes longer than a sentence or two between tool calls now come back as thinking blocks, which are empty at the default display setting. If your interface shows that text to users, it may go quiet.

New refusal stop reason. Responses can stop with refusal and a category such as cyber or general_harms. Handle it, or your app may show an empty answer.

More tokens and pricier images. Sonnet 5.5 uses the Sonnet 5 tokenizer, which Anthropic says produces about 30% more tokens than Sonnet 4.5 for the same text. Images can now reach 2,576 pixels and up to 4,784 tokens each.

What the Sonnet 4.5 move costs

On list price, Claude Sonnet 5.5 is cheaper. Sonnet 4.5 costs $3 input and $15 output per million tokens. Sonnet 5.5 costs $2 and $10.

The tokenizer change eats part of that saving. If the same text produces about 30% more tokens, a per-token price that is a third lower leaves you roughly 13% ahead before anything else.

Default thinking and larger images can erase the rest. A chatbot that never used thinking on Sonnet 4.5 will think by default on Sonnet 5.5 unless you set between_tools. Run a week of real traffic through Claude Sonnet 5.5 and compare bills before you switch everything.

For a broader view of which tier fits which job, see our guide to picking between Opus, Sonnet and Haiku.

A Claude Sonnet 5.5 migration checklist

1. Find every Sonnet 4.5 call. Search code, plugin settings and environment files for `claude-sonnet-4-5`. Include WordPress plugins and no-code tools.
2. Change the model ID to `claude-sonnet-5-5`.
3. Replace `disabled` thinking with `between_tools` at `high` effort or lower, or remove the field and pick an effort level.
4. Replace thinking budgets with adaptive thinking and effort.
5. Remove `temperature`, `top_p` and `top_k`.
6. Remove prefill. Use structured outputs or tools for fixed formats.
7. Switch forced tool use to `auto` plus `strict: true`, or code-side checks on Bedrock.
8. Update computer use to `computer_toolset_20260801` on the API and Google Cloud.
9. Handle the `refusal` stop reason and the new thinking blocks between tool calls.
10. Keep conversations append-only, and do not replay history across accounts.
11. Re-test effort and compare cost on real traffic before November 30.

Frequently Asked Questions