What is Claude Sonnet 5.5, and what changes if you move from Sonnet 5?

What Sonnet 5.5 is

Anthropic released Claude Sonnet 5.5 on September 28, 2026 and calls it the second model in the Claude 5.5 family. Opus 5.5 came first on September 22. The announcement says Claude Haiku 5.5 will follow in the coming weeks.

It is a hosted, proprietary model. The API ID is claude-sonnet-5-5. It is listed on Amazon Bedrock as anthropic.claude-sonnet-5-5 and on Google Cloud, Microsoft Foundry and Claude Platform on AWS as claude-sonnet-5-5. The model page lists a 1M-token context window, 128K max output tokens, a June 2026 knowledge cutoff and a retirement date no sooner than September 28, 2027.

Anthropic describes Sonnet 5.5 as strongest at well-scoped everyday tasks, fixing bugs and producing documents, slides and spreadsheets, and as a faster, lower-cost complement to Opus 5.5.

The price, stated plainly

Per-token prices did not change from Sonnet 5. The claim that Sonnet 5.5 is cheaper is about tokens per task, not the rate card: Anthropic says it typically needs far fewer tokens to do the same work and, in its testing, costs up to 30 percent less per task. Your own savings depend on your prompts and effort setting, so measure them.

• Input: 2 US dollars per million tokens.
• Output: 10 US dollars per million tokens.
• Cache reads: 0.20 per million. Five-minute cache writes: 2.50 per million. One-hour cache writes: 4 per million.
• Batch API: 50 percent off input and output.
• For comparison, Opus 5.5 lists 4 and 20 per million input and output tokens.

What Anthropic says improved

All figures below are vendor-reported from the launch page and have not been independently reproduced by RepoRadar. Benchmark versions are new, so treat the numbers as directional.

Anthropic also says Sonnet 5.5 at its highest effort performs comparably to Opus 5.5 on several evaluations, but that in its own and external testing Opus 5.5 remains clearly stronger at complex, open-ended work requiring sustained judgment. That caveat is worth more than any single score when you pick between them.

• Terminal-Bench 4.0 (agentic coding): 70.6 percent, against 10.3 percent for Sonnet 5 and 66.4 percent for Opus 5.5 at Xhigh effort.
• CursorBench 4.0: 55.5 percent, against 34.1 percent for Sonnet 5 and 57.8 percent for Opus 5.5.
• FrontierCode 1.1: 52.1 percent at Xhigh effort and 46.2 percent at Max, against 42.4 percent for Sonnet 5 and 54.4 percent for Opus 5.5. Anthropic notes the lower Max score came partly from out-of-scope edits.
• OSWorld 2.1 (computer use): 80.1 percent, against 57.0 percent for Sonnet 5.
• Speed: output more than 30 percent faster than Sonnet 5.

Five breaking changes to test before you switch

These come from Anthropic's What's new and migration pages for Sonnet 5.5. Each returns an HTTP 400 error, so they show up quickly in testing.

• Thinking cannot be set to disabled. Send thinking type between_tools instead, which turns off up-front thinking. It works only at low, medium or high effort; at xhigh or max it returns an error.
• Forced tool use is rejected. tool_choice of any or a named tool returns an error. Keep tool_choice on auto and use strict tool use or structured outputs for schema-valid input.
• Thinking blocks are tied to the model and the conversation. Sonnet 5.5 can read thinking from Sonnet 5 and older models, but no other model reads Sonnet 5.5 blocks. For accounts created on or after August 31, 2026, replaying a block after editing earlier history returns an error by default. Keep conversations append-only.
• On the Claude API and Google Cloud, the older computer_20251124 computer use tool is rejected; use the computer_toolset_20260801 toolset. Amazon Bedrock still accepts the older tool.
• The advisor tool (beta) refuses Opus 4.8, Opus 4.7 and Sonnet 5 as advisors for a Sonnet 5.5 executor.

Changes that do not throw an error

Some differences only change behavior, which makes them easier to miss.

• Adaptive thinking is on by default, and thinking tokens are billed as output tokens. Revisit max_tokens.
• Effort levels are recalibrated. Anthropic says a given effort level does not produce the same amount of thinking as on Sonnet 5, and recommends re-running your effort sweep. The default effort on the Claude API is high.
• Longer notes between tool calls now arrive as thinking blocks. At the default display setting their text is empty, so an app that streams those notes to users goes quiet with no error.
• Non-default temperature, top_p or top_k values return an error.
• Thinking blocks work only in the account that produced them or a linked account; blocks sent from another account are silently dropped.
• New safeguards can decline a request in five categories, including cyber, frontier_llm (helping build competing AI models) and reasoning_extraction. Higher-risk cyber tasks visibly fall back to Sonnet 5. Handle stop_reason refusal in your code.

Who should switch, and who should wait

If you use Sonnet through Claude Code, the Claude apps or a managed product, the provider handles the API changes; judge it on speed and results. If you call the API yourself and never set thinking, tool_choice or sampling parameters, the swap is mostly the model ID plus reading content blocks by type.

Wait and test if your code disables thinking, forces tool calls, edits conversation history between turns, uses computer use on the Claude API, streams text between tool calls to users, or routes turns between different Claude models. Those are the paths that break or change silently.

Limits of this answer

This answer is based on Anthropic's launch page and platform documentation as read on September 28, 2026, the release day. RepoRadar has not run Sonnet 5.5 hands-on, and every benchmark and cost figure is vendor-reported. Prices, limits and platform availability can change; check the linked pages before budgeting.

A useful next action

Before switching a production workload, run your own evaluation set on Sonnet 5 and Sonnet 5.5 side by side at two effort levels, log tokens per task, and grep your code for thinking disabled, tool_choice any or tool, and computer_20251124. Our answers on reading AI benchmarks and on what happens when a model you depend on is removed cover the rest of the checklist.