Anthropic Claude Opus 5.5 rejects disabled thinking, causing 400 errors
Anthropic has released Claude Opus 5.5, Fable 5.1, and Sonnet 5.5, introducing breaking changes to the API that may disrupt existing automation. The most significant change is that Claude Opus 5.5 no longer accepts requests to disable or explicitly enable thinking; including thinking: {"type": "disabled"} or thinking: {"type": "enabled"} now returns a 400 error. Developers must omit the thinking…
Key points
- Claude Opus 5.5 returns a 400 error if thinking is disabled or explicitly enabled.
- Developers must omit the thinking field and use the effort parameter for control.
- Forced tool_choice specifications are rejected; users must switch to auto mode.
Additionally, the new model rejects forced tool specifications using tool_choice set to any or specific tool names. Users are required to switch to auto mode and utilize strict tool use configurations. The release also discontinues the fast mode for the previous Opus 4.7 model. While these changes do not affect the Claude Code tool directly, they impact any custom automation or application that calls the Claude API directly, such as news gathering or article generation systems.
Developers are advised to review their code for instances of the deprecated thinking and tool_choice parameters. The article notes that similar code written for Sonnet 5 may break in five places when migrated to Sonnet 5.5. These updates are part of a broader ecosystem shift, coinciding with OpenAI deprecating older model snapshots like gpt-4o-2024-05-13 and o1 on the same date.
Model page: Sonnet 5.5 →
Claude Opus 5.5, disabling thinking results in 400 error
note.com · 23 October 2026
Loading the full article…
This text was published by note.com and written by amu_lab｜AI検証ジャーナル. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Generative AI & Models
All →- How Do AI Video Generators Work? From Diffusion to World Simulation · 1 src
- Ahead of AI releases eight‑hour video course on building reasoning models from scratch · 1 src
- Nvidia and University of Washington unveil open-source protein design model RFD4-Proteina · 2 src
- Moonshot AI releases Kimi K3.1 Pro, a 2.8T open-weight model closing gap with proprietary leaders · 6 src
- NVIDIA releases Nemotron 3.5 Bolt E4B for local voice agents and robots · 2 src
Comments
via GitHub Discussions