Anthropic is removing the fixed-thinking-budget option for Claude Opus 4.6 and Sonnet 4.6. Developers who set a specific token limit for the model's reasoning process will need to switch to adaptive mode, where Claude decides how much to think based on the problem at hand.
A quick explanation of what this means: extended thinking is Claude's ability to work through a problem step-by-step before responding - essentially showing its work before giving a final answer. With a fixed budget, you tell the API exactly how many "thinking tokens" to use per query: always 4,000, always 8,000, regardless of complexity. Adaptive thinking removes that constraint. Claude decides whether a simple query needs 200 tokens of reasoning or a complex one needs 15,000.
For Claude Opus 4.7 and newer models, the fixed-budget option doesn't exist at all. The API uses an effort parameter - low, medium, or high - to signal roughly how much reasoning you want without pinning an exact token count.
Developers who explicitly disabled adaptive thinking in Claude Code will be directly affected - that toggle is going away. Anyone using thinking: {type: "enabled"} with a token budget in the API needs to migrate to thinking: {type: "adaptive"} with an effort parameter.
In practice, this matters most for agentic workflows and structured-output pipelines where consistent reasoning depth was a design assumption. If you tuned a fixed budget to prevent over-thinking simple queries or under-thinking complex ones, that calibration will need to be rebuilt around the effort parameter. Test your prompt behavior after the switch, especially for tasks with strict output format requirements.
Anthropic hasn't published a hard cutoff date, but the direction is clear: adaptive thinking is the standard going forward, and the fixed-budget approach is being retired across the model family.