Anthropic extended thinking: budget_tokens is dead on 4.7 and later, use adaptive thinking instead

Extended thinking configuration: manual vs adaptive

1. On current models (Claude 4.7 and later, e.g. Claude Opus 5.5, Claude Sonnet 5, Claude Fable 5.1), use adaptive thinking: thinking set to {"type": "enabled"} with no budget. Do not send budget_tokens.
2. Extended thinking with an explicit budget is deprecated on the Claude 4.6 models (requests using it still succeed). On Claude 4.7 and later it is not supported: requests that include it are rejected with a 400 error.
3. If your code targets 4.6-era models and needs predictable latency or precise control over the thinking budget, manual mode remains the documented choice there. Gate it on the model, not on a global flag.
4. In manual mode, the budget must be less than max_tokens; the budget has to leave room for the final response. The one exception is interleaved thinking, where the budget can exceed max_tokens because the budget spans all thinking blocks within one assistant turn.
5. Extended thinking cannot be combined with max_tokens set to 0 (cache pre-warming), because the budget must be less than max_tokens.
6. When migrating to adaptive thinking, remove every budget reference from your request builders, prompts, and docs. A stale helper that injects it will 400 on every call to a new model.

Failure modes this prevents: 400s on every request after a model upgrade; copying a budget snippet from training data onto Sonnet 5; setting the budget above max_tokens and starving the final answer.

Anthropic's extended-thinking docs state the budget approach is deprecated on 4.6, rejected with 400 on 4.7 and later, the budget < max_tokens rule with the interleaved-thinking exception, and the adaptive migration path. Verified live at https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking.