Claude API 400 invalid_request_error about thinking: '"thinking.type.enabled" is not supported for this model', 'adaptive thinking is not supported on this model', or '"thinking.type.disabled" is not supported'. Which thinking setting does each model accept?

Cause: Claude models differ in how thinking is configured, and sending the wrong 'thinking' setting returns a 400 invalid_request_error. This usually happens when you change the model id but keep request code written for an older model, or when one code path serves several models. What each error means (per Anthropic's error documentation): - '"thinking.type.enabled" is not supported for this model': Claude 4.7 and later removed extended thinking (type enabled with budget_tokens). Use adaptive thinking instead: omit 'thinking' (adaptive is the default) or send thinking: {"type": "adaptive"}, and control depth with output_config.effort. - 'adaptive thinking is not supported on this model': the reverse. Claude 4.5 and earlier models only support extended thinking: use thinking: {"type": "enabled", "budget_tokens": N} there. - '"thinking.type.disabled" is not supported for this model': on Claude Opus 5.5, Fable 5 / 5.1 and Mythos 5 / 5.1, thinking is always on. Omit 'thinking' or use adaptive with a lower effort. If you just don't want thinking text in responses, set display: "omitted" on the thinking config. - On Claude Sonnet 5.5, use thinking: {"type": "between_tools"} for the lowest setting (no up-front thinking); 'disabled' and 'enabled' are rejected, and between_tools is rejected at xhigh/max effort. Claude Haiku 5.5 accepts disabled, with the same effort limits. Fix: 1. Read the error message: it names the setting the model wants. 2. Make the thinking config depend on the model: adaptive (or omitted) for Claude 4.7+ and 5.x models, enabled+budget_tokens only for 4.5 and earlier; never send disabled to always-on models. 3. When switching models, follow the model's migration guide; related 400s appear for removed sampling parameters (temperature/top_p/top_k), forced tool_choice on Opus 5.5 / Sonnet 5.5, and edited thinking blocks in multi-turn conversations (pass them back unchanged). 4. Catch the SDK's typed BadRequestError instead of matching message strings, and log the request_id for support. Verify: the same request succeeds on the target model, and a test covers each model id you use. Sources: https://platform.claude.com/docs/en/api/errors https://platform.claude.com/docs/en/build-with-claude/thinking-troubleshooting https://platform.claude.com/docs/en/build-with-claude/thinking