Troubleshooting thinking (Claude API docs)

Summary

Anthropic's troubleshooting page for thinking. It opens with a per-model table of which thinking.type values each Claude model supports, its default, and which values return a 400 error, then explains the common thinking-related errors. No publication date is shown; content as read on 2026-10-13.

Key points

• Always on (cannot disable): Fable 5.1, Fable 5, Opus 5.5, Mythos 5.1, Mythos 5, and Mythos Preview. Sending "disabled" returns a 400; use display: "omitted" to hide thinking text instead.
• On by default, can be turned off: Opus 5 and Haiku 5.5 accept thinking: {type: "disabled"}, but only at effort high or below; xhigh or max with thinking disabled returns a 400. On Haiku 5.5, changing effort per message while thinking is disabled also fails. Sonnet 5 also accepts "disabled".
• Sonnet 5.5 rejects "disabled" and instead takes thinking: {type: "between_tools"} (no up-front thinking; short updates between tool calls), at effort high or below.
• Off by default: Opus 4.8 and 4.7 (adaptive only), Opus 4.6 and Sonnet 4.6 (adaptive or deprecated extended). Claude 4.5 models support only extended thinking with budget_tokens; 4.7 and later reject "enabled".
• Replayed thinking blocks on Fable 5.1, Opus 5.5, Sonnet 5.5, and Haiku 5.5 are bound to an unchanged prefix (system, tools, earlier messages); enforced for accounts created on or after 2026-08-31.
• Opus 5 with thinking disabled can leak tool calls or XML tags into text output; Anthropic recommends keeping thinking on and lowering effort instead.

Why it matters

The table settles which Claude models can run without thinking. Among current models, Haiku 5.5 is the only one where a plain "disabled" works, which matters for low-latency classification and routing.