Adaptive thinking and effort ⚠ verify

Definition

Adaptive thinking (thinking={'type':'adaptive'}) lets Claude dynamically allocate its reasoning budget, tuned by an effort level rather than a manual token budget.

Key points

• Config: thinking={'type':'adaptive'} plus effort ∈ low | medium | high | xhigh | max.
• xhigh = best for coding/agentic work; high = minimum for intelligence-sensitive tasks.
• effort defaults to high on Opus 4.8 (all surfaces) and on Sonnet 5 (Claude API + Claude Code). ⚠ verify (2026-07-08).
• Manual budget_tokens is deprecated on 4.6+ and returns a 400 error on Opus 4.8/4.7. ⚠ verify (2026-07-08).
• Thinking budget tokens are a subset of max_tokens, billed as output tokens, and count toward rate limits. With adaptive thinking Claude determines allocation dynamically.
• tool_choice: any/tool is NOT compatible with extended thinking — use auto/none (see Extended thinking with tool use).
• Enables Interleaved thinking automatically on 4.6+ / Sonnet 4.6.

Why it matters for the exam

• D1 and D4 test picking the right effort (xhigh for coding/agentic, high floor for intelligence-sensitive) and knowing that manual budget_tokens now errors on newer models.

Common gotchas

• Setting budget_tokens on Opus 4.8/4.7 → 400 error; migrate to effort. ⚠ verify (2026-07-08).
• Forcing tools with any/tool while thinking is on → error; only auto/none are allowed.

See also

• Interleaved thinking
• Extended thinking with tool use
• Agentic loop

Sources

• netlight-study-guide
• docs-models-pricing