Interleaved thinking ⚠ verify

Definition

Interleaved thinking is reasoning that happens between tool calls during an agentic run, letting Claude reflect on a tool result before deciding the next action.

Key points

• Automatic on adaptive-thinking models: Opus 4.6+ / Sonnet 4.6 with adaptive thinking.
• Opus 4.5 / Sonnet 4.5 / earlier Claude 4 need the beta header interleaved-thinking-2025-05-14. ⚠ verify (2026-07-08).
• Complements the Agentic loop — thinking blocks appear across successive tool_use turns.
• Governed by the same thinking-block rules as Extended thinking with tool use: the unmodified thinking block (with signature) must accompany its tool_result.

Why it matters for the exam

• Tests knowing interleaved thinking is automatic on 4.6+/Sonnet 4.6 with adaptive thinking, and that older models require the explicit beta header.

Common gotchas

• Don't assume you need a beta header on 4.6+ — it's on by default there; the header is only for Opus 4.5 / Sonnet 4.5 / earlier Claude 4.

See also

• Adaptive thinking and effort
• Extended thinking with tool use
• Agentic loop

Sources

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