Source: The "think" tool (Anthropic Engineering, 2025)

Anthropic engineering post on a no-op 'think' tool for mid-task reasoning, compared with extended thinking; later superseded by interleaved thinking.

Contents
  • What: engineering blog post
  • Publisher: Anthropic
  • Date: 2025-03-20, updated 2025-12-15
  • Original: The “think” tool

Summary

Describes a tool whose only effect is to log a thought, giving the model room to reason after tool results arrive mid-task. That was something up-front extended thinking did not cover at the time. The December 2025 update says extended thinking has improved and is now recommended over the think tool in most cases.

Key claims

  • Extended thinking (as of the post) happens before the response; the think tool happens during it, after new tool results.
  • On τ-Bench airline, the think tool with an optimized prompt scored 0.570 pass^1 against 0.370 baseline (54% relative). Extended thinking performed similarly to the unprompted think tool.
  • Best fit: long tool chains, policy-heavy domains, sequential decisions where mistakes are costly. Not useful for single or parallel tool calls or simple instruction following.
  • Domain-specific examples of how to reason help most.
  • Update (2025-12): extended thinking is now preferred in most cases.

Feeds

Extended thinking · Reasoning models

Suggest changes →