85 agentic tools audited for security incidents · Talk to the lab →
MotionticBook a call
Glossary

What is the think tool?

The think tool is a tool definition that does nothing except give a model a place to write down reasoning in the middle of a tool-use sequence, before it takes its next action.

Anthropic described the think tool in March 2025. The definition is a tool whose only input is a thought; calling it changes nothing outside the model. Its value is that it gives the model a sanctioned pause after a tool result arrives, to check the result against the rules and decide whether it has what it needs.

That makes it different from extended thinking, which happens before the model starts its response. The think tool is used during the response and focuses on new information, such as what a tool just returned.

Where it helped

Anthropic's tests on Claude 3.7 Sonnet used τ-bench, which measures customer service agents working under detailed policies. In the retail domain, the think tool alone raised the pass^1 score from 0.783 to 0.812. On SWE-bench, a similar think tool contributed an average improvement of 1.6%.

  • Reading tool output carefully

    Cases where a result has to be checked before acting on it, and backtracking may be needed.

  • Policy-heavy work

    Environments with detailed rules, where each step has to be checked for compliance.

  • Sequential decisions

    Chains where each step builds on the last and mistakes are costly.

  • Little benefit

    Single or parallel tool calls, and simple instruction following with few constraints.

Where it stands now

In a December 2025 update to the post, Anthropic said extended thinking had improved enough that it now recommends using extended thinking instead of a dedicated think tool in most cases. Current models from the major labs can reason between tool calls without a workaround, and OpenAI's GPT-5 prompting guide recommends its Responses API so that reasoning carries over from one tool call to the next.

The idea behind the tool still holds: give the agent an explicit checkpoint to verify a result against policy before it acts. What changed is that the model now provides that space itself.

How Motiontic's engineers apply it

Motiontic's engineers rely on the model's built-in reasoning first, and add explicit checkpoints only where evals show policy slips between steps. Where a check really matters, such as a refund limit, it runs in code rather than in the model's reasoning.

The think tool compared with extended thinking

The think tool compared with extended thinking
Think toolExtended thinking
When the reasoning happensDuring the response, between tool callsBefore the response begins
What it focuses onNew information, such as a tool resultThe task as a whole
How it is enabledA tool definition plus prompt guidanceA model or API setting
Anthropic's guidance as of December 2025Use only in specific casesPreferred in most cases

Common questions

Is the think tool still recommended?

Mostly not. Anthropic's December 2025 update to its original post recommends extended thinking instead of a dedicated think tool in most cases, because the built-in feature has improved.

What is the difference between the think tool and chain of thought?

Chain of thought is reasoning written out as part of the answer. The think tool gives the model a separate place to reason partway through a sequence of tool calls, after it sees each result.

Does the think tool work with models other than Claude?

Any model with tool calling can be given a no-op tool. Anthropic's published results, though, were measured on Claude models, so the benefit on other models needs its own eval.

Running agents that have to hold up?

Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.