τ²-benchBrowse 296

τ²-bench

Can the model plan, use tools, and finish a multi-step task?

Version not specifiedExact reported variant
Agents8 ranked models11 reported values2 reportsHigher is better

Top rankings

One row per model, using its best reported score across effort settings.

8
RankModelBest scoreBest reported setting
1Claude Opus 4.6Anthropic99.3%max effortGoogle report2 values · 2 reportsFeb 19, 2026 · source
2Gemini 3.1 ProGoogle99.3%high effortGoogle report1 value · 1 reportFeb 19, 2026 · source
3GPT-5.2OpenAI98.7%xhigh effortGoogle report2 values · 2 reportsFeb 19, 2026 · source
4Claude Opus 4.5Anthropic98.2%Reported configurationMiniMax report1 value · 1 reportFeb 12, 2026 · source
5Gemini 3 ProGoogle98.0%high effortGoogle report2 values · 2 reportsFeb 19, 2026 · source
6Claude Sonnet 4.6Anthropic97.9%max effortGoogle report1 value · 1 reportFeb 19, 2026 · source
7MiniMax M2.5MiniMax97.8%Reported configurationMiniMax report1 value · 1 reportFeb 12, 2026 · source
8MiniMax M2.1MiniMax87.0%Reported configurationMiniMax report1 value · 1 reportFeb 12, 2026 · source

Effort curve

Every sourced cost-linked effort value for this exact version. Lines connect complete sweeps only.

0
No cost-linked effort sweep for this version.
Definition and comparison boundarypublic methodology

Can the model plan, use tools, and finish a multi-step task?

Agentic execution across tools, environments, or long-running workflows. Higher is better. The value is the percentage reported in this lab's table.

Harness, tools, prompts, attempt count, and benchmark version can materially change the result. Compare within one reporting context.

Method / source