Tau2-Bench Verified: leaderboard

An agent model serves simulated customers (played by gpt-5.1) in airline, retail and telecom settings, following a written policy and calling tools; the score is the unweighted mean of the three domain pass rates on the 278 base tasks (50 airline, 114 retail, 114 telecom); the README says its figures are averaged over 5 runs. Unlike tau2-bench, the tasks whose expected actions broke the policy or the database were corrected. Amazon's own Nova 2 rows were added in a later commit that does not state how they were run.

Metric: Mean of the airline, retail and telecom pass^1 (%). Source: github.com. Saturation forecast: Around February 2027. 14 models tracked. Updated 17 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1Grok 4.1 FastReasoning82.71#429
2Claude Opus 4.581.99#161
3GPT-5.2xHigh80.3#185
4GPT-5Medium79.92#177
5Gemini 3 Pro79.39#145
6Nova 2 Pro78.53#516
7GPT-5.1High76.82#235
8Nova 2 Omni75.7#472
9Claude Sonnet 4.573.32#258
10Nova 2.0 Lite72.43#1125
11GPT-5 MiniMedium69.83#359
12Gemini 2.5 Pro56.38#274
13Claude Haiku 4.553.44#580
14Gemini 2.5 Flash41.57#467

No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.

Interactive version: theaggregate.ai/benchmark?slug=tau2-bench-verified · How It Works · Data refreshed daily, snapshot 2026-11-17.