Benchmarks › τ-bench

τ-bench

Resolve a simulated customer issue using the allowed tools
50.7top score · Grok 4.6
134models ranked
84open-weights models
Pass@1 success ratehigher is better
Oct 20, 2026last updated

How the test works

1

Task

97 tasks

  • What the model doesResolve a simulated customer issue using the allowed tools
  • Kinds of tasksDual-control customer support
  • In → outText → Tool calls, dialogue
2

Run

5 runs per task

  • Agent / harnessUpstream tau2-bench harness
  • Step limit200 steps
  • Token budgetNon-reasoning: up to 16,384 output tokens, adjusted to model limits; reasoning: model-specific maximum
3

Score

Pass@1 (solved on the first try) success rate

  • Graded byBackend world/database-state checks
  • Better isHigher

τ³-Banking leaderboardtop 10 of 134

Full leaderboard with filters →
Modelτ³-BankingPer taskPriceSpeedSupportsContext
50.7high
$1.48$2 / $665500K
50.5max
$1.60$1.25 / $4.251281M
GLM 5.3Z.aiOpen
50.3max
$2.01$1.40 / $4.401861M
Qwen3.8 2.4T A95BAlibaba (Qwen)Open
49.1thinking
$2.16$2 / $6127262K
Qwen3.8 27BAlibaba (Qwen)Open
48xhigh
$1.01$0.42 / $2.55114262K
Qwen3.8 Max (0902)Alibaba (Qwen)
47.8thinking
$5.41$2 / $640984K
47.2max
$7.63$10 / $50691M
47.2max
$0.28$0.15 / $0.501791M
Kimi K3Moonshot AIOpen
46max
$2.00$3 / $151121M
45.8medium
$0.93$0.75 / $3.75—1M

Versions and boards on Model Menu

BoardModels
τ³-BankingShown above
134
21

τ-bench questions

What is τ-bench?›

Resolve a simulated customer issue using the allowed tools. It has 97 tasks. Created by Sierra.

Which model is best on τ-bench?›

Grok 4.6 leads τ³-Banking with 50.7, ahead of Muse Spark 1.3 (50.5) and GLM 5.3 (50.3).

What is the best open-weights model on τ-bench?›

GLM 5.3 with 50.3, then Qwen3.8 2.4T A95B (49.1).

How is τ-bench scored?›

Pass@1 (solved on the first try) success rate; backend world/database-state checks; 5 runs per task per task.

Which versions of τ-bench are there?›

τ³-Banking (134 models), τ³-bench · Banking 1.0.1 (21 models).