Skip to content

Benchmark

tau2-bench v1 retail pass^3

Agentic benchmarks measure how well models carry out tasks that take several steps. Scores are percentages. Higher is better.

Models measured
7
Scores recorded
8
Sources
1
IQ weight
0.08

Score distribution

56.1%
Mean
57.9%
Median
5.5%
Std dev
51.1%
10th percentile
59.9%
90th percentile
47.4%
Min
47.4%66.5%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 0.08 for chat models, measured over 7 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.

Leaderboard

#ModelScore
1Qwen3.5 397B A17Bsubmission-qwen3.5-397b-a17b-think_sierra_2026-03-02, usersim-gpt-5.2, effort-enabled· Sierra tau-bench Leaderboard data docs66.5%
2GPT 5.2submission-gpt-5-2_sierra_2026-02-26, usersim-gpt-5.2, effort-high· Sierra tau-bench Leaderboard data docs59.9%
3Claude Opus 4.5submission-claude-opus-4-5_sierra_2026-02-26, usersim-gpt-5.2, effort-high· Sierra tau-bench Leaderboard data docs58.8%
4Gemini 3 Flashsubmission-gemini-3-flash_sierra_2026-03-02, usersim-gpt-5.2, effort-high· Sierra tau-bench Leaderboard data docs57.9%
5Gemini 3 Pro Previewsubmission-gemini-3-pro_sierra_2026-03-02, usersim-gpt-5.2, effort-high· Sierra tau-bench Leaderboard data docs54.4%
6GPT 5.2submission-gpt-5-2-none_sierra_2026-02-26, usersim-gpt-5.2, effort-none· Sierra tau-bench Leaderboard data docs53.1%
7GLM 5submission-glm-5-think_sierra_2026-03-02, usersim-gpt-5.2, effort-enabled· Sierra tau-bench Leaderboard data docs51.1%
8Claude Sonnet 4.5submission-claude-sonnet-4-5_sierra_2026-02-26, usersim-gpt-5.2, effort-enabled· Sierra tau-bench Leaderboard data docs47.4%

Data credits: Sierra Research tau2-bench leaderboard submissions (github.com/sierra-research/tau2-bench), MIT (licence)