AgentProcessBench - tau2-Bench Step Accuracy: leaderboard
Metric: StepAcc (%): share of assistant steps, pooled over trajectories, whose correct (+1), neutral (0) or incorrect (-1) label from the judge matches the expert annotation, on the 250 tau2-Bench trajectories (customer-service tool use with the official tau2-Bench environments) of AgentProcessBench's 1,000 human-annotated tool-use trajectories (250 per source benchmark, rollouts of five agent models); higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash (Preview) (Non-reasoning) | 81.2 | #78 (Gemini 3 Flash (Preview)) |
| 2 | GPT-5.2 Instant | 80.4 | #205 |
| 3 | Kimi K2.5 (Thinking) | 78.3 | #139 (Kimi K2.5) |
| 4 | DeepSeek V3.2 (Thinking) | 77.2 | #198 (DeepSeek V3.2) |
| 5 | GPT-5.2 (Medium) | 75.2 | #105 (GPT-5.2) |
| 6 | DeepSeek V3.2 (Non-reasoning) | 73.3 | #198 (DeepSeek V3.2) |
| 7 | Kimi K2.5 (Non-reasoning) | 73 | #139 (Kimi K2.5) |
| 8 | Qwen 3 30B A3B 2507 (Thinking) | 71.8 | #366 (Qwen 3 30B A3B 2507) |
| 9 | GPT-5.2 (Non-reasoning) | 70.3 | #105 (GPT-5.2) |
| 10 | Qwen 3 30B A3B 2507 Instruct | 67.2 | #464 |
| 11 | Qwen 3 8B (Thinking) | 66.4 | #667 (Qwen 3 8B) |
| 12 | Qwen 3 4B 2507 (Thinking) | 63 | #525 (Qwen 3 4B 2507) |
| 13 | Llama 3.1 8B Instruct | 61.1 | #1018 |
| 14 | Qwen 3 4B (Reasoning) | 60.3 | #823 (Qwen 3 4B) |
| 15 | Qwen 3 4B 2507 Instruct | 59.2 | #745 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=agentprocessbench-tau2-bench-step-accuracy · How It Works · Data refreshed daily, snapshot 2026-10-11.