AgentProcessBench - tau2-Bench First-Error Accuracy: leaderboard
Metric: FirstErrAcc (%): share of trajectories where the judge locates the first incorrect step at the same position as the experts (or agrees there is none), 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 2028. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V3.2 (Thinking) | 65.6 | #198 (DeepSeek V3.2) |
| 2 | Kimi K2.5 (Thinking) | 65.6 | #139 (Kimi K2.5) |
| 3 | Qwen 3 30B A3B 2507 (Thinking) | 61.6 | #366 (Qwen 3 30B A3B 2507) |
| 4 | GPT-5.2 Instant | 58.4 | #205 |
| 5 | DeepSeek V3.2 (Non-reasoning) | 57.6 | #198 (DeepSeek V3.2) |
| 6 | GPT-5.2 (Non-reasoning) | 56.4 | #105 (GPT-5.2) |
| 7 | Gemini 3 Flash (Preview) (Non-reasoning) | 56 | #78 (Gemini 3 Flash (Preview)) |
| 8 | GPT-5.2 (Medium) | 53.6 | #105 (GPT-5.2) |
| 9 | Qwen 3 30B A3B 2507 Instruct | 49.6 | #464 |
| 10 | Kimi K2.5 (Non-reasoning) | 49.2 | #139 (Kimi K2.5) |
| 11 | Qwen 3 4B 2507 Instruct | 48.8 | #745 |
| 12 | Qwen 3 4B (Reasoning) | 47.6 | #823 (Qwen 3 4B) |
| 13 | Llama 3.1 8B Instruct | 47.2 | #1018 |
| 14 | Llama 3.3 70B Instruct | 46.8 | #520 |
| 15 | Qwen 3 4B 2507 (Thinking) | 46.8 | #525 (Qwen 3 4B 2507) |
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-first-error-accuracy · How It Works · Data refreshed daily, snapshot 2026-10-11.