AgentProcessBench: 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, pooled over all assistant steps of AgentProcessBench's 1,000 human-annotated tool-use trajectories (250 per source benchmark, rollouts of five agent models) (the printed average weights each source by its number of steps); 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) | 78.3 | #78 (Gemini 3 Flash (Preview)) |
| 2 | Kimi K2.5 (Thinking) | 75.9 | #139 (Kimi K2.5) |
| 3 | GPT-5.2 Instant | 74.8 | #205 |
| 4 | DeepSeek V3.2 (Thinking) | 72.8 | #198 (DeepSeek V3.2) |
| 5 | DeepSeek V3.2 (Non-reasoning) | 71.4 | #198 (DeepSeek V3.2) |
| 6 | Kimi K2.5 (Non-reasoning) | 71.4 | #139 (Kimi K2.5) |
| 7 | GPT-5.2 (Medium) | 71 | #105 (GPT-5.2) |
| 8 | GPT-5.2 (Non-reasoning) | 70.1 | #105 (GPT-5.2) |
| 9 | Qwen 3 30B A3B 2507 (Thinking) | 68.5 | #366 (Qwen 3 30B A3B 2507) |
| 10 | Qwen 3 30B A3B 2507 Instruct | 65 | #464 |
| 11 | Qwen 3 8B (Thinking) | 63.2 | #667 (Qwen 3 8B) |
| 12 | Qwen 3 4B 2507 (Thinking) | 60 | #525 (Qwen 3 4B 2507) |
| 13 | Qwen 3 4B 2507 Instruct | 58.9 | #745 |
| 14 | Qwen 3 4B (Reasoning) | 58.8 | #823 (Qwen 3 4B) |
| 15 | Qwen 3 8B (Non-reasoning) | 57.1 | #667 (Qwen 3 8B) |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=agentprocessbench · How It Works · Data refreshed daily, snapshot 2026-10-11.