AJ-Bench: leaderboard
AJ-Bench — agentic judging benchmark; models act as judges over agent trajectories, scored by overall Avg@3 accuracy.
Metric: Overall Avg@3 (%). Source: aj-bench.github.io. 17 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | deepseek-v3.2 | 77.34 | |
| 2 | gemini-3-pro-preview | 75.05 | |
| 3 | gpt-5-mini-low | 72.41 | |
| 4 | claude-sonnet-4.5 | 69.77 | |
| 5 | grok-4 | 69.24 | |
| 6 | gemini-2.5-pro | 68.31 | |
| 7 | claude-opus-4.5 | 66.79 | |
| 8 | glm-4.6 | 64.81 | |
| 9 | deepseek-v3.2 | 64.49 | |
| 10 | kimi-k2-0905-preview | 64.33 |
Interactive version: theaggregate.ai/benchmark?slug=aj-bench · How It Works · Data refreshed daily, snapshot 2026-10-20.