Agent-Diff: leaderboard
Metric: Assertion-weighted score (%) on Agent-Diff's 224 enterprise API tasks (Box, Google Calendar, Linear, Slack replicas) in the no-documentation setting, code-executing agent, three trials per task, graded by state-diff assertions with any unexpected side effect zeroing the task; posterior mean of a task-level Bayesian bootstrap; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V3.2 | 88.1 | #198 |
| 2 | Devstral 2 | 86 | #394 |
| 3 | Kimi K2 0905 | 75.4 | #282 |
| 4 | Grok 4.1 Fast | 74.9 | #208 |
| 5 | Gemini 3 Flash (Preview) | 73.8 | #78 |
| 6 | GPT-OSS-120B | 68.5 | #330 |
| 7 | Claude Haiku 4.5 | 49.3 | #271 |
| 8 | Llama 4 Scout | 38 | #646 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=agent-diff · How It Works · Data refreshed daily, snapshot 2026-10-11.