AARRI-Bench (Mini-SWE-Agent): leaderboard
AARRI-Bench (Act As a Real Research Intern, 2026) is a set of 82 hand-crafted AI-research tasks graded by unit tests, testing whether an agent harness and model behave like a careful human researcher. The board is the percentage of tasks solved in full, every unit test passing, with the Mini-SWE-Agent harness; higher is better.
Metric: Task success rate (%) under the 0/1 reward (a task counts only when every one of its unit tests passes) over the 82 hand-crafted AI-research-intern tasks in four scenario categories; Harbor containerized runs, Mini-SWE-Agent agent harness; a configuration's tasks without a valid run are left out of its denominator; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 | 68.3 | #79 |
| 2 | DeepSeek V4 Flash | 60.5 | #258 |
| 3 | Qwen 3.6 Plus | 59.8 | #222 |
| 4 | MiniMax-M2.7 | 56.8 | #518 |
| 5 | Kimi K2.6 | 56.4 | #178 |
| 6 | GPT-OSS-120B | 43.5 | #700 |
| 7 | Qwen 3 235B A22B 2507 (Thinking) | 40.2 | #463 |
| 8 | Qwen 3 Next 80B A3B Instruct | 39.9 | #766 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=aarri-bench-mini-swe-agent · How It Works · Data refreshed daily, snapshot 2026-11-13.