Agent2 RL-Bench - HumanEval: leaderboard
Metric: HumanEval pass rate (%) on the 82 held-out problems of the Qwen3-8B-Base model after the coding agent (driver LLM in a CLI scaffold) post-trains it for 12 hours in an isolated GPU workspace, best score the grading server returned over the run, single run; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 6 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Opus 4.6 | 84.15 |
| 2 | GPT-5.4 | 72.25 |
| 3 | GPT-5.2 | 69.51 |
| 4 | Claude Sonnet 4.5 | 65.85 |
| 5 | GPT-4o | 64.63 |
| 6 | Gemini 2.5 Flash | 10.98 |
Interactive version: theaggregate.ai/benchmark?slug=agent2-rl-bench-humaneval · How It Works · Data refreshed daily, snapshot 2026-10-07.