Agent2 RL-Bench - GSM8K: leaderboard

Agentic post-training engineering on Agent^2 RL-Bench (2026): a coding agent driven by the model has 12 hours in an isolated GPU workspace to post-train Qwen3-8B-Base, and the grading server keeps the best checkpoint score of the run. GSM8K test accuracy of the trained model, in percent; higher is better.

Metric: GSM8K test accuracy (%) 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: Estimated already saturated. 6 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GPT-5.289.46#185
2Claude Opus 4.688.78#113
3Claude Sonnet 4.588.7#261
4GPT-5.484#135
5GPT-4o83.32#751
6Gemini 2.5 Flash70.13#463

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=agent2-rl-bench-gsm8k · How It Works · Data refreshed daily, snapshot 2026-11-21.