PostTrainBench — leaderboard
AI R&D automation benchmark: agents get 4 small LLMs, one H100, and 10 hours to improve model performance through post-training. Measures dataset curation, training strategy, and constraint awareness.
Metric: Weighted Avg Score. Source: posttrainbench.com. Status: saturation imminent. 12 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Opus 4.6 | 23.2 |
| 2 | Gemini 3.1 Pro (Preview) | 21.6 |
| 3 | GPT-5.2 | 21.4 |
| 4 | GPT-5.4 | 20.2 |
| 5 | GPT-5.1 Codex Max | 19.7 |
| 6 | Gemini 3 Pro | 18.1 |
| 7 | GPT-5.3 Codex | 17.8 |
| 8 | GPT-5.2 Codex | 17.2 |
| 9 | Claude Opus 4.5 | 17.1 |
| 10 | Claude Sonnet 4.6 | 16.4 |
| 11 | GLM-5 | 13.9 |
| 12 | Claude Sonnet 4.5 | 9.9 |
Interactive version: theaggregate.ai/benchmark?slug=posttrainbench · How the rankings work · Data refreshed daily, snapshot 2026-07-22.