Updated daily at 05:00 UTC
Benchmark
PostTrainBench
Measures: A CLI agent gets a small base model (1 to 4B parameters), one H100 and 10 hours to improve it by post-training; the score averages the result over four base models and seven evaluations.
- Maintainer
- Rank et al. (ELLIS Institute Tübingen, Max Planck Institute for Intelligent Systems, University of Tübingen)
- Introduced
- 2026
- Status
- Active
- Saturation
- Headroom
- Typical provenance
- Independent
- In scoring set
- No
Leaderboard
Every result we hold, whoever ran it. Only independent and third-party rows count towards a score.
| Model | Average score | Provenance | Source | Evaluated |
|---|---|---|---|---|
| Claude Fable 5 Anthropic · Max effort | 41.8% | Source not stated | Epoch AI via Epoch AI | — |
| GPT-5.6 Sol OpenAI · Max effort | 36.2% | Source not stated | Epoch AI via Epoch AI | — |
| Claude Opus 5 Anthropic · Unknown effort | 35.0% | Source not stated | Epoch AI via Epoch AI | — |
| Claude Opus 4.8 Anthropic · High effort | 33.8% | Source not stated | Epoch AI via Epoch AI | — |
| Claude Opus 4.8 Anthropic · Max effort | 32.9% | Source not stated | Epoch AI via Epoch AI | — |
| Kimi K3 Moonshot AI · Unknown effort | 32.0% | Source not stated | Epoch AI via Epoch AI | — |
| GLM-5.2 Z.ai (Zhipu AI) · Max effort | 31.7% | Source not stated | Epoch AI via Epoch AI | — |
| Claude Opus 4.7 Anthropic · Xhigh effort | 28.6% | Source not stated | Epoch AI via Epoch AI | — |
| GPT-5.5 OpenAI · Xhigh effort | 27.2% | Source not stated | Epoch AI via Epoch AI | — |
| Grok 4.5 SpaceXAI (xAI) · High effort | 23.4% | Source not stated | Epoch AI via Epoch AI | — |
| Gemini 3.1 Pro | 22.0% | Source not stated | Epoch AI via Epoch AI | — |
| GPT-5.4 OpenAI · High effort | 19.0% | Source not stated | Epoch AI via Epoch AI | — |
- Epoch AI, 'Capabilities & benchmarking', epoch.ai/benchmarks, CC BY 4.0.