MatFlowBench Pass Rate
Science · 2026-10-09
Share of all 130 MatFlowBench rollouts that pass the task verifier (25 computational tasks x 2 plus 20 experimental tasks x 4), the publisher's ranking metric. Experimental tasks supply 80 of the 130 rollouts, so this pooled rate weights them more than Pass@1 does; rollouts of one task are not independent items.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6 Astra | 60.8 |
| Claude Opus 5.5 | 53.8 |
| Claude Sonnet 5.5 | 47.7 |
| GPT-6.1 Sol | 44.6 |
| Grok 4.7 | 26.2 |
| Gemini 3.8 Flash | 25.4 |
| Kimi K3 | 23.8 |
| Hy4 Preview | 22.3 |
| Muse Spark 1.3 | 20.0 |