VariantBench V2
Science
VariantBench V2 clean-v1 reruns on 118 variant-analysis evaluations. Macro-mean pass rate on updated evaluation sets, distinct from V1.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5 | 52.8 |
| Claude Opus 5.5 | 50.6 |
| GPT-6 Astra | 46.6 |
| Grok 4.7 | 45.5 |
| Opus 4.8 | 42.4 |
| Grok 4.6 | 40.4 |
| Sonnet 5 | 39.8 |
| GPT-6 Sol | 38.7 |
| GPT-5.6 Sol | 35.3 |
| Opus 4.7 | 35.0 |
| Gemini 3.7 Flash | 31.4 |
| Gemini 3.8 Flash | 31.4 |
| Sonnet 4.6 | 31.1 |
| GPT-5.5 | 30.5 |
| GPT-6 Luna | 27.7 |
| GPT-5.6 Luna | 27.1 |