VariantBench
Science · 2026-07-16
VariantBench evaluates agents on 118 real-world genetic-variant discovery, quality-control, statistical genetics, population genomics, and clinical interpretation tasks. Scores average three independent trials per task, for 354 gradeable task runs; repeats are not distinct evaluation items.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5 | 49.7 |
| GPT-6 Astra | 48.6 |
| Grok 4.6 | 46.0 |
| Opus 4.8 | 42.1 |
| GPT-5.6 Sol | 42.1 |
| Sonnet 5 | 38.7 |
| GPT-5.6 Terra | 38.7 |
| Grok 4.5 | 38.1 |
| Opus 4.7 | 37.0 |
| Gemini 3.5 Flash | 35.3 |
| Gemini 3.1 Pro Preview | 34.5 |
| Kimi K3 | 32.8 |
| GPT-5.5 | 30.2 |
| GPT-5.6 Luna | 29.9 |
| Grok 4.3 | 20.3 |