This page reports published VariantBench scores from benchmarks.bio. VariantBench contains 118 evaluations built from real experimental data, each one graded deterministically against the biological result the original analysis reached, so a run counts as a pass only when the agent recovers that result rather than when it produces plausible-looking output. Results are reported per model and per harness rather than per model alone, because the harness an agent runs under changes its score substantially. Scores use a 0-100 percentage scale and are not comparable across benchmarks; see the linked repository and paper for task construction, grading, and confidence-interval methodology.
| Rank | Model | Harness | Provider | Score |
|---|---|---|---|---|
| 1 | Claude Opus 5 | Claude Code | Anthropic | 49.7% |
| 2 | Claude Opus 5 | Pi | Anthropic | 49.4% |
| 3 | GPT-6 Astra | Codex | OpenAI | 48.6% |
| 4 | GPT-6 Astra | Pi | OpenAI | 47.7% |
| 5 | Grok 4.6 | Grok Build | SpaceXAI | 46.0% |
| 6 | Claude Opus 4.8 | Pi | Anthropic | 42.1% |
| 7 | GPT-5.6 Sol | Codex | OpenAI | 42.1% |
| 8 | GPT-5.6 Sol | Pi | OpenAI | 41.2% |
| 9 | Claude Opus 4.8 | Claude Code | Anthropic | 39.8% |
| 10 | Claude Sonnet 5 | Claude Code | Anthropic | 38.7% |
| 11 | GPT-5.6 Terra | Codex | OpenAI | 38.7% |
| 12 | Claude Sonnet 5 | Pi | Anthropic | 38.1% |
| 13 | Grok 4.6 | Pi | SpaceXAI | 38.1% |
| 14 | Grok 4.5 | Pi | SpaceXAI | 37.6% |
| 15 | Claude Opus 4.7 | Pi | Anthropic | 37.0% |
| 16 | Claude Opus 4.7 | Claude Code | Anthropic | 35.6% |
| 17 | Gemini 3.5 Flash | Pi | 35.3% | |
| 18 | GPT-5.6 Terra | Pi | OpenAI | 34.7% |
| 19 | Kimi K3 | Pi | Moonshot | 32.8% |
| 20 | GPT-5.6 Luna | Codex | OpenAI | 29.9% |
| 21 | GPT-5.5 | Pi | OpenAI | 29.7% |
| 22 | GPT-5.5 | Codex | OpenAI | 28.8% |
| 23 | GPT-5.6 Luna | Pi | OpenAI | 28.2% |
Methodology and provenance
Scores use a 0–100 percentage scale. See Paper PDF, Original publication, Repository for task construction, grading, aggregation, and uncertainty details.