LABBench2
Can the model reason accurately about difficult scientific material?
Top rankings
One row per model, using its best reported score across effort settings.
| Rank | Model | Best score | Best reported setting |
|---|---|---|---|
| 1 | Gemini 3.8 FlashGoogle | 86.2% | Reported configurationGoogle DeepMind report1 value · 1 reportSep 2, 2026 · source |
| 2 | Claude Opus 5Anthropic | 84.2% | Reported configurationGoogle DeepMind report1 value · 1 reportSep 2, 2026 · source |
| 3 | Gemini 3.7 FlashGoogle | 82.1% | Reported configurationGoogle DeepMind report2 values · 2 reportsSep 2, 2026 · source |
| 4 | GPT-5.6 SolOpenAI | 82.1% | Reported configurationGoogle DeepMind report1 value · 1 reportSep 2, 2026 · source |
| 5 | GPT-5.6 TerraOpenAI | 81.2% | Reported configurationGoogle DeepMind report2 values · 2 reportsSep 2, 2026 · source |
| 6 | Claude Sonnet 5Anthropic | 80.1% | Reported configurationGoogle DeepMind report2 values · 2 reportsSep 2, 2026 · source |
| 7 | Gemini 3.6 FlashGoogle | 76.1% | Reported configurationGoogle report1 value · 1 reportAug 13, 2026 · source |
Effort curve
Every sourced cost-linked effort value for this exact version. Lines connect complete sweeps only.
Definition and comparison boundarypublic methodology
Can the model reason accurately about difficult scientific material?
Scientific knowledge and reasoning under the benchmark's published question set. Higher is better. The value is the percentage reported in this lab's table.
Harness, tools, prompts, attempt count, and benchmark version can materially change the result. Compare within one reporting context.
Google-computed for Gemini and GPT-5.6 Terra with a Linux terminal, bioinformatics tools, Python, R, and internet access. Biology real-world research tasks; exact model-card cells.
Method / source