OncologyBench Strict (25 tasks, October 9 rubric)
Science · 2026-10-09
OncologyBench's October 9, 2026 strict release covers 25 selected real experimental oncology evaluations, with nine cases re-graded under a new decision rubric. The published score requires both task and decision success and reports a macro-mean over three trials per evaluation. The same exact cohort is exported in both the legacy V0 and clean V1 result catalogs; duplicate model/harness observations are counted once. Task, rubric and cohort revisions must remain distinct.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5 | 40.0 |
| Claude Opus 5.5 | 37.3 |
| Claude Sonnet 5.5 | 32.0 |
| Kimi K3 | 24.0 |
| Muse Wonder | 24.0 |
| GPT-6 Astra | 21.3 |
| GPT-6 Sol | 20.0 |
| GPT-5.6 Sol | 18.7 |
| Grok 4.6 | 18.7 |
| Grok 4.7 | 18.7 |
| Opus 4.8 | 16.0 |
| GPT-6 Luna | 16.0 |
| Sonnet 5 | 14.7 |
| DeepSeek V4.1 Flash | 14.7 |
| Gemini 3.7 Flash | 14.7 |
| GPT-5.5 | 14.7 |
| Gemini 3.8 Flash | 13.3 |
| Muse Spark Open | 12.0 |
| GPT-5.6 Luna | 10.7 |
| Mercury 3 | 8.0 |
| Sonnet 4.6 | 5.3 |