LifeSci Diamond Benchmark
Explore score history ↗55 life-science research tasks; 10 public, 45 held out. Strict pass requires every rubric line fully met and no penalty fires. Scores are publisher-reported by Arca, which publishes the benchmark and fields evaluated systems, so they are not independent measurements. Sample scores are not comparable to the full benchmark. Only 44 of 55 answer keys are specialist-written or reviewed; 11 rest on primary sources without expert review, and 6 November additions are not second-reviewed.
- Evaluates
- health
- Openness
- unknown
- Importer
- Model reports
LifeSci Diamond links
LifeSci Diamond results and reported scores
Reported benchmark scores · Model reports
Reported scores: 22
| Model | Organization | Reported value | Reported | Evidence |
|---|---|---|---|---|
| Arca Auto | Arca | 20.0 | 2026-11-04 | evidence |
| Arca with Claude Opus 5 | Arca | 9.1 | 2026-11-04 | evidence |
| Arca with Claude Opus 5.5 | Arca | 12.7 | 2026-11-04 | evidence |
| Arca with GPT-6 Astra | Arca | 29.1 | 2026-11-04 | evidence |
| Claude Fable 5 | Anthropic | 3.6 | 2026-11-04 | evidence |
| Claude Fable 5.1 | Anthropic | 1.8 | 2026-11-04 | evidence |
| Claude Opus 4.8 | Anthropic | 0.0 | 2026-11-04 | evidence |
| Claude Opus 5 | Anthropic | 1.8 | 2026-11-04 | evidence |
| Claude Opus 5.5 | Anthropic | 3.6 | 2026-11-04 | evidence |
| Claude Sonnet 5 | Anthropic | 1.8 | 2026-11-04 | evidence |
| Gemini 3.1 Pro Preview | 0.0 | 2026-11-04 | evidence | |
| Gemini 3.7 Flash | 1.8 | 2026-11-04 | evidence | |
| Gemini 4 Argon | 0.0 | 2026-11-04 | evidence | |
| GPT-5.6 Sol | OpenAI | 5.5 | 2026-11-04 | evidence |
| GPT-5.6 Terra | OpenAI | 1.8 | 2026-11-04 | evidence |
| GPT-6 Astra | OpenAI | 20.0 | 2026-11-04 | evidence |
| GPT-6.1 Sol | OpenAI | 14.5 | 2026-11-04 | evidence |
| GPT-6.1 Terra | OpenAI | 9.1 | 2026-11-04 | evidence |
| Grok 4.6 | xAI | 3.6 | 2026-11-04 | evidence |
| Kimi K3 | Moonshot AI | 1.8 | 2026-11-04 | evidence |
| Muse Spark 1.2 | Meta | 0.0 | 2026-11-04 | evidence |
| Muse Wonder | Meta | 9.1 | 2026-11-04 | evidence |
Scores are partitioned by the source that reported them and are never merged into a single cross-source ranking, because the sources measure different things and say so.