Benchmarks
Biopharma Bench is the benchmark we built. APEX-Agents, OfficeQA Pro, and SpreadsheetBench are public benchmarks we run through our harness.
Biopharma Bench
81 expert tasks across regulatory, CMC, quality, and clinical work, built on top of 24 real drug and device products. Graded against criteria drawn from the actual submissions and internal documents behind those products.
Leaderboard and method →1GPT-6 AstraCodex · high59.5%
2Claude Fable 5.1Claude Code · high58.9%
3GPT-6.1 SolCodex · high58.8%
Public benchmarks, our harness
| Model | Pass rate |
|---|---|
| Gemini 3 Flash | 48.3% |
| GPT-5.4 mini | 39.6% |
| GPT-5.4 nano | 35.9% |
| GPT-5.5 | 35.8% |
| GPT-5.4 | 35.6% |
| Gemini 3.1 Pro | 34.4% |
| Kimi K2 | 24.4% |
| Model | Pass rate |
|---|---|
| GPT-5.4 | 71.4% |
| Model | Pass rate |
|---|---|
| Gemini 3.1 Pro | 88.0% |
| GPT-5.4 | 81.9% |
Related research
Workplace evaluation methods
EvalScience publishes methods for how the tested setting changes what a score means, and how to construct knowledge-work benchmarks. Biopharma Bench uses those construction checks.