OpenAI releases MedChemBench: GPT-6.2 Astra scores 84.5% but nails only half the key drug-design calls
rohanpaul_ai · x · 2026-11-21
OpenAI introduced MedChemBench, a benchmark testing whether AI models can read molecular structures, interpret experimental evidence, and decide which compounds to synthesize next.
- Problems drawn from 117 papers and patents, spanning 761 graded molecular structures, with researchers' later-reported results withheld
- Each prompt stops at a real decision point: biological target, compounds made so far, and selected assay results
- At xhigh reasoning, GPT-6.2 Astra earned 84.5%, vs 83.3% for GPT-6 Astra, 82.0% for GPT-6.1 Sol, and 58.2% for GPT-5.6 Sol
- Models are near-perfect at reading chemistry, but right only about half the time on the decision that actually matters: which compound to pursue next
Takeaway: AI understands medicinal chemistry well, but the key judgment calls still lag behind human chemists.
Related event: OpenAI Launches MedChemBench for AI Drug Chemistry Reasoning(4 posts)→
More from Research
- Vishal Sikka on reported Riemann Hypothesis proof with Claude: 'an incredible day' — vsikka · 2026-11-21
- Thinking Machines' 26B looped-MoE Inklings 2 Mini beats GLM-5.2 and Opus 4.5 on benchmarks — ziqiao_ma · 2026-11-21
- Bloomberg reporter claims AI may have cracked the Riemann Hypothesis — bgreene · 2026-11-21
- Interpretability competition launched to test logit lenses on Thinking Machines' Inkling — benno_krojer · 2026-11-21
- NDIF hosts Thinking Machines' Inkling 2 Mini and launches a Looped Lens interpretability challenge — benno_krojer · 2026-11-21
- OpenAI's GPT-Rosalind agents surface new drug discovery lead for neurodegenerative disease — OpenAI · 2026-11-21