The Information Machine
FollowingDay 2in the edition since 21 Nov 20267 sourcesupdated 21 Nov 2026

OpenAI's MedChemBench and Rosalind model series

The gist

Latest: GPT-6.2 Astra Leads MedChemBench at 84% Overall, 50% on Key Decisions

MedChemBench provides a publicly grounded measure of where AI stands on the actual judgment calls in early drug discovery, and the gap between overall accuracy and decision-point accuracy shows the limits of current models on the hardest tasks. The Rosalind line and associated benchmarks represent OpenAI's sustained push into life sciences research workflows.

When stories like this move, I write about them in an email each morning.

One free email each morning, around 8 AM US Eastern.

The full picture

MedChemBench, OpenAI's medicinal chemistry benchmark, is drawn from 117 papers and patents covering 761 graded molecular structures, with each prompt stopping at a real decision point and providing the biological target, prior compounds, and selected assay results. The benchmark withholds results that researchers reported later, testing genuine predictive ability. GPT-6.2 Astra is the top-performing model at roughly 84% overall accuracy, described as near-perfect at reading chemistry but correct only about half the time on the decision that actually matters.

OpenAI has released a series of Rosalind-branded models alongside the benchmark. GPT-Rosalind-5.5, designed specifically for medicinal chemistry, scores 27.5% on MedChemBench versus GPT-5.5's 25.1% while using 7.2% fewer tokens. GPT-6 Rosalind, priced at $15/$75 per million tokens, raised MedChemBench v7 scores from 28.9% to 49.6% at the highest reasoning budget compared to GPT-Rosalind-5.5, improved GeneBench Pro v13 scores from 16.7% to 37.7%, and raised LifeSciBench Gold scores from 44.0% to 61.3%, while also increasing the refusal rate on a biorisk safety test. LLM Stats tracks four OpenAI models on MedChemBench; GPT-5.6 Sol leads that leaderboard at 0.483 on a 0-to-1 scale, with the field average at 0.4.

OpenAI also developed LifeSciBench, an externally expert-judged benchmark covering six workflow areas in life sciences research, described as taking an end-to-end view of scientifically valuable work rather than evaluating single components in isolation.

In a separate application, GPT-Rosalind working with Perturb-AI and Scripps Research analyzed gene-expression data from millions of mouse brain cells in 32 minutes and surfaced a hypothesis that disrupted cholesterol transport may drive neurodegeneration linked to the DCTN1 gene, which is associated with Perry syndrome. That collaboration is described as expanding to identify new drug targets for neurodegenerative diseases.

How it developed
First in the edition · 21 November 2026
20 November 2026in the 21 Nov edition

OpenAI released MedChemBench, drawn from 117 papers and patents covering 761 graded molecular structures, testing real decision points in medicinal chemistry.

GPT-6.2 Astra leads at roughly 84% overall but correct only about half the time on its critical decisions, while GPT-6 Rosalind raised the top MedChemBench v7 score from 28.9% to 49.6%. GPT-Rosalind with Perturb-AI and Scripps Research analyzed millions of mouse brain cells in 32 minutes, surfacing a hypothesis that disrupted cholesterol transport may drive DCTN1-associated neurodegeneration.

1 November 2026

GPT-6 Rosalind launched at $15/$75 per million tokens with improved scores across MedChemBench v7, GeneBench Pro v13, and LifeSciBench Gold

11 July 2026

LLM Stats leaderboard for MedChemBench published, tracking four OpenAI models with GPT-5.6 Sol leading at 0.483

Sources
2 more sources

Get the daily email

When stories like this move, I write about them in an email each morning.

One free email each morning, around 8 AM US Eastern.

See the latest edition · Privacy