Atlas

Benchmarks

MedChemBench

OpenAI's MedChemBench (November 20, 2026) poses 99 medicinal-chemistry problems built from 117 recent papers and patents (median publication year 2025), each giving the biological context, compound structures and experimental evidence (structure drawings, assay tables, plots, protein–ligand views) up to a decision point in a drug program; a model answers each in one turn, without tools. Answers reconstruct molecules as SMILES (761 distinct graded structures), explain results, rank or select compounds and predict measured properties. Deterministic checks (RDKit-canonical isomeric SMILES matches, pairwise ranking and Jaccard set credit, fold-error credit for numbers) and a GPT-6 Astra judge at medium effort for rationale criteria give partial credit, and the score is the share of each task's rubric budget earned, averaged over ten responses per task (completed invalid responses score zero) and then equally over the 99 tasks. OpenAI built it and reports only its own GPT models. It is not the 'MedChemBench (Internal)' of OpenAI's launch posts (medchembench-openai-internal-reported), whose values differ.

Models
12
Top score
84.5
Leader
GPT-6.2 Astra
Updated
Nov 20, 2026
Resolution
~1 mo
Sharp
Saturation
~Feb 2027
Projected
Signal
93rdpercentile
141 benchmarks from 0.8 to 2.9196 benchmarks from 2.9 to 5.0131 benchmarks from 5.0 to 7.1105 benchmarks from 7.1 to 9.282 benchmarks from 9.2 to 11.365 benchmarks from 11.3 to 13.327 benchmarks from 13.3 to 15.421 benchmarks from 15.4 to 17.58 benchmarks from 17.5 to 19.68 benchmarks from 19.6 to 21.76 benchmarks from 21.7 to 23.84 benchmarks from 23.8 to 25.95 benchmarks from 25.9 to 28.03 benchmarks from 28.0 to 30.12 benchmarks from 30.1 to 32.14 benchmarks from 32.1 to 34.21 benchmark from 34.2 to 36.33 benchmarks from 36.3 to 38.41 benchmark from 38.4 to 40.52 benchmarks from 40.5 to 42.61 benchmark from 42.6 to 44.71 benchmark from 44.7 to 46.81 benchmark from 48.9 to 50.9

57th strongest of 818

Signal-to-noise +16.94

Difficulty
66thpercentile
1 benchmark from 5.8 to 14.71 benchmark from 23.5 to 32.42 benchmarks from 32.4 to 41.21 benchmark from 41.2 to 50.12 benchmarks from 50.1 to 58.99 benchmarks from 58.9 to 67.88 benchmarks from 67.8 to 76.718 benchmarks from 76.7 to 85.523 benchmarks from 85.5 to 94.441 benchmarks from 94.4 to 103.232 benchmarks from 103.2 to 112.141 benchmarks from 112.1 to 121.049 benchmarks from 121.0 to 129.875 benchmarks from 129.8 to 138.760 benchmarks from 138.7 to 147.5122 benchmarks from 147.5 to 156.4135 benchmarks from 156.4 to 165.3124 benchmarks from 165.3 to 174.139 benchmarks from 174.1 to 183.021 benchmarks from 183.0 to 191.88 benchmarks from 191.8 to 200.72 benchmarks from 200.7 to 209.54 benchmarks from 209.5 to 218.4

278th hardest of 818

Capability needed 159.6

Organizations
12 dated models · 6 frontier modelsClosed weightsOpen weights
Data: OpenAIatlas.village.dev

MedChemBench leaderboard

Related benchmarks

Most correlated

Benchmarks outside its family whose model rankings track it most reliably (Spearman ρ over at least 10 shared models).

Loading Atlas data…