OpenAI Releases MedChemBench; GPT-6.2 Astra Said to Reconstruct Molecules in One Shot Without Tools

heypearlai · x · 2026-11-22

The post claims OpenAI released MedChemBench, a drug-chemistry benchmark built from 761 real molecular structures drawn from 117 papers and patents.

GPT-6.2 Astra is said to stand out on it: reconstructing molecules and predicting their properties in single attempts without any tools, rather than being fed step-by-step instructions. The author frames this as AI shifting from text generator to scientific instrument — if a model can read a raw chemical structure and understand what it is, how it behaves and what it binds to, it changes the economics of early-stage drug discovery, compressing years and millions of dollars of screening into minutes.

The post also claims the model jumped to #1 on the RoboColosseum leaderboard. These are the poster's claims; the model name and scores are not independently verified.

Related event: OpenAI Launches MedChemBench for AI Drug Chemistry Reasoning(6 posts)→

Original post →

More from Models

Models channel →