OpenAI is testing whether artificial intelligence can help scientists identify promising drug candidates.
The company on Friday unveiled MedChemBench, a test designed to evaluate whether AI models can interpret molecular structures, analyze experimental results and determine which compounds researchers should investigate next.
The benchmark is built around 99 problems drawn from published research and is designed to replicate the decisions medicinal chemists face when developing new drugs. Improving one characteristic of a compound, such as its ability to target a disease-related protein, can come at the expense of other properties, including how well it is absorbed or how quickly it breaks down in the body.
MedChemBench draws on 117 research papers and patents and covers 761 distinct molecular structures. Each problem gives models information that researchers might have at a particular stage of drug development, including the biological target, research objectives, previously tested compounds and selected experimental results.
OpenAI evaluates models across four areas: reconstructing molecular structures, explaining scientific evidence, ranking and selecting compounds, and predicting molecular properties based on experimental measurements.
How OpenAI’s AI Models Performed
At the highest shared reasoning setting, GPT-6.2 Astra earned 84.5% of available credit, followed by GPT-6 Astra at 83.3%, GPT-6.1 Sol at 82% and GPT-5.6 Sol at 58.2%.
But the results also revealed persistent weaknesses. While GPT-6 Astra excelled at reconstructing molecular structures and explaining experimental findings, selecting the best compounds and predicting their properties proved more challenging.
A problem involving PRMT5 inhibitors, which are being studied for cancer treatments, illustrates the difficulty.
Models were asked to predict how three related compounds would perform across several measures, including their ability to inhibit PRMT5 in cells, cross cell barriers and resist breakdown by liver enzymes.
Follow-up experiments, withheld from the models, revealed trade-offs: One compound showed stronger PRMT5 activity, while another crossed cell barriers more easily and resisted breakdown longer.
The results underscore a fundamental challenge in drug development: Strength in one area does not necessarily make a compound the best overall candidate.
OpenAI also tested whether models could accurately interpret molecular diagrams, including atom connections, ring structures and stereochemistry.
In one example, GPT-6 Astra correctly reconstructed a molecule with an 18-atom ring in all 10 attempts at the highest shared reasoning setting. GPT-5.6 Sol and GPT-5.6 Luna failed all 10, producing structures with incorrect ring sizes or disconnected molecular fragments.
AI Drug Discovery Still Has Limits
Despite the results, OpenAI cautioned against treating the benchmark as evidence that AI can independently discover new medicines.
The company said the evaluation measures reasoning against historical research rather than real-world drug discovery. It also acknowledged that models may have previously encountered some of the published source material and that each response was a single attempt without access to tools, feedback or additional experiments.
The benchmark is intended to identify where AI models are improving and where further progress is needed before they can become more reliable tools for medicinal chemistry research.
Photo Courtesy: Evolf on Shutterstock.com
This content was partially produced with the help of AI tools and was reviewed and published by Benzinga editors.
Market News and Data brought to you by Benzinga APIs© 2026 Benzinga.com. Benzinga does not provide investment advice. All rights reserved.





