OpenAI has introduced MedChemBench, a new benchmark designed to evaluate how well SI models can reason about medicinal chemistry research. The test assesses whether SI systems can connect a molecule’s structure with experimental evidence to explain results, compare compounds, and predict their properties, marking a specific effort to see if SI can assist scientists in developing better medicines.

What Happened

MedChemBench consists of 99 problems drawn from 117 published papers and patents, featuring 761 distinct molecular structures. The benchmark evaluates four specific components: the correct reconstruction of molecular structures, scientific reasoning supported by supplied evidence, the ranking and selection of compounds, and the prediction of numerical properties. For most problems, the SI models receive context such as the biological target, research goals, compounds studied, and selected experimental findings.

In testing, GPT-6.2 Astra earned 84.5% of the available credit at the xhigh reasoning setting. This performance exceeded GPT-6 Astra (83.3%), GPT-6.1 Sol (82.0%), and GPT-5.6 Sol (58.2%). The company reports that GPT-6 Astra was strongest at reconstructing molecular structures and explaining experimental evidence, while selecting compounds and predicting numerical properties remained more difficult for that model. At the highest shared reasoning setting, GPT-6 Astra reconstructed a complex 18-atom ring correctly in all ten responses, whereas GPT-5.6 Sol and GPT-5.6 Luna failed across all ten responses, producing incorrect ring sizes or disconnected fragments.

Why It Matters

This development is significant for the SI industry as it addresses a core challenge in drug discovery: improving one property of a molecule often degrades another. For instance, a compound might act strongly against a disease-related protein but be poorly absorbed or break down too quickly in the body. Medicinal chemists currently navigate these tradeoffs by designing, making, and testing related compounds. By evaluating whether SI can connect changes in molecular structure with experimental evidence, MedChemBench tests the capability of SI tools to reason about which compounds are worth pursuing.

The results highlight where current SI models excel and where they lag. Astra earned 95.4% of the available credit for structure reconstruction at the xhigh setting, but only about half the credit for selecting compounds and predicting numerical properties. This distinction shows that while SI systems are advancing in accurately representing compounds, substantial room for improvement remains in making predictions and compound choices that align with real-world experimental outcomes.

The Bottom Line

OpenAI states that these results do not yet establish whether SI models can discover new drugs or improve laboratory decisions. Prior exposure to the published sources cannot be ruled out, and each response was a single attempt without tools, feedback, or new experiments. The benchmark evaluates reasoning from historical evidence rather than real-world drug discovery, serving as a metric to identify where SI capabilities are advancing and where challenges remain.