OpenAI and Molecule.one demonstrate near-autonomous AI chemist improving medicinal chemistry reaction yields

OpenAI and Molecule.one published research on June 17, 2026 showing a near-autonomous AI system successfully improved a challenging reaction in medicinal chemistry — the first time an AI has tackled an open-ended synthetic chemistry problem from problem selection through to experimental verification.

What's new

The system combined GPT-5.4 with Molecule.one's Maria AI, which handles wet-lab execution. The AI selected the research area, generated proposals, rated them, and ran the experiments autonomously — with human chemists stepping in only to verify the results.

Key results:

• Tested 10,080 reactions under optimized conditions
• Yield improvements in 88% of boronic acids and 83% of sulfonamides tested
• Human chemists independently repeated 14 representative reactions; 11 showed higher yields, including 8 with more than twofold improvement
• The full autonomous process took about 2.5 months, plus another half month for human write-up

The target reaction — an optimization problem in fragment-based drug discovery — had resisted easy improvement for years. The AI identified conditions that human chemists had not tried systematically.

Context

This work follows OpenAI's earlier GPT-Rosalind initiative, which brought specialized life-science reasoning to drug discovery workflows. Molecule.one, the Czech-founded retrosynthesis startup, provides the Maria AI platform — a cloud lab environment where AI can propose, schedule, and execute reactions.

Prior autonomous chemistry demonstrations tended to limit the AI's role to suggesting conditions within a predefined search space. This work had the model define the problem itself.

Why it matters

Medicinal chemists routinely spend months optimizing reactions that determine whether a drug candidate can be manufactured at scale. An AI that can compress that timeline — and identify yield improvements that humans missed — has direct implications for drug discovery cost and speed.

The 2.5-month autonomous cycle, if generalizable, suggests AI could parallelize exploration across dozens of reaction classes simultaneously. OpenAI framed the result as a proof of concept, not a product launch, but the quantitative yield data and human-verified replication make it one of the more concrete AI-in-chemistry results published to date.