On March 4, 2026, a chemistry AI received a single open-ended instruction: pick one of several important reactions in medicinal chemistry and make it better. No specific molecule. No predetermined answer. Just a hard, real problem and access to a robotic laboratory.
Three months later, on June 4, the system handed its best idea to a panel of outside chemists. The proposal, logged internally as OAI-M1-03, targeted a reaction that has frustrated drug chemists since it was first reported in 1998. On June 17, OpenAI published the result from its three-month collaboration with the Polish chemistry startup Molecule.one.
What makes the announcement different from the usual "AI in science" press cycle is narrow and specific. This is the first publicly documented case of a frontier AI model acting as a near-autonomous agent inside a real wet-lab workflow, proposing an unexpected hypothesis, helping run the experiments to test it, and arriving at a finding that independent experts confirmed was worth publishing. Chemistry does not grade on a curve. The yields either went up or they did not.
They went up.
The Reaction That Resisted Chemists for Decades
The target was Chan-Lam coupling, a copper-catalyzed method for forming carbon-nitrogen bonds, the kind of bond that appears throughout modern medicines. Chemists like it because it uses cheap copper catalysts and tolerates air and moisture, unlike the fussier palladium-based alternatives.
It has one stubborn weakness. When chemists try to couple primary sulfonamides, a chemical group found in more than 91 FDA-approved drugs across cancer, antimicrobial, and heart medicines, the reaction tends to fail. The sulfonamide's electron-pulling structure slows a key step in the copper cycle, side reactions eat up the starting material, and yields stay low. That makes a whole class of useful drug molecules harder to build, which is a real bottleneck in early drug discovery.
GPT-5.4 zeroed in on that exact substrate class as a high-value problem. Then it proposed something chemists had largely missed: a mild oxidant called TEMPO, a stable organic radical that has been sold for decades and used mostly in alcohol chemistry, could speed up the sluggish step and rescue the reaction. TEMPO had appeared in the Chan-Lam literature only in scattered, isolated contexts. Its systematic use for this difficult sulfonamide problem was not established practice. The model made the connection between two adjacent corners of the literature, and the lab tested whether the connection held.
How GPT-5.4 and a Robotic Lab Ran the Loop
OpenAI connected GPT-5.4 to Maria, Molecule.one's agentic chemistry system, which is wired into a microliter-scale, high-throughput experimentation lab. The setup ran as a structured loop rather than a single query.
The cycle worked like this:
- Molecule.one scientists wrote the steering and grading prompts that defined the goal and standards.
- GPT-5.4 generated and ranked thousands of research proposals.
- Human chemists reviewed only the top-ranked proposals and picked four to send to the lab.
- Maria translated each approved proposal into precise experimental instructions and ran the reactions.
- The system analyzed the results and proposed the next round of experiments.
This is the same agentic pattern, plan, act, observe, and revise, that now drives the most capable AI coding and research agents, pointed at physical chemistry instead of software.
Humans stayed in the loop the entire time. They designed the prompts, chose which proposals reached the bench, corrected experimental details (including a deliberate decision to exclude one solvent that might react badly with the oxidants under test), prepared reagents, and independently repeated key reactions by hand to check the machine's work. What made the process near-autonomous was not the absence of people. It was the origin of the scientific ideas. The model proposed the research area, named the high-value substrate class, and suggested the oxidant. The people steered and verified.
What 10,080 Reactions Actually Showed
Maria ran 10,080 reactions across two experimental cycles, a volume that would take a bench chemist running three experiments a day roughly a decade to match. The scale was the point. A reaction that looks great on a handful of test cases often collapses across a broader set, so the large count let the system find TEMPO among ten candidate oxidants, watch the effect repeat across many substrate combinations, and map where it worked and where it did not.
| Metric | Before | After TEMPO |
|---|---|---|
| Mean reaction yield | 16.6% | 25.2% |
| Share of reactions clearing 30% yield | 15.6% | 37.5% |
| Boronic acids that improved | — | 88% |
| Primary sulfonamides that improved | — | 83% |
Then came the test that matters most. Tiny microliter reactions can produce results that vanish at normal scale, so human chemists repeated 14 representative substrate pairs at standard bench scale. Yields improved in 11 of the 14 pairs, and 8 of those more than doubled. A second cycle turned up a practical bonus: a cheaper structural cousin of TEMPO, called 4-hydroxy-TEMPO, delivered nearly the same benefit at lower cost, which matters for any lab that wants to use the method at scale.
Tim Cernak, Associate Professor of Medicinal Chemistry at the University of Michigan and one of four outside experts who reviewed the preprint, was direct about the significance.
"The merger of high throughput experimentation and modern AI represents a new frontier of scientific discovery. This new reaction is a powerful demonstration, where exceptionally mild conditions and a practical oxidant enable a nicely general substrate scope for one of the more popular reactions in drug synthesis." — Tim Cernak, Associate Professor of Medicinal Chemistry, University of Michigan (via Tech Times, Jun 18, 2026)
Why This Lands Differently for Practitioners
For machine learning practitioners, the interesting part is not the chemistry. It is the shape of the system.
Most AI tools in chemistry until now predicted molecular properties, suggested synthesis routes, or screened virtual compound libraries, all in software. This system did something categorically different. It ran real physical reactions, learned from real experimental data, and decided what to try next inside a closed loop that touched actual lab hardware. The line being crossed is between AI that simulates chemistry and AI that participates in it.
It also fits a pattern OpenAI has been building all year. In February, working with Ginkgo Bioworks, the company connected GPT-5 to a cloud lab and ran more than 36,000 experimental conditions to cut the cost of cell-free protein synthesis by 40 percent. In April it launched GPT-Rosalind, a life-sciences reasoning model. This week it also introduced LifeSciBench, a benchmark where the best AI passed only 36 percent of expert science tasks, a reminder that autonomous science is still early. The Chan-Lam result used the general-purpose GPT-5.4, not the specialized Rosalind model, which suggests the frontier models can contribute to real domain science when paired with the right infrastructure and expert oversight. That same thesis is fueling billions in AI drug-discovery investment, from Isomorphic Labs to OpenAI's own pharma partnerships.
The Other Side: A Preprint, Not a Cure
OpenAI was unusually careful about what the result does not show, and the caveats are real.
The work is a preprint, not yet formally peer-reviewed. Four external experts supported its novelty, but independent replication by labs not involved in the project has not happened yet, and the mechanism behind TEMPO's improvement is not fully characterized. The finding covers one reaction class on one experimental platform. It does not automatically generalize to other couplings, other substrates, or manufacturing-scale conditions. Bench validation covered 14 substrate pairs, not the full chemical universe.
The autonomy has limits too. Human judgment was essential at multiple points, and the whole thing depended on specialized high-throughput lab infrastructure that very few organizations have. This is not a chatbot inventing drugs from a prompt. It is a tightly supervised collaboration that happened to let the model own the key scientific ideas.
And the broader timeline has not moved. A typical drug program still takes 10 to 15 years and costs an estimated $2.8 billion from target to approval, and no AI-designed drug has yet cleared a large clinical trial. A better way to run one reaction is a real improvement to one slow step, not a shortcut around the whole pipeline.
The Bottom Line
A frontier model was handed a vague goal and a robotic lab, and it produced a specific, unexpected, experimentally validated chemistry result that human experts called worth publishing. Strip away the hype and that is what happened. The model did not just summarize known chemistry. It proposed a hypothesis that working chemists had overlooked, and 10,080 reactions later, the idea held.
The honest framing is that this is one narrow win that now has to survive independent replication, mechanism studies, and real adoption in discovery labs, all measured in years. The provocative framing is harder to dismiss. If an AI can reliably connect scattered corners of the scientific literature into hypotheses that bench experiments confirm, the bottleneck in science stops being ideas and becomes how fast we can test them.
Chemistry does not grade on a curve. This time, the machine's idea passed.
Sources
- AI Drug Discovery Chemistry Hits Wet Lab: GPT-5.4 Boosts Chan-Lam Yields in 10,080 Reactions — Tech Times (Jun 18, 2026)
- A near-autonomous AI chemist improves a challenging reaction (Official) — OpenAI (Jun 17, 2026)
- OpenAI on the Chan-Lam coupling result — OpenAI on X (Jun 17, 2026)
- OpenAI's AI Chemist Improves Drug Discovery Chemistry — Techs Current (Jun 2026)
- Ginkgo Bioworks autonomous laboratory driven by OpenAI's GPT-5 — PR Newswire (Feb 2026)
- Chan-Lam coupling (background) — Wikipedia