Report 249 · AI in the Lab
What the AI chemist's yield gain measures
In June 2026 OpenAI announced that GPT-5.4, wired to a robotic chemistry lab run by the startup Molecule.one, proposed an additive that improves a notoriously stubborn reaction used to make drug molecules. Across 10,080 reactions, the mean yield rose from 16.6% to 25.2%. The chemistry looks solid and the write-up is unusually candid. But the headline number is an estimate from a detector model, the additive was first tried in this reaction in 1998, and the chemistry paper itself never mentions the AI.
The reaction, in one paragraph
Chan–Lam coupling uses a cheap copper catalyst to join a boronic acid to a nitrogen atom, making the carbon-nitrogen bonds that show up all over medicinal chemistry. It runs in air and tolerates moisture, which is why chemists like it. Its weak spot is primary sulfonamides (R–SO₂NH₂), a group found in many drugs: the sulfonyl group makes the nitrogen a poor nucleophile, and the boronic acid tends to decompose before it couples. One of those decomposition routes, oxidative deboronation, turns the boronic acid into a phenol, which then eats more boronic acid. Yields stay low.
What OpenAI says happened
Per OpenAI's post, the project ran from a first prompt on 4 March 2026 to sharing results with outside experts on 4 June. GPT-5.4, running inside a harness with prompts written by Molecule.one scientists, generated and ranked thousands of research proposals for improving one of several reaction classes. Human chemists reviewed the top-ranked few and picked four to test. Molecule.one's agent, Maria, turned those into lab instructions and ran them on a high-throughput robotic platform. The winning proposal, labeled OAI-M1-03, picked primary sulfonamides as the target and suggested mild oxidants, including TEMPO (a stable, shelf-bought radical best known from alcohol oxidation), as an additive.
Of the other three proposals, OpenAI says one was disproven and two were "experimentally proven," with analysis still ongoing. So the scorecard so far is one written-up result out of four tested ideas, picked by people from thousands generated.
What the 16.6% to 25.2% actually is
The preprint describes two screening campaigns: 5,088 reactions testing ten oxidants under 53 condition sets, then 4,992 reactions optimizing TEMPO loading, copper, base, solvent and temperature. Each campaign crossed 12 sulfonamides with 8 boronic acids, giving 96 substrate pairs. Every reaction ran in a 96-well plate at 28.6 microliters.
Under the best condition (2 equivalents of TEMPO, 20 mol% copper acetate), the mean estimated product yield across those pairs was 25.2%, against 16.6% for the same setup with no oxidant. The share of reactions clearing 30% rose from 15.6% to 37.5%. That is a real improvement on a hard reaction. It also means, by simple subtraction, that about 62% of reactions under the best condition still came in at 30% or below. Better, not solved.
The word that matters is estimated. This is the part I can speak to from the instrument side. The yields came from LC-PDA-MS: separate the mixture by liquid chromatography, read each peak's UV absorbance with a photodiode array, and compare it with an internal standard. The catch is that different molecules absorb UV light very differently, so a peak's area doesn't convert to an amount unless you know that specific product's response. Making and calibrating a pure standard for every product across 96 pairs isn't practical at this scale, so the team corrected each peak using a DFT-predicted absorption profile for that product, a computed estimate of how strongly it should absorb. That's a reasonable choice for ranking conditions against each other. But it means every absolute yield in the screen carries the error of a quantum-chemistry prediction, and the screen's numbers are best read as relative.
The bench checks, read closely
OpenAI's post says human chemists "reproduced representative reactions manually at bench scale" and saw higher yields for 11 of 14 substrate pairs. The preprint shows those 14 came from two different tests.
The first four were chosen, in the preprint's words, as pairs "that achieved relatively high HTE yields for TEMPO-containing conditions." In other words, they were selected because TEMPO had already looked good on them. They also didn't use the same copper loading in both arms: the TEMPO runs used 10 mol% copper and the no-oxidant controls used 40 mol%, each set to its best screening condition. Three of four improved; in the fourth, the no-oxidant control actually did slightly better (product-to-standard ratio 2.60 versus 2.12). These are the only bench runs with qNMR yields, a more direct measurement, and the gains there are large: 10% to 33%, 27% to 59%, and 36% to 99%.
The next ten are the stronger test. Ten boronic acids, nine of them never used in the screens, each paired with one sulfonamide, same 20 mol% copper in both arms. Eight improved, with the clearest gains on electron-poor boronic acids. Two didn't, and in one the TEMPO run made noticeably less product (ratio 0.56 down to 0.13). These are reported as product-to-standard ratios, not yields.
So "11 of 14" is accurate, but the most convincing slice is 8 of 10 on fresh substrates, measured as ratios, and the biggest yield numbers come from pairs picked because they already worked.
The idea is older than the AI
Here is the detail the "AI found an overlooked additive" framing skips, and it comes from the preprint itself. TEMPO was in the original 1998 Chan–Lam paper by Patrick Lam and colleagues, where, the preprint says, it "only occasionally outperformed molecular oxygen" across a small number of examples. Air won by default and TEMPO faded. The preprint found only two later studies that deliberately used TEMPO in Chan–Lam work. The screen's starting conditions came from a 3,904-reaction dataset on this exact sulfonamide problem published in 2025 by Abigail Doyle's group.
None of that diminishes the result. Spotting a lead that a field dropped 28 years ago and then testing it systematically, across 96 pairs and ten competing oxidants, is useful science, and the preprint makes a narrower and well-supported novelty claim: the first broad evaluation of TEMPO in Chan–Lam coupling and the first TEMPO protocol built for primary sulfonamides. The side-product data are a nice touch too: TEMPO was the one oxidant that raised product yield while lowering oxidative deboronation, which hints at why it works (the authors label this a hypothesis needing mechanistic study). But "connected two neglected corners of the literature" is a more accurate description than "discovered."
The paper never mentions the AI
This one surprised me. The chemistry preprint, "TEMPO Improves Generality and Decreases Oxidative Deboronation in Chan–Lam Couplings of Primary Sulfonamides," lists authors from Molecule.one and OpenAI. I searched the text: it does not mention GPT-5.4, a language model, or AI anywhere. It reads as a conventional methods paper ("We therefore asked whether mild redox additives could improve the balance..."). The account of what the model proposed and what the humans did lives in OpenAI's blog post and a separate, rewritten chain-of-thought document, which I did not read.
That splits the claim in two. The chemistry claim (TEMPO helps this reaction) rests on data in the preprint that other labs can test. The AI claim (the model generated the idea) rests on the company's own description of its process. Both may be true. They just have different kinds of evidence behind them. And as of my reading, the preprint is a PDF hosted on OpenAI's site. OpenAI says four outside experts reviewed it; that's not journal peer review.
Where the humans were
Give OpenAI credit for spelling this out. People wrote the steering and grading prompts, chose which four of thousands of proposals reached the lab, made corrections to the experimental plans (the biggest: dropping DMSO as a solvent because it might react with the stronger oxidants), helped prepare reagents and consumables, and repeated key reactions by hand. The company's own limitations section says it plainly:
It does not show that AI can independently run a chemistry research program from end to end.
That's the right sentence, and it's the one to remember when the coverage says "near-autonomous." The pattern matches what I wrote about in why AI scientists fail: the systems that work have people choosing, correcting and verifying at the steps that matter.
How I'd read it
A frontier model plus a robotic lab surfaced a dropped lead, a high-throughput platform tested it at a scale no human team would bother with, and chemists confirmed it on most substrates they tried by hand. That's a credible, modest advance in one reaction, and a good example of AI used as a literature-and-search engine inside a tightly supervised loop. It is not an AI chemist discovering new chemistry on its own, the screen's yields are model-corrected estimates, and the most striking bench numbers come from hand-picked pairs.
What would change my read: an independent lab reproducing the TEMPO effect on its own substrates with isolated yields, a mechanism study, or a peer-reviewed version of the paper. If one lands, I'll update this post in place.
Sources
- Jan Rzymkowski, Shuyuan Zhang, Artur Chołuj, Aleksander Szkółka, Mateja Dud, Mateusz Bruno-Kamiński, Jan Busz, Michał Sadowski et al. (Molecule.one and OpenAI), "TEMPO Improves Generality and Decreases Oxidative Deboronation in Chan–Lam Couplings of Primary Sulfonamides" (preprint PDF hosted by OpenAI, linked from its 17 June 2026 post). (Primary source. Full 20-page PDF read; the supporting information was not read. Source of the two campaigns, the 96 substrate pairs, well volume, the LC-PDA-MS method with DFT-predicted absorption correction, the yields and thresholds, Tables 1 and 2, the copper loadings, the selection criterion for the first four bench pairs, the history of TEMPO in Chan–Lam coupling, the novelty claims, the deboronation hypothesis and the quoted phrases. Not peer reviewed as of reading.)
- OpenAI, "A near-autonomous AI chemist improves a challenging reaction in medicinal chemistry," 17 June 2026. (Read in full. Source of the workflow, the 4 March to 4 June timeline, the human roles and the DMSO correction, the four-proposal scorecard, the "11 of 14" summary, the four external reviewers, and the quoted limitations sentence.)
- P. Y. S. Lam, C. G. Clark, S. Saubern, J. Adams et al., "New aryl/heteroaryl C–N bond cross-coupling reactions via arylboronic acid/cupric acetate arylation," Tetrahedron Letters 39, 2941–2944 (1998). (Not read; bibliographic details checked on Crossref. Cited for the preprint's account of TEMPO's role in the original study.)
- S. S. Gandhi, G. Z. Brown et al. (Doyle group), "Data Science-Driven Discovery of Optimal Conditions and a Condition-Selection Model for the Chan–Lam Coupling of Primary Sulfonamides," ACS Catalysis (2025). (Not read; bibliographic details checked on Crossref. Cited for the preprint's statement that its baseline conditions came from this 3,904-reaction dataset.)
"About 62% at or below 30%" is my subtraction from the preprint's 37.5%. The explanation of why UV peak areas need a per-compound response correction is my own, from general analytical-chemistry practice; the preprint states only that it corrected with DFT-predicted absorption profiles. My search for AI terms covers the preprint's main text and references, not the supporting information.