The most potent molecule is not automatically the most promising drug candidate. OpenAI’s MedChemBench, announced on 20 November 2026, gives life-science students and research teams a concrete reason: a molecule’s strength against its target can pull in a different direction from its permeability and breakdown.
In a historical example used by the benchmark, compound 23 blocked a protein at a lower concentration. Compound 24 crossed a cell barrier more readily and had lower reported clearance in a liver-enzyme assay. Neither result, alone, settles which compound deserves further investigation.
The useful lesson for AI-assisted research is a comparison, not a coronation. This analysis argues that an assistant should explain which properties fit the research goal, which conflict, and which remain unknown before it recommends a candidate.
What the two compounds actually show
These were not molecules invented by the new AI benchmark. Kevin Cottrell and colleagues’ 2024 chemistry paper describes them as part of research targeting PRMT5, a protein studied in cancer drug discovery. Table 3 reports four revealing comparisons:
Potency: compound 23’s cellular IC50 was 0.017 micromolar, versus 0.029 for 24. IC50 is the concentration needed for 50% inhibition of the measured activity; lower means more potent in that assay.
Permeability: 23 measured 11 and 24 measured 17, in units of 10⁻⁶ cm/s, across the tested cell layer.
Breakdown: human liver microsomal clearance was 11 for 23 and less than 10 for 24, in microlitres/minute/milligram. Lower clearance here indicates slower breakdown in that laboratory preparation.
Efflux: 24 also had a higher reported efflux ratio: 2.4 versus 1.0. Efflux can move compounds back out of cells, complicating the permeability picture.
These are laboratory measurements, not patient doses, human half-lives or proof of brain exposure. The results do not establish either compound as the better medicine.
Why “better” needs a research goal
The FDA’s explanation of drug development puts potency in a wider setting. Researchers also investigate how a compound is absorbed, distributed, metabolized and excreted, alongside toxicity, interactions and potential benefits. A strong effect in one test answers only part of that larger question.
For the two compounds, a potency-only ranking would put 23 first. A permeability-only ranking would reverse the order. That reversal is the point: the ranking changes because the question changes, not because one of the measurements must be wrong.
Nor can a sensible comparison simply count how many columns each molecule wins. As an illustrative decision principle, one unresolved property could matter more to a particular project than several favourable measurements. An assistant that combines everything into a single score without explaining those priorities hides the judgment the researcher needs to inspect.
Even a tiny bit of notation matters. “Less than 10” is not an exact value of 10. Replacing the bound with a neat number would make an AI-generated comparison look more precise than its source. Preserving that uncertainty is useful work, not a failure to give a decisive answer.
What the benchmark score tells you
OpenAI built MedChemBench from 99 historical problems. Its announcement reports 83.3% overall credit for GPT-6 Astra and 84.5% for GPT-6.2 Astra at the xhigh reasoning setting. Neither number is a drug-selection accuracy rate.
The paper’s detailed results for GPT-6 Astra show why separate skills matter. At xhigh, it earned 95.4% of structure credit and 91.1% of scientific-rationale credit, but 50.4% for ranking and selection and 47.0% for numerical prediction. Those component figures belong to GPT-6 Astra, not GPT-6.2 Astra.
The categories use different tasks and scoring rules, including partial credit. Earning about half the available points is not the same as making half the choices correctly, and does not establish chance-level performance.
Strong structure reconstruction is nevertheless valuable. As OpenAI’s examples illustrate, changing a ring or disconnecting fragments changes the molecule being discussed. Getting the identity right is a necessary foundation, even when deciding what to investigate next remains harder.
The study design also limits the conclusion: researchers averaged ten single-turn responses per task and model-effort condition, without tools or iterative experimental feedback. Prior exposure to the published sources cannot be ruled out. Historical reference answers may also miss workable alternatives. The evaluation does not establish improved laboratory decisions or a newly discovered medicine.
What to ask an AI comparison to explain
A proposed approach is to start with the project’s objective, then ask for a property-by-property account. Keep each compound’s identity attached to its measurements, assay conditions, units and source. Separate measured values from predictions, and leave missing information visibly missing. This is a way to organize a research discussion, not a validated candidate-selection protocol.
For this example, the useful follow-up is specific: what additional evidence would determine whether the permeability result outweighs the potency difference for the stated objective? The answer should expose the assumption behind a recommendation, not quietly supply an arbitrary weight for each property.
There is a role for software in assembling that evidence. Google’s 3 November Antigravity account describes a language model coordinating literature and database work while established chemistry tools perform calculations. That older demonstration is not a replication of MedChemBench or a head-to-head comparison. The guide to Antigravity’s chemistry research tools covers this evidence-gathering role.
The separate question of how new experiments change a proposal is covered in AI molecular discovery with laboratory feedback. Here, the immediate decision is narrower: what do the available properties justify investigating next?
A less potent molecule can remain interesting without becoming the winner. The most useful AI answer would make that distinction explicit—and show the scientist exactly which tradeoff still needs resolving.
