
What happened
OpenAI introduced MedChemBench, a benchmark built from 99 problems, 117 papers and patents, and 761 molecular structures. At the xhigh setting, GPT‑6.2 Astra earned 84.5% of available credit, ahead of GPT‑6 Astra at 83.3%, GPT‑6.1 Sol at 82.0%, and GPT‑5.6 Sol at 58.2%.
Why it matters
The scores are a first public read on how well the current GPT models connect a molecule's structure with experimental evidence, compared with each other. OpenAI itself cautions that these results do not establish whether models can discover new drugs or improve laboratory decisions.
WHO IT HITSMedicinal chemists and the data teams that support them get a common yardstick for how much of their reasoning work current AI models can reliably handle — with OpenAI's own caveat that the benchmark does not prove any real drug-discovery gains.
Summaries like this, every morning.
OpenAI frames MedChemBench as a response to a specific difficulty in medicinal chemistry: making one property of a molecule better can make another worse, such as stronger activity against a disease-related protein coming with poorer absorption or faster breakdown in the body. The benchmark therefore supplies the context a chemist would have at a chosen point in a research program — the biological target, the researchers' goals, the compounds studied so far, and selected experimental findings — and asks models to carry the reasoning forward.
The results OpenAI reports are graded per component rather than as one pass-or-fail number. Astra's clearest strengths were reconstructing molecular structures and explaining the scientific evidence. When a problem required assembling a molecule while preserving an 18-atom ring, GPT‑6 Astra reconstructed the compound correctly in all ten responses, while GPT‑5.6 Sol and GPT‑5.6 Luna failed across all ten; the illustrated answers showed a 15-atom ring and two disconnected fragments, respectively, errors that change the identity of the molecule being analyzed. Selection and prediction, by contrast, still leave substantial room to improve.
OpenAI is explicit about the limits of the exercise. The results do not establish whether 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, so its value appears to sit in separating four abilities — reconstructing molecules, explaining findings, selecting compounds, and predicting measurements — and showing where errors remain.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
Pharmaceutical Technology published an item titled "AI and the digitalisation of pharma manufacturing", but th…

At the Open Source AI for Science Summit, Meta's FAIR team said the Digital Brain Project, run with the Rothsc…

Details have emerged about the personnel at Prometheus, a physical AI company co-run by Amazon founder Jeff Be…

AgeAI, led by CEO Gen Mikami, joined the Gotanda Valley Acceleration Program 2026 from its September 17 kickof…

ConvergePartners said it agreed to adopt Paravellum's data-infrastructure technology to build a shared foundat…

Boltz raised $110 million in a Series A led by Bessemer, with backers including Greycroft and Andreessen Horow…