MedChemBench's vendor-reported 84.5% and OpenAI's Space access for Plus users
As OpenAI's Space reaches Plus users, our forecast is that MedChemBench still has no public dataset or evaluation-code link by 18 December 2026.
On 20 November 2026, OpenAI announced that GPT-6.2 Astra earned 84.5% of the available credit at xhigh effort on 99 medicinal chemistry problems. That reads like a strong credential. But the published breakdown for the earlier GPT-6 Astra at the same effort setting gives 95.37% for molecular structures and only 50.42% for ranking and selecting compounds. Numerical prediction earned 47.02%. The breakdown describes an earlier model, rather than GPT-6.2 Astra's performance by component. We would monitor independent selection and numerical prediction results for compound choices alongside the aggregate grade.
Imagine hiring a scientist whose employer sets the examination and announces the result, while keeping the questions from competing employers. That is the position a buyer faces here. We predict MedChemBench still has no public dataset or evaluation-code link by 18 December 2026 on OpenAI's announcement, its paper or the llm-stats listing. That is a 28-day window after the 20 November 2026 release. The reproduction gap is the interval in which a seller publishes a score before offering the questions or scoring machinery through its public release materials. A buyer has a credential, while the examiner controls the public terms of comparison.
The score is precise in a way the comparison is not. At xhigh reasoning effort, GPT-6 Astra earned 83.3%, leaving GPT-6.2 Astra ahead by 84.5 minus 83.3, or 1.2 percentage points. In the 9 November 2026 paper, GPT-6 Astra at max effort recorded 83.83%, with a 95% bootstrap interval from 81.27% to 86.20%. Those endpoints are 86.20 minus 81.27, or 4.93 points apart, for the older model at a different effort setting. That interval is no significance test for GPT-6.2 Astra: the paper contains no score for that model. It is a warning against turning a small apparent increment into a procurement premium without a comparable uncertainty test.
MedChemBench is substantial work. The research paper describes 61 model and effort combinations, ten responses per problem and 60,390 responses in all. Its 99 problems draw on 117 papers and patents and involve 761 distinct molecular structures. Six OpenAI authors tested eleven OpenAI models. They used deterministic checks for structure and numerical questions, and a separate GPT-6 Astra judge at medium effort for scientific rationales. The benchmark was first named on 3 June 2026, before this detailed release. A large, partly mechanical evaluation deserves attention. Yet the buyer still needs to know how the same questions treat a rival system.
The internal score also bundles different kinds of success. GPT-6 Astra at xhigh scored 91.05% on rationales, beside 95.37% on structures, 50.42% on selection and 47.02% on numerical prediction. Structure and selection therefore sit 95.37 minus 50.42, or 44.95 percentage points, apart on the same model and setting. Even the best explanation of a bad compound choice leaves the buyer with the bad choice. Our spending priority would follow the decision criteria, rather than the easiest high marks, until an independent medicinal chemist helps validate the grading.
OpenAI makes that distinction itself. Its 20 November 2026 post says, "These results do not establish whether models can discover new drugs or improve laboratory decisions." The sources were previously published; the company says earlier exposure remains possible. Each answer was a single attempt without tools, feedback or new experiments, and the paper calls the benchmark "close to saturation in its current form". Its authors ask for agreement with independent medicinal chemists and prospective cases. We value the retrospective score as a research signal, while refusing to convert it directly into laboratory savings.
An early public link is a genuine possibility, given OpenAI's own publishing record. At the 25 September 2025 GDPval launch, it released a 220-task gold subset and a public grading service while scoring Claude Opus 4.1, Gemini 2.5 Pro and Grok 4 alongside GPT-5, with Claude ahead. At the 12 May 2025 HealthBench launch, data and evaluation code appeared on GitHub. FutureHouse's LAB-Bench also offered a public dataset of more than 2,400 questions on 14 July 2024. Those are examples of immediate public access, well inside our 28-day horizon.
That opposing case matters most because openness serves the vendor too: outside comparisons make a benchmark more useful to customers and competitors. In five comparable launches, GDPval, FrontierScience, HealthBench, PaperBench and LAB-Bench, five of five scored models from rival vendors at release. MedChemBench starts on different ground. One day into the 28-day window, on 21 November 2026, its llm-stats listing marked both "Dataset SOON" and "Code SOON"; its four listed models were OpenAI's, with scores labelled self-reported and unverified. OpenAI's announcement and paper carried neither public release link. Keeping 99 questions private protects against memorization; it says little about withholding scoring code.
Publication speed should not be confused with independent evaluation speed. Across seven searched benchmark cases, zero independent outside-group scores appeared within 28 days; the scores for rivals at launch were run by benchmark owners. Google's MultiMedQA used existing public datasets when announced on 26 December 2022; the first outside score found came in Microsoft's GPT-4 paper on 20 March 2023, 84 days later. An independent HealthBench analysis arrived about 9.5 months after its launch and put 81.8% of physician disagreement variance at the case level. Those delays explain why our forecast targets the public link rather than a finished outside paper. Access can change before a research group publishes anything.
The same 20 November 2026 release notes brought OpenAI's Space to Plus users aged 18 and older, extending collaborative pages and presentations. Pages had arrived on 29 September 2026 and Slides on 16 October 2026 for Pro, Business and Enterprise. Space can update pages from connected apps, while Free accounts cannot open shared pages and public links remain unavailable. Neither product supplies evidence of the other's value, although both place access in the seller's hands.
A public dataset or code link would narrow the reproduction gap and change our access assessment to partial reproducibility. A full rerun still requires both questions and scoring instructions. A private access arrangement with a pharma company would sit beyond this public test. For a research buyer, the next expenditure should purchase independent selection and numerical prediction results, preferably on prospective compounds, rather than an extra fraction on a vendor's own grade. The indicator to watch is a public download or code repository, not a new marketing score.
A research buyer should monitor OpenAI's announcement, its paper and the llm-stats listing for a public MedChemBench dataset or evaluation code link, which would change the access assessment to partial reproducibility. Our forecast that MedChemBench still has no public dataset or evaluation-code link by 18 December 2026 would be wrong if any such link appears on OpenAI's announcement, its paper or the llm-stats listing by 18 December 2026.