OpenAI’s GPT-6.2 Astra Scores 84.5% on Drug-Chemistry Test

OpenAI’s GPT-6.2 Astra earned 84.5% of the available credit on MedChemBench, a medicinal-chemistry benchmark OpenAI released on Friday. GPT-5.6 Sol scored 58.2% at the same setting.

The 26.3-point gap, by our arithmetic, comes from OpenAI’s own tests of its own models, at the xhigh reasoning setting.

Business Pill: The Cost of Checking

An evergreen Business Engineer explainer on the cost of checking work. It teaches the general idea only and does not discuss MedChemBench.

The key insight: On OpenAI’s new 99-problem medicinal-chemistry benchmark, GPT-6.2 Astra earned 84.5% of the credit against 58.2% for GPT-5.6 Sol, and GPT-6 Astra rebuilt molecules at 95.4% but earned only about half the credit for choosing compounds and predicting properties.

What the Test Asks

MedChemBench has 99 problems drawn from 117 papers and patents, with 761 distinct molecular structures, OpenAI said.

For most problems, models get the context available at a chosen point in a research program: the biological target, the goals, the compounds tested so far and selected results. They must reconstruct molecules, explain results, rank compounds and predict measured properties.

One problem uses research on PRMT5, a protein studied in cancer drug discovery. The model must predict how three related compounds block the protein, cross a cell barrier and break down in a liver-enzyme assay.

MedChemBench scores reported by OpenAI for its own models at the xhigh reasoning setting.
MedChemBench scores reported by OpenAI for its own models at the xhigh reasoning setting.

How the Models Scored

At xhigh, GPT-6.2 Astra scored 84.5%, GPT-6 Astra 83.3%, GPT-6.1 Sol 82.0%, GPT-6.1 Terra 75.1% and GPT-5.6 Sol 58.2%, OpenAI reported. At max effort, GPT-6 Astra reached 83.8%.

OpenAI evaluated eleven GPT models in all. Higher reasoning settings improved scores, “while large differences between models remained”, it wrote.

Strong on Structures, Weaker on Choices

GPT-6 Astra earned 95.4% of the available credit for reconstructing molecular structures. It earned “only about half the credit” for selecting compounds and predicting numerical properties, OpenAI said.

In one structure task, Astra rebuilt a molecule with an 18-atom ring correctly in all ten responses. GPT-5.6 Sol and GPT-5.6 Luna failed in all ten.

Bar chart: GPT-6 Astra earned 95.4% of structure credit but about half the credit for selecting compounds and predicting properties on MedChemBench
Where GPT-6 Astra scored on MedChemBench, by component, as reported by OpenAI.

The Structural Read

As we read OpenAI’s results, the gain is in reading chemistry correctly; the weaker scores sit in the decisions a lab actually pays for.

As we read the caveats, this is a self-reported benchmark on published work, so it measures recall and reasoning on known programs rather than discovery.

OpenAI, 20 Nov 2026

“These results do not establish whether models can discover new drugs or improve laboratory decisions.”

Three Implications

FOR DRUG DISCOVERY TEAMS GPT-6 Astra earned about half the credit on compound selection and property prediction, OpenAI said.

FOR AI LABS OpenAI tested eleven of its own models; no other lab’s models are scored yet.

FOR EVALUATORS OpenAI says prior exposure to the published sources cannot be ruled out.

What Is Not Established

OpenAI says 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”, it added, and each answer was a single attempt without tools, feedback or new experiments. Only OpenAI models were tested, so the benchmark gives no comparison with other labs yet.

Separately today, the evaluation group METR reported that AI agents appear to trail human experts on judgment-heavy tasks without clear feedback.

This Week In The AI Supercycle

AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source

This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.

Read the full issue →

Business Engineer Framework

The AI Supercycle

The Business Engineer map of where AI capability turns into business leverage.

Read the Map of AI →

The Bottom Line

On OpenAI’s own 99-problem test, GPT-6.2 Astra tops the table at 84.5%, while GPT-6 Astra earns 95.4% on molecular structures but about half the credit for choosing compounds and predicting their properties, OpenAI reported.

97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. All scores and benchmark details are OpenAI’s, from its 20 November 2026 MedChemBench release, which reports results for OpenAI models only. The 26.3-point gap is our arithmetic. OpenAI states the benchmark does not show whether models can discover new drugs.

Sources: OpenAI, ‘Introducing MedChemBench’, 20 Nov 2026 · OpenAI, MedChemBench paper (PDF)

Scroll to Top
FourWeekMBA