News

OpenAI's New Drug Exam Shows GPT-6.2 Astra Can Rebuild Molecules and Still Miss the Pick

OpenAI published MedChemBench on November 20. GPT-6.2 Astra scored 84.5 percent at xhigh, with about half credit on compound picks and property predictions.

OpenAI just published a chemistry test where its newest model can redraw a molecule from a paper and still cannot tell you which analog to make next.

The company introduced MedChemBench on November 20: 99 problems drawn from 117 papers and patents, covering 761 distinct molecular structures. At the highest shared reasoning setting, which OpenAI calls xhigh, GPT-6.2 Astra earned 84.5 percent of the available credit. GPT-6 Astra scored 83.3 percent, GPT-6.1 Sol 82.0 percent, and GPT-5.6 Sol 58.2 percent. GPT-6 Astra was strongest at reconstructing structures and explaining the evidence already in the prompt. Selecting compounds and predicting their numerical properties were harder.

OpenAI is careful about what that score is not. The results do not show that a model can discover a drug, or that it would improve a lab's next experiment. Each answer was a single attempt, with no tools, no feedback, and no new wet-lab data. Prior exposure to the published sources cannot be ruled out.

95 percent on the drawing, about half on the pick

Most problems give the model the context a chemist would have had at a chosen point in a program: the biological target, the goal, the compounds tried so far, and selected measurements. The bench grades four things: atom connections and stereochemistry, explanations backed by the supplied evidence, rankings that meet the stated requirements, and agreement with the numbers in the source.

Structure is the easy part. Astra took 95.4 percent of the structure credit at xhigh. In one task that asks the model to assemble a molecule around an 18-atom ring, GPT-6 Astra reconstructed the compound correctly in all ten responses. GPT-5.6 Sol and GPT-5.6 Luna failed all ten. Sol shrank the ring to 15 atoms. Luna split the molecule into two disconnected fragments. Those are different chemicals.

The harder half is the actual medicinal-chemistry job. Astra earned about half the credit for selecting compounds and predicting numerical properties. One problem, based on PRMT5 inhibitors studied in cancer research, asks the model to forecast three related compounds: how strongly they block the protein in cells, how readily they cross a cell barrier, and how quickly liver enzymes break them down. Later experiments, withheld from the model, showed a tradeoff. Compound 23 blocked PRMT5 at a lower concentration. Compound 24 crossed the barrier more readily and broke down more slowly.

That is the point of the bench. Stronger activity does not come bundled with better absorption or slower metabolism. OpenAI already published GPT-6.2 Astra as a model it holds at a self-improvement threshold. MedChemBench is the same lab asking whether that model can read a reaction scheme and still miss which analog is worth making.

A test of yesterday's notebooks

OpenAI says higher reasoning settings lifted scores, and large gaps between models remained. At max effort, GPT-6 Astra reached 83.8 percent. The company evaluated eleven GPT models across 51 combinations of model and reasoning setting.

What the bench cannot say is whether any of this holds on a live program, with new assays and a chemist in the loop. It grades reasoning from historical evidence. The next molecule, the one that is not in a 2024 paper, is still untested.

Some offers on this page may be paid placements or contain affiliate links.

Subscribe to Techpresso

Free daily newsletter, read in 5 minutes.

Subscribe free