On November 20, OpenAI introduced MedChemBench, an evaluation of medicinal chemistry tasks connecting molecular structures with experimental evidence. It contains 99 problems drawn from 117 papers and patents. The company reports 84.5% of available credit for GPT-6.2 Astra at xhigh. That is a score under a specific grading scheme, not the proportion of successful drugs developed.
What this kind of evaluation shows
The useful question is whether a system makes verifiable connections between available evidence and a proposed next action. In scientific work, a plausible explanation may be insufficient if it concerns the wrong structure or omits an important condition. Evaluation therefore needs to distinguish different kinds of mistakes.
Our editorial interpretation is that the aggregate score starts the discussion. A scientific team needs to understand which part of its task a model handles well and where human checking remains essential. Averaging can conceal a weak component that is critical to a particular project.
A proposal for checkable use
We would begin with a completed internal case whose answers can be checked. Give the model only the information available at a chosen historical point. Keep the later outcome hidden until assessment. This makes a useful prediction easier to distinguish from a persuasive explanation written with hindsight.
For every answer, we would require a clear connection to the supplied evidence and a separate statement of assumptions. A proposed compound lacking a required property should remain an uncertain possibility. This is our example procedure; SIExpo has conducted neither chemistry experiments nor an original MedChemBench evaluation.
We would not use such a pilot to approve a laboratory or clinical stage automatically. Actual testing requires other kinds of evidence and expertise. A convenient interface and a high language-model score cannot substitute for those decisions.
Keeping rejected proposals and the reasons for rejection would be particularly useful. Reporting only the successful ideas makes the system appear more reliable than everyday operation may justify. Such a record also lets a later model version be compared on the same cases, with changes in the evidence clearly identified.
Before expanding use, the team should specify what success means operationally: fewer factual corrections, better prioritization or a shorter review process. These are different goals. A tool can improve one while leaving another unchanged, and a benchmark score alone cannot decide which matters most.
MedChemBench is a new instrument for measuring scientific capabilities. Its value depends on transparent tasks, understood grading limitations and subsequent scrutiny by researchers, rather than on turning one evaluation result into a medical promise.



