The premise that just failed a test
Chemical foundation models exist because of a data problem. Molecular structures are abundant; experimental measurements are not. Models such as ChemBERTa and MoLFormer were pretrained on large collections of molecular structures precisely so that a researcher with fewer than a hundred measurements — the realistic case in most chemistry labs — could fine-tune one and get a usable property predictor.
A paper posted to arXiv on 30 October tested whether that still holds against general-purpose frontier models, and reports that it does not.
Sixty datapoints
On five physicochemical and quantum-chemical regression tasks, with only 60 experimental datapoints available, GPT-6 Astra outperformed the chemical foundation models the authors assessed as currently eligible.
That is the data-scarce regime the specialist models were built for. Losing there is not a marginal result; it is a loss on home ground.
The exchange rate
The more useful number is how much data it takes for the specialists to catch up. The authors report roughly a factor of ten.
Given 1,000 datapoints, the best chemical models reach the performance GPT-6 Astra manages on 60. Given 10,000, they reach what Astra manages on 1,000. The specialist models are not worse in the abstract — they are behind by about an order of magnitude of experimental effort, which in a wet lab is the difference between a semester and several years.
What the authors think is happening
The explanation offered is about what the two kinds of model were trained on. Chemical foundation models learned from molecular structures. General-purpose models were trained on close to all publicly available text, which includes the accumulated human literature on molecular properties and the physical laws that govern them.
The implication is that the prior a frontier model brings to a chemistry task is not structural pattern-matching but something closer to the field’s written knowledge — and that in data-scarce settings, that prior is worth more than a structure-only pretraining run.
What the paper does not settle
The comparison names GPT-6 Astra and Claude Opus 5.5 as the general-purpose side. These are closed, API-only models, so the result is not reproducible in the way a comparison between two open chemical models would be, and it is sensitive to whichever versions were served during the evaluation window. It also covers regression tasks on physicochemical and quantum-chemical properties, not synthesis planning, reaction prediction or anything requiring a lab.
A general-purpose model that has read the chemistry literature may also have read the specific measurements in a benchmark. The paper’s framing does not resolve the contamination question for readers.
What to watch
If the ten-times figure holds up, the practical advice for a lab with sixty measurements changes: start with a frontier model, not a chemical one. Watch for independent replication on held-out measurements collected after the models’ training cutoffs, which is the only version of this test that settles the contamination objection.
