Preprint finds general-purpose model beats chemistry models on sparse data

A preprint reports that GPT-6 Astra outperformed the chemical foundation models tested on five physicochemical and quantum-chemical prediction tasks using only 60 experimental data points. The comparison tests whether models trained on molecular structures provide better predictions than general-purpose models when experimental measurements are scarce. The best chemical models matched Astra's performance with 1,000 versus 60 data points and 10,000 versus 1,000 points, respectively. The findings concern these prediction tasks and training-data comparisons, rather than demonstrated improvements in experimental chemistry.