Recent tests show that large language models trained on all publicly available knowledge can predict molecular properties better than chemistry‑specialised models when only a few studies are available. Researchers compared general‑purpose foundation models—GPT‑6 Astra, Gemini 4 Argon and Claude Opus 5.5—to chemistry‑focused models ChemBERTa and MoLFormer. Each general‑purpose model received a prompt with 60 labeled molecules and 40 unseen queries. The specialised models were fine‑tuned on 60 to 10,000 labeled data points.

Across five regression tasks—the total energy U₀, ε LUMO, FreeSolv, ESOL, and Lipophilicity—the general‑purpose models scored higher R² values. For example, GPT‑6 Astra achieved R² of 0.99 on U₀ versus 0.67 for ChemBERTa, and 0.88 on ESOL versus 0.73 for MoLFormer. On the BACE activity classification, GPT‑6 Astra scored 0.47 against 0.70 for ChemBERTa. When 1,000 training points were supplied, chemistry‑only models caught up.

Negative‑control experiments shuffled the target labels. Even then, GPT‑6 Astra and Claude Opus 5.5 maintained high accuracy, suggesting these models may be memorising distribution patterns rather than learning chemistry rules. The models also sometimes declined to answer prompts on certain tasks.

These findings point to general‑purpose models as powerful starting points for chemical predictions, but caution is needed because high accuracy does not guarantee true generalisation.”