Anthropic's general-purpose language model, Claude Opus 4.7, has achieved performance on par with dedicated nuclear magnetic resonance spectroscopy software. A research report published June 5, titled "Making Claude a chemist," shows Opus 4.7 matches specialized tools like ChemDraw 25.0.2 and MestReNova 17.0.0 in chemistry tasks.
The study tested the model across 20 compounds from recent synthetic chemistry preprints, evaluating both forward prediction and inverse structure elucidation. On hydrogen NMR shifts, Opus 4.7 recorded the lowest average error at plus or minus 0.079 ppm. For carbon shifts, it tied with MestReNova at plus or minus 1.37 ppm. Errors under 0.1 ppm on hydrogen data represent high-quality predictions.
The model also outperformed on consistency in predicting peak splitting patterns and J-coupling values. In inverse tasks, it recovered all simpler target structures on every attempt. With hints from starting materials, it succeeded on four out of seven denser structures across all runs.
Notably, Opus 4.7 was not fine-tuned on chemistry-specific data. Chemists can copy NMR data into a chat window and receive structural proposals without a proprietary software license or 2D NMR data, potentially streamlining a workflow largely unchanged for decades.