Anthropic reports Claude matching NMR software on spectral analysis

Category: ML research
Status: developing
Sources: Anthropic research, HN discussion
Summary: In a research write-up dated 2026-06-05, Anthropic reports evaluating Claude on nuclear magnetic resonance (NMR) spectral analysis using 20 compounds drawn from synthetic-chemistry preprints published after the models' training cutoff, across four structural families. The post states Opus 4.7 reached average errors of about 0.079 ppm for hydrogen and 1.37 ppm for carbon shift prediction, comparable to or better than the ChemDraw and MestReNova chemistry tools, and that on inverse structure elucidation it recovered all 8 simpler structures from spectra alone and 4 of 7 more complex structures when given starting-material context. The result is the lab's own internal evaluation and has not been independently reproduced.
Why it matters: If a general model performs routine NMR prediction at parity with dedicated cheminformatics software, it shifts part of the structure-elucidation workflow from specialized tools toward LLM tool-use pipelines, a direct interest for AI-for-science tooling.
Follow-up: Watch for independent reproduction, a method paper or dataset release, and evaluation on larger or harder compound sets.