Anthropic is integrating specialized chemistry expertise to improve Claude's ability to handle complex molecular data. Chemist David Kamber led a study testing how the models interpret NMR spectra, a tedious manual process where chemists match spectral peaks to atoms. This work aims to reduce the research burden of translating between hand-drawn sketches, database strings, and instrument readouts.

Researchers compared Opus 4.7, Opus 4.6, and Sonnet 4.6 against industry standards like ChemDraw and MestReNova. Using 20 novel compounds from recent preprints to avoid training bias, the team measured prediction errors in parts per million. Opus 4.7 outperformed others on hydrogen predictions with an average error of ±0.079 ppm, falling well within the standard tolerance window.

While traditional software handles forward prediction, Claude attempted the harder task of proposing a molecular structure from an experimental spectrum. This capability leverages multimodality and step-by-step reasoning to allow chemists to audit the AI's logic. This approach bypasses some long-standing data problems like paywalled journals and sparse null-results that hindered previous machine-learning tools.

The results show that frontier models can now match or exceed the accuracy of dedicated chemistry software for specific spectral tasks. This shift allows chemists to automate the translation and integration work that typically consumes significant lab time. The study proves that multimodal reasoning can handle NMR spectroscopy without relying on pre-curated databases.