A general-purpose language model just walked into the chemistry lab and held its own against software that’s been purpose-built for molecular analysis. Anthropic published a research report on June 5 titled “Making Claude a chemist,” demonstrating that Claude Opus 4.7 can perform nuclear magnetic resonance spectroscopy tasks at a level that matches, and in some cases exceeds, dedicated NMR tools like ChemDraw 25.0.2 and MestReNova 17.0.0.
The numbers tell the story
Anthropic’s study tested Opus 4.7 across 20 compounds sourced from recent synthetic chemistry preprints, evaluating both forward prediction (simulating what a spectrum should look like given a molecular structure) and inverse structure elucidation (working backward from spectral data to figure out the molecule).
On hydrogen NMR shifts, Opus 4.7 posted 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. To translate that into something meaningful: parts per million is the standard unit for measuring chemical shifts in NMR, and errors under 0.1 ppm on hydrogen data represent genuinely high-quality predictions.
The model also outperformed on consistency when predicting peak splitting patterns and J-coupling values, two features that chemists rely on heavily to distinguish between similar molecular structures.
On the inverse side of things, where the model had to deduce structures from 1D NMR and high-resolution mass spectrometry data, Opus 4.7 successfully recovered all simpler target structures on every attempt. When the team added hints from starting materials for more complex targets, the model succeeded on four out of seven denser structures across all runs.
Why this is different from typical AI benchmarks
What makes the Anthropic result unusual is that Opus 4.7 wasn’t fine-tuned on chemistry-specific data for this task. It operates on routine chemist-pasted readouts with no specialized setup required. In English: a chemist can copy their NMR data into a chat window and get a structural proposal back, no proprietary software license needed.
The study also notably didn’t require 2D NMR data, which is typically considered essential for complex structure elucidation. Two-dimensional NMR experiments take longer to run and generate more data to interpret. Bypassing that requirement, even for simpler compounds, streamlines a workflow that has remained largely unchanged for decades.