Anthropic Research · 2026-06-05 · notable
Anthropic Research's 'Making Claude a Chemist' — Claude Opus 4.7 Hits ±0.079 ppm Hydrogen NMR Prediction Error, Ties MestReNova on Carbon Shifts, and Recovers All 8 Simpler Molecular Structures From Spectra Plus Formula Alone
Anthropic chemist David Kamber benchmarks Claude on NMR analysis: Opus 4.7 averages ±0.079 ppm hydrogen-shift error, ties MestReNova on carbon, beats classical software on peak splitting, and elucidates 8/8 simpler structures from 1D spectra plus formula.
Anthropic's first AI-for-Science post puts Claude up against ChemDraw and MestReNova on routine NMR work — and Opus 4.7 holds its own.
Key specs
| Opus 4.7 1 h nmr avg error | ±0.079 ppm |
|---|---|
| Opus 4.7 13 c avg error | ±1.37 ppm |
| Mest re nova 13 c avg error | ±1.48 ppm |
| Forward structures solved | 8 / 8 |
| Inverse hard problems solved | 4 / 7 |
What is it?
A research note from Anthropic chemist David Kamber, the first in a series Anthropic is publishing under its AI for Science effort. The post evaluates Claude — primarily Opus 4.7 — on hydrogen and carbon nuclear-magnetic-resonance (NMR) tasks across 20 small molecules from four structural families drawn from publications after the model's training cutoff.
How does it work?
The team ran two task shapes. In the forward direction, Claude predicted 1H and 13C chemical shifts from a molecule's structure; results were compared to ChemDraw and MestReNova. In the inverse direction, Claude was given the experimental 1H and 13C spectra plus the molecular formula and asked to propose the structure. For harder targets, the team added small starting-material hints. Predictions were scored against literature values.
Why does it matter?
On forward 1H prediction, Opus 4.7's ±0.079 ppm error is well under half the tolerance window that organic chemists use to call peak assignments correct, and it predicts sub-peak spacing within half a hertz ~80% of the time versus 26–35% for classical software. The inverse task — going from spectrum to structure on a 1D NMR alone — is a job no off-the-shelf chemistry package does well; Opus 4.7 cleared all 8 simpler structures on every attempt and 4 of 7 harder ones across 3 runs.
Who is it for?
Synthetic chemists, analytical chemistry labs, AI-for-science researchers.
Try it
claude.ai