Anthropic published a white paper on June 5, 2026, showing that Claude Opus 4.7, a general-purpose model with no chemistry-specific training, holds its own against the dedicated NMR software chemists use every day. The test pitted three Claude models against ChemDraw and MestReNova on 20 compounds, and on some measures Claude came out ahead.
What they actually tested
NMR spectroscopy is how a chemist confirms which molecule they actually made. It is also tedious work: matching every peak in a spectrum to an atom in a proposed structure, by hand, for every single compound. Anthropic chemist David Kamber set up the test using 20 compounds pulled from ChemRxiv preprints posted after the models' training cutoff, which sidesteps the obvious worry that the models had simply memorized the answers.
The first task was forward prediction: feed the software a structure, ask it to predict where the hydrogen and carbon peaks land. The research post reports that on hydrogen, Opus 4.7 was the most accurate of the lot, averaging an error of about 0.079 ppm, well inside the window a chemist would accept. On carbon it was effectively tied with MestReNova. So far, a draw against tools built for exactly this.
Where it pulled ahead
Peak shape is the more interesting result. Claude matched the reported splitting pattern more often than any other tool, and all three Claude models nailed the sub-peak spacing to within half a hertz roughly 80% of the time. ChemDraw and MestReNova managed that 26 to 35% of the time. That is a wide gap, and splitting patterns carry real structural information, so it is not a cosmetic win.
The numbers come from Anthropic testing its own model, which is worth keeping in mind. The compound set is also tiny, and the authors say so plainly.
The harder direction
Forward prediction is the easy half. The reverse, working out a structure from a spectrum, is what existing software mostly leaves to the human. Dedicated elucidation tools exist, but they typically want 2D NMR, a paid license, and someone who knows how to drive them.
Claude was given 15 elucidation problems, three attempts each, working only from a molecular formula and 1D hydrogen and carbon spectra. On the eight simpler targets it recovered every structure on every attempt. On the seven denser ones, given the starting material as a hint, it got four right on all three runs and the rest on two of three. The starting-material hint is doing some lifting here; the post notes that without it, the model would sometimes spin through its reasoning without committing to an answer.
"Our claim is a modest one," the team writes, and for once that reads as accurate rather than false humility. Twenty compounds is a sketch, not a verdict.
So what now
The catch is sample size, and Anthropic frames the results as indicative rather than precise. They list what they left out: 2D experiments, stereochemistry, natural products, most solvents beyond DMSO, chloroform, and water. The team says it would want several hundred compounds across 20 to 30 scaffold classes before drawing firm conclusions.
Anthropic is opening its AI for Science program to chemistry researchers, with applications and a contact address in the post. The full white paper is linked there for anyone who wants the per-compound breakdowns.
