Anthropic says its Claude models are beginning to take on a task that sits at the center of everyday chemistry: reading nuclear magnetic resonance, or NMR, spectra and using them to identify molecular structures.
In a new white paper, the company reported that one of its general-purpose models, Claude Opus 4.7, performed competitively with widely used chemistry software on a small benchmark of synthetic molecules. Anthropic also said the model was able to work in reverse, suggesting likely structures from spectral data and a molecular formula.
NMR is one of the standard tools chemists use to confirm what a compound actually is. Because molecules cannot be directly seen in a microscope, researchers rely on the pattern of peaks in a spectrum to connect atoms in a proposed structure with the experimental data. That process is often slow and labor-intensive, especially when each peak has to be matched by hand.
To test whether Claude could help with that work, Anthropic compared three of its models, Opus 4.7, Opus 4.6 and Sonnet 4.6, against ChemDraw and MestReNova on 20 compounds pulled from synthetic chemistry preprints published after the models’ training cutoff. The company said it chose compounds spanning four structural families, each selected to present different NMR challenges.
For the forward-prediction task, each tool received a molecule in SMILES format and was asked to predict where the hydrogen and carbon peaks would appear in a 1D NMR spectrum. Anthropic then compared the predicted values with published experimental results. Because language models can vary from run to run, the Claude models were queried three times per compound.
On hydrogen predictions, Opus 4.7 produced the strongest results, with an average error of 0.079 ppm. On carbon, Opus 4.7 and MestReNova were close, with average errors of 1.37 ppm and 1.48 ppm respectively. Opus 4.6 came in behind those two, while Sonnet 4.6 performed worst of the three Claude models.
Anthropic said the biggest differences showed up in difficult edge cases, including an NH proton in one chloropyridazine compound that the lower-tier models placed far from its actual position.
The company also said Opus 4.7 was better than the other tools at predicting peak splitting and sub-peak spacing, features chemists use when reading spectra. According to the benchmark, the model matched the reported splitting pattern more often than ChemDraw or MestReNova, and all three Claude models predicted sub-peak spacing within half a hertz about 80 percent of the time.
Anthropic also tested whether Claude could identify a molecule from its spectrum alone. In that part of the study, Opus 4.7 was given 15 structure-elucidation problems along with the exact molecular formula and 1D hydrogen and carbon NMR spectra. For the more complex cases, the model was also given the structure of the starting material.
The company said Opus 4.7 identified all eight of the simpler targets correctly on every attempt using only spectra and formula. For the seven harder targets, the model got the right answer on all three tries for four molecules when given the starting-material hint, and on two of three tries for the rest.
Anthropic framed the results as evidence that a general-purpose model can now assist with a common chemistry workflow without being specifically fine-tuned for the field. The company said Claude can handle both routine prediction and the reverse problem of proposing a structure from data that a chemist might paste into a chat interface.
Still, Anthropic was careful to note limits. The benchmark was small, covering 20 compounds in the forward task and 15 in the inverse task. The company said it did not evaluate 2D NMR techniques, stereochemistry, or a broader range of solvents and scaffold types. It also said the harder inverse problems remained challenging without extra context.
Even with those caveats, the study signals a broader push by Anthropic to make Claude more useful in chemistry, not only for reading spectra but also for structure rendering, reaction reasoning and mechanism analysis.