Anthropic: Claude as a Chemist — Opus 4.7 Predicts NMR Spectra With ±0.079 ppm Error

Anthropic published a white paper on June 5, 2026, about the capabilities of the Claude model in chemistry, with an emphasis on NMR spectroscopy (nuclear magnetic resonance — a method for determining molecular structure). In the research, the Claude Opus 4.7 model achieved an average error of ±0.079 ppm in hydrogen prediction, demonstrating accuracy that holds up against specialized chemistry tools.

What is NMR spectroscopy and why does it matter?

NMR spectroscopy is one of the foundational methods in chemistry for determining the structure of molecules. It relies on the behavior of atomic nuclei in a magnetic field, and the results are expressed in ppm (parts per million). Chemists use it daily to confirm that a synthesized compound is indeed the one they intended to produce.

The ability of a language model to work reliably with NMR data is not trivial, because it requires understanding the relationship between a molecular structure and the physical signal it produces.

How accurate was Claude Opus 4.7?

In predicting hydrogen signals, Claude Opus 4.7 achieved an average error of ±0.079 ppm, which the researchers describe as a value well within half the tolerance window. For carbon, the error was ±1.37 ppm.

These results are comparable to established industry tools. The ChemDraw and MestReNova programs yield an error of ±1.48 ppm on the same task, which Claude even surpasses for carbon. The model thus shows it can match specialized software rather than merely providing rough estimates.

What does it mean that Claude works bidirectionally?

A particularly interesting finding is that Claude works bidirectionally. In one direction, it predicts the expected spectrum from a known structure (forward prediction). In the other direction, it performs inverse structural elucidation — attempting to reconstruct the molecular structure from the spectrum alone.

In the inverse task, the model reconstructed all 8 simpler molecular structures from spectra alone. For more complex targets, it succeeded with 4 of 7, with additional context about the starting material. This shows it is capable of reasoning in both directions, which is closer to the way an experienced chemist thinks.

How was the evaluation conducted?

The evaluation covered two groups of tasks. For forward prediction, 20 compounds across four scaffold families (groups of related molecular backbones) were used. For inverse structural elucidation, 15 examples were used.

This setup allows the model to be tested on diverse chemical structures rather than just a narrow sample. The results suggest that Claude can serve as a useful auxiliary tool in chemical work, especially in stages where one needs to quickly assess the agreement between a predicted and a measured spectrum.

What do these results mean for chemists?

For chemists in the lab, a language model’s ability to reliably predict and interpret NMR spectra could speed up daily work. Confirming the structure of a synthesized compound often requires comparing the measured spectrum with the expected one, and a tool that does this quickly and accurately reduces the time needed to verify results.

It is particularly important that Claude does not replace existing specialized tools, but rather matches them and in some measures surpasses them, for example in carbon prediction. In addition, its ability to perform inverse structural elucidation opens the possibility of proposing a candidate structure from an unknown spectrum, a demanding task even for experienced specialists. With this white paper, Anthropic shows that general language models are increasingly entering domains that until recently were reserved for narrowly specialized software.