Anthropic taught Claude chemistry: Opus 4.7 outperforms specialized NMR software

Anthropic has published a new entry on its science blog about using Claude Opus 4.7 in chemistry. The results are unexpected: the model handles NMR spectroscopy tasks on par with specialized software and in a number of cases outperforms it.

Chemistry and NMR spectroscopy

NMR spectroscopy (nuclear magnetic resonance) is a foundational method of organic chemistry. It lets scientists "look inside" a molecule and determine the arrangement of atoms, the types of chemical bonds and spatial configurations. Without it, developing new drugs, making polymers and synthesizing specialized materials would be impossible, as would most fundamental research in organic chemistry.

NMR data looks like a set of peaks on a graph. Each peak corresponds to a particular atom or group of atoms in a specific chemical environment. Interpreting such spectra takes deep knowledge, years of experience and an understanding of subtle chemical effects. Specialized software has been developed for years precisely to automate this process.

What the study showed

Anthropic tested Opus 4.7 on standard NMR tasks and got unexpected results:

• On most tasks, the model delivered results comparable to specialized NMR software
• On some tasks, Opus 4.7 beat traditional tools in interpretation accuracy
• The model successfully analyzed spectra, interpreted peaks and predicted molecular structures
• Unlike specialized software, Claude explains its reasoning in natural language

Context matters: specialized NMR programs are built over years by teams of chemists and engineers for specific tasks. Claude Opus 4.7 is a general-purpose language model that did not undergo specialized fine-tuning for chemistry. That is why the results look especially impressive.

Why it matters

"To manipulate a molecule, chemists first need to understand its structure.

Their main tool is NMR spectroscopy," Anthropic's blog says.

Claude's capabilities open up several practical scenarios for science and education. In research, this means less dependence on expensive licenses. Specialized NMR software often costs tens of thousands of dollars a year, which is especially significant for small labs and universities in developing countries. An AI assistant able to perform the same functions could democratize access to these tools.

For education, Claude's interactivity is a major advantage. Traditional software gives an answer but does not explain it. A chemistry student learning to interpret spectra can ask follow-up questions, ask for a specific peak to be explained or two similar signals to be compared, and get detailed answers in plain language.

What this means

Claude's success on NMR tasks is part of a broader trend: language models are increasingly moving into specialized scientific niches. If it was once assumed that AI would help with text and code but stay out of real science, these results suggest otherwise.

The next step is integrating such capabilities into real laboratory workflows and, possibly, into chemists' standard toolkits.