Claude can now read a molecule’s fingerprint about as well as the specialist software chemists have leaned on for decades. That’s the headline from Anthropic’s new science report, “Making Claude a chemist,” which tested the latest model, Claude Opus 4.7, on one of the core tasks in a chemistry lab: interpreting NMR spectroscopy.
What stands out here is the comparison. Anthropic reports that Opus 4.7 matches, and on some tasks beats, dedicated NMR software like MestReNova. For a general-purpose AI model going up against purpose-built scientific tools, that’s a real result.
What NMR is and why it matters
Before a chemist can change a molecule, they have to know its structure. Nuclear magnetic resonance (NMR) spectroscopy is their main tool for that. It produces a spectrum, a kind of signal pattern, that reveals how atoms are arranged. Reading those spectra correctly is slow, skilled work, and it’s where a lot of bench time goes.
That’s the job Anthropic pointed Claude at.
How they tested it
Anthropic benchmarked Opus 4.7 on predicting NMR signals and compared its accuracy against MestReNova, a standard in the field. According to Anthropic:
- Claude made the smallest hydrogen prediction errors in the comparison.
- It nearly matched MestReNova on carbon predictions.
- Most striking, it can run the problem backwards, inferring the molecule from the spectrum rather than just predicting the spectrum from a known molecule.
That last point is the one chemists noticed. Working a spectrum back to an unknown structure is closer to what actually happens in a lab, and it’s harder than the forward direction.
The bigger picture
NMR is one piece of a wider effort. Anthropic says it’s also testing Claude on chemical literature understanding, the messy reality that the same molecule might be drawn as a diagram, written out by name, abbreviated, or referenced by a code, all in the same paper. Tying those together is its own challenge.
Other tasks are further behind. Anthropic notes that spectral analysis is far enough along to benchmark, while areas like retrosynthesis planning, working out how to build a target molecule step by step, are still being scoped. So this isn’t a claim that Claude is a finished chemist. It’s a map of where the model is already useful and where it isn’t yet.
What this means for practitioners
If you work in or around a chemistry lab, the practical takeaway is specific:
- For NMR signal prediction and structure interpretation, Claude is now worth testing against your existing software, not as a toy but as a real second opinion.
- For verification, an AI that can reason both directions, molecule to spectrum and spectrum to molecule, can help catch errors in structure assignment faster.
- For literature work, a model that recognizes the same compound across names, drawings, and codes could cut down the manual cross-referencing that eats research hours.
Anthropic frames the goal plainly: it wants working chemists to know where Claude can save them time and where they still need to rely on their own expertise. That’s an unusually honest pitch. It’s drawing the line itself rather than claiming the model does everything.
The limitations
Nobody should read this as “AI replaces chemists.” The report is careful about scope. The strong results are on spectral analysis specifically. Retrosynthesis and other harder reasoning tasks aren’t there yet, by Anthropic’s own account. And matching specialist software on a benchmark is not the same as being trusted unsupervised on novel, high-stakes structures. The expertise of a trained chemist is still the backstop.
What’s significant is the direction. A general model is closing in on tools that were built for one job, and it can do something those tools weren’t designed for. As Anthropic keeps benchmarking the harder tasks, expect the useful-versus-not line to keep moving. For now, NMR is the part chemists can actually try today.
Full methodology and the detailed numbers are in Anthropic’s report.