Anthropic Claims Claude Opus 4.7 Can Rival Specialist NMR Software for Chemical Structure Analysis

Anthropic has published a case study claiming that its general-purpose AI model Claude Opus 4.7 performs as well as — and in some tests better than — established specialist chemistry software on nuclear magnetic resonance (NMR) spectroscopy tasks. The post, titled “Making Claude a chemist,” appears on Anthropic’s newly launched Science Blog and sets out internal benchmark results comparing Opus 4.7 against ChemDraw and MestReNova, two widely used tools in computational chemistry.

The claim is striking because Opus 4.7 has not been fine-tuned on proprietary chemistry datasets. Anthropic says the competitive performance comes from general training combined with prompt and workflow design alone.

What the Benchmarks Show

According to Anthropic’s own reporting, the evaluation covered two broad task types: predicting NMR spectra from known molecular structures, and working backwards — inferring a structure from spectral data. That second category, which chemists call the “inverse” problem, is where the claims are most pointed.

For NMR spectrum prediction, Anthropic tested Opus 4.7 on 20 compounds, according to secondary technology reporting that summarises Anthropic’s study design. The model outperformed rival tools on proton (hydrogen) NMR prediction and matched MestReNova on carbon NMR prediction. Anthropic states that Opus 4.7 is “as good as or better than” ChemDraw and MestReNova on average for routine prediction tasks.

The inverse-structure results are where the numbers get specific. Anthropic reports that Opus 4.7 recovered all eight of a set of simpler molecular structures on every attempt, using only a molecular formula and spectral data. For seven harder target molecules, the model was given a starting-material hint; it returned the correct structure on all three runs for four of those targets, and on two out of three runs for the remaining ones.

These figures come from Anthropic’s own evaluation. No independent standards body — not the Royal Society of Chemistry, nor UK Research and Innovation, nor any national standards laboratory — has published external verification of these results.

How It Works in Practice

NMR spectroscopy works by probing how atomic nuclei respond to magnetic fields and radiofrequency pulses. The technique is a core analytical method across organic, medicinal and materials chemistry, used daily by bench chemists to confirm what molecule they’ve made or to work out what an unknown compound actually is.

Dedicated software like ChemDraw and MestReNova has handled this work for decades. Both require licensed software, specialist training and, for harder problems, full two-dimensional NMR datasets. Anthropic argues that Claude can accept the kind of inputs a bench chemist already has — a molecular formula, a high-resolution mass spectrum, a list of 1D NMR peaks — through a standard chat interface, and return candidate structures alongside reasoning and commentary.

That’s the accessibility argument. Rather than replacing specialist tools outright, Anthropic positions Opus 4.7 as lowering the barrier to computational support, chiefly for smaller labs or teaching settings where licensed specialist software may not be readily available.

Broader Ambitions

The “Making Claude a chemist” post sits within a wider Anthropic push it calls “Anthropic Science,” a programme targeting researchers at established institutions across chemistry, biology, medicine and environmental science. A parallel initiative, Claude for Life Sciences, aims to support biologists and R&D teams with literature review, experimental design and protocol drafting.

Anthropic is clear that the Science Blog work is presented as research, not a commercial product launch. The benchmarks, tasks and success criteria were designed and interpreted by Anthropic and collaborating chemists — not by an independent standards body.

Critics have already raised questions about the scope of the evaluation. The compound sets are small: eight simpler structures and seven harder ones for inverse problems. The task definitions and what counts as a correct result are Anthropic’s own. And Opus 4.7 has not been formally validated for regulated environments such as good laboratory practice, or pharmaceutical quality control, where strict documentation standards apply.

There are also dual-use concerns. A general-purpose model that can reason about molecular structures and spectral data is, by definition, a tool that could be applied well beyond legitimate research. Regulators and safety specialists have begun raising questions about governance and usage monitoring for AI systems with detailed chemical-reasoning capabilities.

Unanswered Questions

Several things remain unclear. Anthropic has not published the full dataset, evaluation protocol or raw results in a peer-reviewed journal, so independent replication isn’t yet possible. It’s also not known how Opus 4.7 performs on larger, more structurally complex molecules, or on the kind of messy, real-world spectra that bench chemists routinely encounter rather than clean benchmark inputs.

Whether performance holds without the starting-material hint — the cue given for the harder inverse problems — is another open question. And how the model behaves when it’s wrong, whether it signals uncertainty clearly or presents incorrect structures with unwarranted confidence, isn’t addressed in detail in the published case study.

What This Means for Kent Residents

Chemistry departments at the University of Kent and any contract research or quality-control labs in the county that currently rely on licensed NMR software may find Anthropic’s claims worth watching, though independent validation would be needed before any research or industrial workflow could responsibly incorporate AI-generated structural proposals. Over a longer horizon, faster AI-assisted drug discovery — if these methods prove out in regulated settings — could feed into pharmaceutical pipelines that ultimately benefit patients served by NHS Kent and Medway. For now, any adoption would be at the discretion of individual institutions, subject to data-protection, intellectual property and research-integrity rules.