Making Claude a Chemist
Anthropic says Claude is starting to help chemists with NMR analysis and structure work. The first white paper compares Claude with standard chemistry tools on 20 novel compounds.
Intelligence analysis by GPT-5.4 Mini
Anthropic is pitching Claude as a chemistry assistant for the translation work chemists do across sketches, spectra, and paper notation. Its first white paper tests Claude on NMR prediction and structure elucidation, comparing several Claude models with ChemDraw and MestReNova on post-training-cutoff compounds.
Anthropic is teaching Claude to help with chemistry homework for grown-ups. It is like giving a detective a better magnifying glass so it can match puzzle pieces from a spectrum to the right molecule.
Analysis
Anthropic frames chemistry as a domain where the same molecule appears in many forms: hand-drawn structures, spectra, databases, patents, and papers. The company argues that this translation burden is central to lab work and difficult to scale manually, especially given the size of chemical space and the steady flow of new substances.
The post says frontier models now have two capabilities that change the chemistry use case. First, they are multimodal, so they can read chemical structures directly from figures or sketches rather than relying only on pre-curated databases. Second, they can show reasoning step by step, which gives chemists something they can inspect and audit.
The first concrete test in the white paper focuses on NMR, one of the most common analytical inputs in synthetic chemistry. Anthropic says it compared three Claude models, Opus 4.7, Opus 4.6, and Sonnet 4.6, against ChemDraw and MestReNova on 20 compounds drawn from ChemRxiv preprints published after the models' training cutoff. The selection was designed to avoid training-data leakage and included compounds from four structural families, five per family, chosen to cover different NMR challenges.
The evaluation covered both forward prediction and the harder reverse task. For forward prediction, the tools received a structure encoded as SMILES and had to predict where each hydrogen and carbon peak would appear on a 1D NMR spectrum. For structure elucidation, Claude was also asked to work backward from an experimental spectrum to a proposed structure, a task that existing software still leaves to chemists.
The article's broader claim is modest: Claude is not replacing chemists, but it is beginning to help with the repeated recall, translation, and integration tasks that sit around expert judgment. Anthropic presents the white paper as an initial step in a longer effort to make that assistance more useful.
Key points
- Anthropic says Claude is being developed to assist chemists with translation and analysis work, not just text tasks.
- The first white paper focuses on NMR, a core but time-consuming chemistry technique.
- Claude models were tested against ChemDraw and MestReNova on 20 novel compounds from ChemRxiv preprints.
- The benchmark included both forward NMR prediction and the harder reverse task of inferring structure from a spectrum.
- Anthropic argues multimodal models and step-by-step reasoning make chemistry problems more tractable.
If the results hold up, Claude could save chemists time on routine interpretation work and make it easier to move between drawings, spectra, and papers. That would be especially useful in labs that do not have access to lots of specialized software or large teams.
The article also makes clear that chemistry AI still faces hard data problems, including sparse negative results and messy published formats. If Claude's reasoning is wrong or hard to verify in edge cases, chemists will still need to do the final interpretation by hand.
