Making Claude a Chemist
Anthropic says it is working to make Claude better at chemistry, starting with a test of how well it handles NMR spectra.
Intelligence analysis by GPT-5.4 Mini
Anthropic frames chemistry as a translation problem across sketches, spectra, databases, and publications, then tests whether Claude can help. The first white paper compares Claude models with ChemDraw and MestReNova on NMR tasks using 20 novel compounds from recent preprints.
Anthropic is teaching Claude to help chemists read hard-to-understand clues from lab tests, like learning to match puzzle pieces in a jigsaw. The first test checks whether Claude can help with NMR spectra, which are like fingerprint patterns for molecules.
Analysis
Anthropic says it is working with synthetic, computational, and analytical chemists to make Claude more useful in chemistry, where people constantly move between hand-drawn structures, instrument readouts, database syntax, and publication notation. The company argues that this translation work is time-consuming and difficult to scale, even though chemistry underlies drugs, materials, and many everyday products.
The first white paper in this effort focuses on NMR spectroscopy, which Anthropic describes as one of the most time-consuming parts of synthetic chemistry. In practice, chemists must match peaks in a spectrum to atoms in a proposed structure by hand. The paper tests 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 evaluation uses compounds from four structural families, with five compounds in each family, chosen because each family presents a different NMR challenge. For forward prediction, Claude and the specialist tools receive structures in SMILES format and must predict where each hydrogen and carbon peak should appear in a 1D NMR spectrum. Anthropic also tests the harder reverse task: starting from an experimental spectrum and proposing the structure behind it. The article says existing software generally leaves that step to the chemist.
Anthropic’s broader claim is careful rather than sweeping. The company says frontier multimodal models can read structures directly from figures or sketches, can ingest methods sections as published, and can show reasoning step by step so chemists can audit outputs. The post argues that this does not solve chemistry’s data problems, but it does make some problems more tractable. The overall message is that Claude is beginning to assist with the daily translation, recall, and integration work that supports human judgment, rather than replacing that judgment.
Key points
- Anthropic says it is working with chemists to make Claude more useful for chemistry tasks.
- The first white paper focuses on NMR spectroscopy, a key but time-consuming step in synthetic chemistry.
- Claude models are tested against ChemDraw and MestReNova on 20 compounds from recent ChemRxiv preprints.
- The study covers both forward NMR prediction and the harder task of inferring a structure from a spectrum.
- Anthropic argues that multimodal models can help with translation and reasoning even though chemistry data remain messy.
If Claude can reliably help read NMR spectra, it could save chemists a lot of manual checking and speed up routine structure work. That would make AI more practically useful in a field where people still spend substantial time translating between different chemical formats.
The article also makes clear that chemistry still has major data and workflow limits, so progress may be uneven. If the model misreads spectra or overstates confidence, chemists would still need to verify every result by hand.
