← Back to feed

Anthropic published a white paper on June 5 comparing Claude Opus 4.7 against ChemDraw and MestReNova on nuclear magnetic resonance spectroscopy, chemistry’s primary structural characterization tool. The results show a general-purpose language model now performs at the level of dedicated NMR software on prediction tasks — and outperforms standard tooling on a harder task the tools don’t formally support.

What NMR Measures and Why It Matters

Nearly every small molecule undergoes NMR analysis during research and quality control — drugs, pesticides, polymers, food additives, and DNA subunits. The technique works by placing a molecule in a magnetic field and reading the radio frequencies at which its nuclei resonate. Each nucleus in a molecule generates a distinct signal that encodes structural information: bond types, spatial arrangement, neighboring atoms.

The forward problem — given a structure, predict its spectrum — is what ChemDraw and MestReNova handle. The inverse problem — given a spectrum, identify the structure — is what organic chemists do. Existing mainstream tools do not solve the inverse step automatically. That work requires expert interpretation.

The Benchmark

Anthropic’s evaluation, run by company chemist David Kamber, tested three capabilities:

1H (hydrogen) NMR prediction: Claude Opus 4.7 achieved the smallest prediction errors of any model tested. Chemical shift prediction for hydrogen nuclei is the most common NMR task in synthetic chemistry — getting it wrong misleads the chemist about what atom environments are present.

13C (carbon) NMR prediction: Claude Opus 4.7 nearly matched MestReNova. Carbon NMR provides the clearest map of a molecule’s carbon skeleton, and MestReNova is the field’s commercial standard for interpreting it.

Structure elucidation (inverse NMR): Given a spectrum, Opus 4.7 could propose the molecular structure. This is the task existing NMR software tools generally leave to the human expert. Anthropic’s report describes it as working the problem backwards — from the spectral shadow back to the molecule.

No Fine-Tuning

Claude Opus 4.7 carries no chemistry-specific fine-tuning. The performance comes from a general-purpose model’s ability to reason across representations: it can read a spectrum as structured data, relate it to chemical shift tables and bonding rules, and produce a structural hypothesis. The Anthropic post notes that frontier models’ multimodal and explicit reasoning capabilities change which problems are tractable “despite” the data scarcity that has limited narrower ML tools in chemistry for years.

The Structural Problem SpaceX Doesn’t Have

Chemistry has a data problem: most experimental results are negative, inconsistently formatted, and locked behind journal paywalls. Narrow ML tools trained on curated databases have struggled with coverage. Claude sidesteps the curation requirement — it was trained on the scientific literature in its published form, including supporting information sections where NMR data appears.

What Changes for Labs

The inverse NMR result matters for practical chemistry workflows. Characterizing an unknown compound, confirming a synthesis product, or debugging a reaction that produced an unexpected outcome all require reading a spectrum and proposing a structure. If that step can be partially automated by a general model rather than routed to a specialist chemist, the throughput of analytical chemistry increases.

CAS now catalogs more than 290 million disclosed substances and adds 15,000 daily. The gap between chemical knowledge and the human capacity to translate it into testable structures is growing. A model that can interpret NMR spectra at near-expert level closes part of that gap without requiring a new specialist model trained from scratch.

Anthropic frames the current result as a “modest claim” — Claude is starting to assist with translation, recall, and integration tasks that complement expert judgment. The NMR white paper is the first in a planned series of chemistry capability reports.