Anthropic's Claude Beats ChemDraw at Reading Molecular Spectra Without Chemistry Training

Anthropic has published its first chemistry-focused research as part of an expanded science effort, and the headline finding is striking: a general-purpose language model with no chemistry fine-tuning is now competitive with the specialized desktop software chemists have leaned on for decades. The white paper pits Claude against ChemDraw and MestReNova on nuclear magnetic resonance spectroscopy, the workhorse technique for figuring out what molecule you actually have in a flask.

Claude for Life Sciences already exists, but chemistry has been a notably harder nut to crack. The data is messy, locked behind paywalls, and scattered across PDFs and supporting information. Anthropic's argument is that multimodal frontier models change which problems are tractable despite that data shortage, because they can read structures off a figure or sketch and reason through experimental sections in the form they were actually published.

What NMR is, and why it eats chemists' time

NMR spectroscopy is one of the most time-consuming steps in synthetic chemistry; for every compound, a chemist has to match each peak in the spectrum to an atom in the proposed structure by hand. The technique probes a molecule with magnetic fields and radio waves, producing a row of peaks whose positions (measured in parts per million, or ppm) encode where each hydrogen and carbon atom sits in the structure.

There are two directions you can run this. Forward prediction takes a drawn structure and simulates what the spectrum should look like. Inverse prediction, also called structure elucidation, goes the other way: given an experimental spectrum, propose the molecule. Both ChemDraw and MestReNova do forward prediction, using a drawn structure to simulate what NMR spectrum will be produced. The inverse problem is much harder and has traditionally been left to the chemist.

The benchmark setup

To avoid contamination, the team pulled fresh compounds from ChemRxiv preprints posted after the model training cutoffs. The 20 span four structural families, five compounds each, with each family selected because it involves a different category of NMR challenge. Each tool received a SMILES string (the line-of-text notation for a molecule) and had to predict where every hydrogen and carbon peak would fall, in the same solvent the original paper used.

Three Claude models were tested: Opus 4.7, Opus 4.6, and Sonnet 4.6. Because a language model's output varies between runs, each Claude model was queried three times per compound and averaged; ChemDraw and MestReNova return the same answer every time and were run once. The acceptance window was tight: ±0.20 ppm for hydrogen or ±1.0 ppm for carbon.

How the numbers came out

On hydrogen shifts, Claude pulled ahead. Opus 4.7 was most accurate, with an average error of ±0.079 ppm, well under half the tolerance window, and the highest share of peaks landing inside it. On carbon, it was a near-tie: Opus 4.7 and MestReNova were effectively tied, at ±1.37 and ±1.48 ppm; the remaining tools kept the same rank order on both elements.

The more interesting gap shows up on splitting patterns, which describe how a peak breaks into sub-peaks based on neighboring atoms. That shape carries structural information a chemist reads alongside the position. Opus 4.7 matched the experimentally reported splitting pattern more often than any other tool, and all three Claude models predicted the sub-peak spacing to within half a hertz roughly 80% of the time, against 26 to 35% for ChemDraw and MestReNova.

The inverse task is where the result gets unusual. Each supplied the compound's exact molecular formula (from high-resolution mass spectrometry) and its hydrogen and carbon NMR spectra. Opus 4.7 recovered all eight simpler structures on every attempt from spectra and formula alone. On the seven harder targets, given the starting-material hint, it returned the correct structure on all three runs for four of them and on two of three runs for those that remained.

Why this matters more than a benchmark

Dedicated structure-elucidation software is not new. What is new is the setup cost. Dedicated structure-elucidation software has existed for decades, but it typically requires 2D NMR (a spectrum with two axes, and the output is a contour map rather than a row of peaks), specialized training, and licensed tools. Claude does it from the same high-resolution mass spectrum and 1D peak list a chemist would paste into a chat, with no setup.

That is the practical breakthrough. Bench chemists who never adopted retrosynthesis tools or commercial elucidation packages can paste a peak list into a chat window and get a candidate structure back. The model also shows its reasoning, which means a chemist can audit each step.

Where it still falls short

The team is candid about the limits:

• The evaluation was small, 20 compounds across four scaffolds for the forward task, 15 for the inverse task, and each scaffold contributes a single class of failure modes.
• On the densest inverse targets, without the starting material as an additional input, the model could loop through its reasoning without committing to a final structure.
• 2D experiments (COSY, HSQC, HMBC) and stereochemistry are out of scope by design, since 1D NMR alone cannot fix configuration. That rules out the complex natural products where structure elucidation gets genuinely hard.
• Solvent coverage was limited to DMSO-d₆, CDCl₃, and D₂O, so methanol-d₄, benzene-d₆, and acetone-d₆ are not assessed.

Sonnet 4.6 was notably weaker than the Opus models on hard peaks, scattering its guesses by several ppm on one notorious NH proton. So the result is specifically a frontier-model story, not a generic LLM story.

The broader chemistry roadmap

NMR is the first piece. Anthropic flagged several other bottlenecks it wants to chip away at: converting chemical structures from figures, slides, and patents into machine-readable form; reaction and retrosynthetic reasoning; mechanism explanations with electron arrows and intermediates; and pulling chemistry out of method sections, supporting information, and patents where the same molecule may be drawn, named, abbreviated, or referenced by code.

This sits inside a broader push. Opus 4.7 itself shipped with sharper vision capabilities aimed at exactly this kind of work, with image processing up to 2,576 pixels on the long edge, more than three times the fidelity of prior models, which is critical for agents that need to read dense screenshots or extract data from complex technical diagrams, such as chemical structures or life-science patent workflows. Anthropic is also working closely with scientists through its AI for Science program, which provides free API credits to leading researchers working on high-impact scientific projects around the world, and is now opening that program to chemistry researchers explicitly.

For anyone working at a bench, the read is straightforward: routine 1D NMR interpretation and first-pass elucidation are now things you can offload to a chat window with reasonable confidence on the easy cases, knowing that complex stereochemistry and natural products still need 2D experiments and a trained eye.