Claude takes on chemistry

If you have ever spent a late night in an organic chemistry lab, you know the quiet torture of NMR interpretation.

Matching messy instrument peaks to molecular drawings is tedious, and while software has simulated spectra for decades, working backward from an experimental readout to an unknown structure has remained a job left entirely to human brains.

Anthropic wants Claude to take over that burden. In a new white paper, Anthropic chemist David Kamber evaluated Claude Opus 4.7, Opus 4.6, and Sonnet 4.6 against standard chemistry staples ChemDraw and MestReNova.

Beating dedicated software at its own game

To ensure the models were not just reciting memorized literature, Anthropic pulled 20 novel compounds across four structural families from ChemRxiv preprints published after the models' training cutoff.

Given only a SMILES string and the reaction solvent, Opus 4.7 edged out both legacy tools on 1D hydrogen chemical shift predictions, averaging an error of just ±0.079 ppm. On carbon predictions, Opus 4.7 effectively tied MestReNova at ±1.37 ppm versus ±1.48 ppm.

The real divide showed up in the hard cases. On a notoriously tricky NH proton in a chloropyridazine family, Opus 4.7 stayed tightly anchored near the true 6.8 to 7.9 ppm window. Opus 4.6 scattered guesses across several ppm, while Sonnet 4.6 hallucinated way out in the 10 to 13 range.

Claude also crushed peak splitting metrics. All three Claude models predicted sub-peak spacing within half a hertz around 80% of the time, compared to just 26% to 35% for ChemDraw and MestReNova.

Cracking molecules in reverse

Forward prediction is one thing, but inverse structure elucidation is the real prize. Dedicated tools usually demand complex 2D NMR datasets, expensive site licenses, and specialized training.

Anthropic fed Opus 4.7 a high-resolution mass spec formula alongside 1D hydrogen and carbon peak lists for 15 compounds, asking it to rank candidate structures.

For eight simpler targets, Opus 4.7 nailed the exact structure on every single run with zero clues. On seven denser targets like fused rings and spirocycles, given only the reaction starting material as a hint, it got four right across all three runs and solved the remaining three on two out of three attempts.

The study is small, covering just 20 compounds for forward prediction and 15 for inverse tasks, meaning chemists should treat these as early indicators rather than definitive benchmarks.

Why it matters

What makes this remarkable is that Opus 4.7 is a general-purpose model without chemistry-specific fine-tuning. Yet right out of the box, it matched or beat software that analytical chemists have relied on for decades.

For bench chemists and biotech startups, the immediate value is massive. Instead of paying steep seat licenses for specialized elucidation packages or wrestling with tricky 2D experiments, researchers can paste messy 1D peak lists into a chat window and get an instant second opinion. It is absolutely worth testing on your backlog of unassigned spectra today.

For scientific software incumbents, this is a serious warning shot. Their moat has always been handcrafted heuristics and curated chemical databases. A frontier model with strong step-by-step reasoning just demonstrated that broad multimodal intelligence can leapfrog dedicated analytical engines.

We think Anthropic is proving that frontier models do not need bespoke niche architectures to solve hard science. The obvious caveat is the modest sample size, but the trajectory looks clear: AI is finally ready to do actual chemistry.