Claude Just Matched 40 Years of Chemistry Software

Claude just matched chemistry software that took 40 years to build, with no chemistry training at all.

Anthropic published a benchmark on 5 June. They put Claude Opus 4.7 against ChemDraw and MestReNova, the specialist NMR tools every medicinal chemist has used for decades, on 20 novel compounds pulled from preprints after the model's training cutoff.

On hydrogen shift prediction Opus 4.7 was the most accurate tool in the test. On carbon it tied the best of the dedicated software. A general model, no fine-tuning, level with instruments built specifically for this one job.

Then they ran it backwards. Given only a spectrum and a molecular formula, could the model name the structure that produced it. This is the harder direction, the one chemists still do by hand. Opus 4.7 recovered all eight of the simpler structures on every attempt, and four of the seven harder ones cleanly. I want to be careful about what this does and does not mean. The sample was small, 20 compounds and four scaffold classes. The authors say so plainly. 2D experiments and stereochemistry were out of scope. This is one analytical step, not a chemist.

But step back and the direction is the story. The value in tools like ChemDraw was decades of encoded domain knowledge. The result here is that a great deal of that knowledge now sits inside a general model that nobody pointed at chemistry on purpose. The moat was the specialty. The specialty is becoming a default capability.

I wrote in December that the real bottleneck in AI drug discovery was never the models. It was turning medicinal chemistry from a craft held in a few senior heads into a machine-readable system. Every manual translation step that collapses into something a model does in seconds is the craft becoming the system, one step at a time.

Structure elucidation is one step. There are many. The interesting question for anyone building or buying techbio tooling is which specialty is load-bearing in three years, and which is about to become a feature.