Anthropic: pre-release Claude models planned synthesis routes for 4 unpublished molecules
In brief
- Three pre-release Claude models planned routes for four unpublished molecules, per Anthropic's November 9 paper.
- Models worked from images with local code tools only, without internet access or literature.
- Chemistry professors who supplied targets rated outputs comparable to first-year graduate students, Anthropic said.
- Routes ranged from 10 to 35 steps across 12 runs, each taking 26 to 84 minutes.
- One run produced a diastereomer error, traced to a misread bond hash in the input image.
What Anthropic says it tested
Retrosynthesis means working backward from a finished molecule to simple, purchasable starting chemicals. For complicated molecules, Crypto Briefing reported, that recipe can stretch across dozens of reactions.
According to Anthropic's paper (authored by David Kamber, who's part of the company's chemistry initiatives), the models tested were pre-release versions of Opus 5.5, Mythos 5.2 and Mythos 5.1. Chemistry professors at Stanford, Yale and Caltech supplied the four targets, and the company said none of them had ever been published. That's the point of the design. Because the targets were unpublished, Crypto Briefing noted, the models couldn't have memorized answers from their training data.
The models didn't have internet access or external literature. Their only help came from local code tools such as RDKit (a software toolkit for working with chemical structures), the paper said.
The numbers, as Anthropic reported them
Across 12 separate evaluations, the models proposed complete routes ranging from 10 to 35 steps, depending on the target and the model. Each run took between 26 and 84 minutes, and all outputs were reviewed blind. One target was a 68-membered macrocyclic archaeal lipid with 22 stereocenters.
The runs weren't flawless.
One of the 12 produced a diastereomer error, caused by a misread bond hash in the input image, according to the report. It's not a cosmetic slip: in drug chemistry, Crypto Briefing noted, a subtle difference in 3D shape can separate a medicine from an inert powder. The paper still pointed to the models' awareness of stereochemical complexity overall, along with signs of thorough reasoning and self-correction during the runs. These are Anthropic's own results, as described in its paper.
Part of a wider chemistry push
The work follows earlier Anthropic research in 2026 that tested Claude on NMR predictions (a technique chemists use to determine molecular structure), plus a separate effort looking at the model's ability to coordinate biomolecular tools. Crypto Briefing reported that the broader initiative aims to push Anthropic's models into scientific work that requires critical thinking and knowledge synthesis, not simple recall.
Frequently asked questions
What is retrosynthesis?
Retrosynthesis means working backward from a finished molecule to simple, purchasable starting chemicals. Crypto Briefing reported that for complicated molecules, the resulting recipe can stretch across dozens of reactions.
How did Anthropic set up the Claude chemistry tests?
According to Anthropic's paper, chemistry professors supplied four unpublished target molecules. Pre-release versions of Opus 5.5, Mythos 5.2 and Mythos 5.1 worked from images without internet access or external literature, using only local code tools such as RDKit. All outputs were reviewed blind.
Did the Claude models make any mistakes in the tests?
Yes. Anthropic's paper says one of the 12 runs produced a diastereomer error, caused by a misread bond hash in the input image. Crypto Briefing noted that in drug chemistry, a subtle difference in 3D shape can separate a medicine from an inert powder.
