Anthropic has tested three pre-release Claude models, including Opus 5.5, Mythos 5.2, and Mythos 5.1, on their ability to plan synthesis routes for complex molecules. In blind tests conducted with unpublished targets provided by professors from Stanford, Yale, and Caltech, the models successfully generated full synthesis plans.
The models operated without internet access or external literature, relying solely on local code tools like RDKit. Across 12 evaluations, the AI proposed routes consisting of 10 to 35 steps, with each process taking between 26 and 84 minutes to complete. The results were assessed by chemistry professors as comparable to the work of first-year graduate students.
The research, authored by David Kamber, noted that the models demonstrated reasoning and self-correction capabilities. However, one of the 12 runs resulted in a diastereomer error due to a misread bond hash in an input image. This study marks a shift for Anthropic from data interpretation to autonomous laboratory strategy planning.
One of the target molecules featured in the research was a 68-membered macrocyclic archaeal lipid containing 22 stereocenters. By using unpublished targets, the company aimed to address concerns regarding AI benchmarks where models might rely on memorized training data rather than critical thinking.
