Anthropic / Wikimedia Commons (Public domain)
Anthropic says three pre-release Claude models planned routes to four unpublished molecules in blind tests judged by professors at Stanford, Yale, and Caltech
Anthropic says its Claude models can now do something that usually takes years of graduate training: plan how to build a complex molecule from scratch.
In a research paper published on November 9, 2026, the company described blind tests in which three Claude models produced full synthesis routes for four complex target molecules. The targets had never been published. The models worked from images, with code as their only tool.
The verdict from the chemistry professors who supplied the molecules was measured. They rated the outputs as comparable to the work of first-year graduate students in retrosynthetic planning.
What Anthropic actually tested #
Retrosynthesis is the art of working backward from a finished molecule to the simple, buyable chemicals you would start with. For complicated molecules, that recipe can stretch across dozens of reactions. Each step has to work, and each one has to avoid wrecking the parts of the molecule you already built.
The models tested were pre-release versions of Opus 5.5, Mythos 5.2, and Mythos 5.1. Leading chemistry professors at Stanford, Yale, and Caltech supplied the four target molecules. Because the targets were unpublished, the models could not have memorized an answer from their training data.
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.
The setup was deliberately restrictive. The models had no internet access and no external literature. Their only help came from local code tools such as RDKit, a software toolkit for working with chemical structures.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
All of the model outputs were reviewed blind. One of the targets was a 68-membered macrocyclic archaeal lipid with 22 stereocenters.
The one mistake, and why it matters #
The results were not spotless. One of the 12 runs produced a diastereomer error, caused by a misread bond hash in the input image.
In chemical drawings, hashed and wedged bonds show whether an atom points toward or away from the viewer. Misread one of those marks and you plan a route to a molecule that has the same atoms in a subtly different 3D shape. In drug chemistry, that subtle difference can separate a medicine from an inert powder.
Still, the evaluations reportedly highlighted the models’ awareness of stereochemical complexity overall. The research also pointed to signs of thorough reasoning and self-correction during the runs.
The paper’s author is David Kamber, who is part of Anthropic’s chemistry initiatives.
How this fits Anthropic’s science push #
This is not Anthropic’s first chemistry experiment of the year. Earlier research in 2026 tested Claude’s performance on NMR predictions, a core technique chemists use to figure out the structure of a molecule. Another effort looked at the model’s ability to coordinate biomolecular tools.
Those earlier projects were largely about interpreting data. The new work moves into planning: taking a real laboratory target and laying out an autonomous strategy to make it.
The broader initiative aims to push Anthropic’s models deeper into scientific work that demands critical thinking and the synthesis of knowledge, rather than simple recall. Blind, unpublished targets are a direct answer to the most common criticism of AI benchmarks: that the model has simply seen the test before.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.