Anthropic tests Claude on chemistry syntheses absent from its training data

Anthropic has documented blind tests of Claude planning total syntheses of molecules that were never published, a probe of real chemistry reasoning.

Anthropic has released a document describing blind tests of Claude on unpublished total syntheses, meaning routes to complex molecules whose syntheses do not appear in any published literature the model could have seen in training.

The design matters because a model that reproduces a known route is recalling, while one that plans a route to a molecule it has never seen described is doing something closer to reasoning. Testing on unpublished work is a way to tell the two apart.

For abundance, the relevance is in drug and materials discovery. Making a new molecule is often a slow, expert-heavy bottleneck, and planning how to build it is a large share of that effort. If AI can do that planning reliably, new medicines and materials could be explored faster and at lower cost, which touches both knowledge and health.

Caveats are large. We have only the document's title and overview framing, not its results, so we cannot say how well Claude performed, how many cases were tested, or how the routes were judged. A plan on paper is also not a molecule made in a flask. The full account sits with Anthropic's document.

Source: Anthropic