AXI Chemotype 1.0: Evaluating Molecular Intelligence in Frontier AI

August 18, 2026

From chemical language to molecular intelligence

Frontier AI can speak confidently about chemistry. But reliable molecular reasoning requires something deeper: interpreting structure, tracking subtle changes, and understanding when nearly identical molecules represent meaningfully different chemical identities.

AXI Chemotype 1.0 is an evaluation set designed to measure different models' ability to distinguish these subtle structural differences, serving as a direct test of drug molecule understanding. Using molecular images and SMILES, the benchmark directly tests essential chemical-reasoning capabilities that general-purpose benchmarks rarely measure, including stereochemistry, symmetry, structural comparison, and the identification of meaningful substructures.

High-Precision Molecular Recognition: Model distinguishes the reference drug from closely related stereochemical variants, preserving fine-grained structural and stereochemical information required for downstream analog identification and drug-similarity reasoning.

Among structures A–D, only one represents ramipril, an approved ACE-inhibitor prodrug. The other three are stereochemical variants created for evaluation. AXI Chemotype 1.0 tests whether AI models can distinguish the known drug from these nearly identical alternatives by recognizing the precise stereochemical differences that define molecular identity.

The evaluation reveals an important gap: a model may recognize a chemical feature in isolation, yet struggle to locate and interpret that same feature within a larger molecular structure. This distinction matters because molecular discovery depends on relationships, context, and subtle structural differences, not simply labels. This challenge is reflected in the leaderboard for our curated 200-question set: GPT-5.5 achieved 18.75% accuracy, while Gemini 3.5 Flash achieved 12.25%.

Our work is now extending beyond structural recognition toward drug intuition: whether AI can connect molecular variation with properties, analog relationships, supporting evidence, and biological context. The workflow is built around automated, reproducible, and modular pipelines, enabling consistent evaluation and efficient expansion across new molecules, targets, and discovery settings. AXI Chemotype 1.0 represents our broader vision for scientific evaluation: moving beyond what AI can say about chemistry toward measuring what it can reliably understand and reason through.