Intology says its Locus agent beat Claude Code and Codex at building models for molecular design
On six targets covering property prediction, small-molecule potency and antibody design, Locus reached a mean Spearman of 0.622 against 0.569 for Codex and 0.450 for Claude Code, by the company's own tests.
Why it matters
Locus is an agent that runs machine-learning research on its own: it writes, trains and tests models rather than answering chemistry questions directly. Intology compared it with general coding agents and the Boltz-2.1 structure model on the same tasks and compute, and says it also beat BioHub and Boltz baselines. Spearman measures how well predicted rankings of molecules match lab results. The results are the company's own and have not been independently checked; watch for external replication.
Key facts
- Mean Spearman 0.622 across six targets
- Codex (GPT-6 Astra Max) 0.569; Claude Code (Opus 5 Max) 0.450; Boltz-2.1 API 0.225
- Tasks: property prediction, small-molecule potency, antibody design
- Matched compute against baselines
- Company-run tests, not independently verified