Latch researchers introduce TxBench antibody benchmark

Researchers built TxBench–Antibody Development, a benchmark featuring 60 real decisions from seven stages of antibody drug development, to test how well AI agents can handle complex therapeutics workflows. Evaluating 34 model and harness setups, they found that the highest pass rate was 39.4% by Grok 4.7, and that decisions became significantly harder as they approached patient testing, with toxicology and first-dose decisions passing only 13% of the time. Open article →