Valen Labs drops FrontierBioML biology benchmark

Standard biology benchmarks are saturated multiple-choice trivia tests that fail to test actual scientific execution. FrontierBioML forces agents into end-to-end bioinformatics workflows over multi-hour horizons where current frontier models barely clear a 40% success rate.

• –Multi-hour execution horizons test real agentic stamina and pipeline debugging rather than single-shot prompt memorization.
• –Sealed held-out test sets prevent evaluation contamination from publicly available genomic datasets and Kaggle solutions.
• –Claude Mythos 5.2 topping out at 40.3 demonstrates that autonomous computational biology remains a major bottleneck for frontier LLMs.
• –Prime Intellect Environments Hub hosting provides an immediate, standardized reinforcement learning environment for teams training specialized science agents.
• –Tight clustering between Claude Mythos and Gemini 4 models indicates current foundation models share common failure modes in long-context scientific problem solving.