Claude Mythos 5 in the Lab: 10x Faster Drug Design and Autonomous Genomics Across 138 Species
Anthropic reports Claude Mythos 5 accelerated drug design roughly 10x, produced strong candidates for 9 of 14 protein targets, and ran autonomous genomics across 138 species.
While Claude Fable 5 dominated launch-day headlines with coding benchmarks, the more provocative results in Anthropic's June 9 announcement came from the laboratory. Claude Mythos 5, the restricted variant with safeguards lifted in some areas, posted scientific results that suggest frontier models are becoming genuine research instruments.
Drug Design at 10x Speed
In Anthropic's reported experiments, Mythos 5 made drug design roughly 10x faster. Across a panel of 14 protein targets, 9 yielded strong candidates - a hit rate that, if it holds up in independent work, would meaningfully change early-stage discovery economics, where candidate generation against difficult targets is a major bottleneck.
Hypotheses Scientists Actually Prefer
Generation speed matters little if the output is mediocre, so Anthropic ran blind comparisons. In molecular biology, hypotheses produced by Mythos 5 were preferred roughly 80% of the time over those from Opus-class models. More striking: one of the model's hypotheses about E. coli was independently corroborated - an instance of a model proposing biology that turned out to be true before humans had established it.
Autonomous Genomics Across 138 Species
The most autonomous result came from comparative genomics. Mythos 5 carried out genomics work across 138 species, and in the course of that work built a custom machine-learning model 100x smaller than comparable approaches that nonetheless outperformed a publication in the journal Science. A frontier model designing a compact, superior ML tool as a subtask of its own research agenda is precisely the kind of recursive capability that both excites and unsettles observers.
Who Gets Access
These capabilities are not publicly available. Mythos 5 is restricted to cyberdefenders and infrastructure providers via Project Glasswing, with select biomedical researchers gaining access later through a trusted access program. The public Fable 5 shares the same underlying model but is wrapped in safety classifiers, including biology and chemistry classifiers that route sensitive queries to Claude Opus 4.8.
The dual-track release reflects the dual-use reality: the same capabilities that accelerate drug design raise biosecurity questions, which is why Mythos-class models also carry a 30-day data retention requirement for safety monitoring, with all human access logged.
For the life sciences, the message of these results is unambiguous: the next generation of research acceleration has arrived - it is just arriving through a controlled door.