What Do the New AI Model Releases by Anthropic and OpenAI Change for Life Sciences?

Anthropic led with a laboratory result. Its restricted model, Mythos 5.1, was given open-source protein design tools and asked to produce binders (the molecules that attach to a biological target, and the starting point for most biologic drugs). Its designs were sent to an external organisation for physical testing. 

Roughly half of them bound their targets, against a field norm of 10-15%, and on three targets the binding was ten times stronger than the best entries submitted to Adaptyv Bio's public design competitions.

This is not the first time Claude has done this. Two weeks earlier, Adaptyv Bio published a wet-lab benchmark of an earlier Claude model showing a 26.8 per cent hit rate. What changed is the rate, not the capability. The two studies used different target sets, so the jump from roughly a quarter to roughly a half is a trend rather than a clean comparison, but Anthropic describes it as the best hit rate it has measured.

Anthropic also reported that Mythos 5.1 rewrote the underlying code of seven widely used genomics and protein models (i.e. Evo 2, Enformer and ProGen2 among them) making them run 1.4 - 2.5 times faster and cutting GPU costs by 30 - 60% on genome-wide analyses. The company says it plans to release that work publicly. This is less eye-catching than designing molecules, but probably more broadly useful: most academic labs run unoptimised code because nobody on staff writes GPU kernels.