Anthropic opened Mythos to biologists. Claude sped their tools up 4x

On Wednesday afternoon, @AnthropicAI posted that applications are open for something called the Life Sciences Verification Program. Vetted biology teams can now use its models, "including, for the first time, Mythos," under a looser set of safeguards built for lab work.

Eleven minutes later, Nathan Frey, who works on life sciences at Anthropic, added the plain-English version. @nc_frey wrote that verified teams get "access to our most capable models for professional biology and drug development work," whether they are an academic lab, a startup, or a pharma company.

That evening, @adaptyvbio announced a joint contest with Anthropic that it called "the biggest Protein Design Competition in the world." Biology used to be a category Anthropic's models refused. It is now a customer segment.

What the verification program changes

Until now, Anthropic's public models declined a slice of ordinary biology work because the same requests could come from someone building a weapon. The new program flips the default for people who prove who they are. Applicants submit research credentials, security standards, and an ethics oversight setup, plus a statement of what they intend to do.

There are two grant types. Standard use is team-wide and renews yearly. High-risk use is tied to a single project, renews every six months, and removes the remaining life-science blocks on Opus and Sonnet, with Mythos held back for that tier "pending government coordination."

The safeguard moves rather than disappears. Anthropic says verified users get offline monitoring instead of real-time blocking, with 30 days of retention walled off from training. The beta runs in Anthropic's own console and in the Enterprise and Team plans. It is not on the third-party clouds yet, individuals on Pro and Max plans come later, and organizations with a healthcare data agreement are excluded for now.

The research behind the launch

The same day, Anthropic published the work that makes the program worth applying to. An internal research model, supervised by two staff members who knew biomolecular modeling but had "no prior experience in inference optimization or kernel engineering," rewrote the inference code for more than 30 open-source biology models in under four weeks.

The headline number is a 4x average speedup for structure prediction. The mechanism is a set of custom GPU kernels Claude named FlashPairformer, plus a low-memory mode that folds molecular machines of more than 10,000 tokens, a bacterial ribosome among them, on a single GPU node instead of a cluster.

The cost line is the one I keep rereading. Anthropic says an earlier protein design campaign used about 2,500 H100 hours and a budget of up to $10,000 per target. A single Claude model with one GPU and 24 hours now matches those in-silico scores "with a combined spend of approximately $150."

Anthropic put the 36 optimization kits on GitHub under Apache 2.0. The repo is a reference release, marked not maintained, and the default fast mode introduces small numeric differences that the authors document but that a published paper would need to disclose.

The contest that turns it into practice

Adaptyv's post lays out the terms. Five design problems drop one at a time between September 28 and October 31, aimed at hard targets such as GPCRs and pH-sensitive binders. Anthropic and Adaptyv are paying for wet-lab testing of more than 5,000 designs at no cost to entrants, Anthropic adds $1 million in Claude credits, and every result, failures included, gets published on Proteinbase.

Three tracks run from expert teams down to high-school students. What Anthropic gets out of it is a public record of what its models do in biology, tested in a lab, and produced by people outside the company.

My read

The interesting move is the shape of the access. Anthropic is saying that for dual-use work, the right control is to verify the person and then watch after the fact, rather than refusing everyone at the prompt. That is closer to how a pharmacy or a firearms dealer operates than how a chatbot has.

Two caveats. The speedup and the $150 figure are Anthropic's own measurements against Anthropic's own earlier baseline, and in-silico scores are not binders in a tube. And the work was done by an "internal, general-purpose research model," which you cannot buy today. The kits it produced are real and free. The workflow that produced them is not on the price list.

What this means for your business

Two things carry over to a business nowhere near a lab. First, expect more products that unlock capability only after you verify who you are and what you intend. Have that answer ready, and know which vendors keep the audit trail.

Second, notice what the two supervising researchers did. They didn't buy a new model. They pointed one at the tools they already ran and got the same answers four times faster and far cheaper. For most companies, the AI cost problem sits in the software they already pay for. New Face Design's free process audit looks at what you already run and finds where that kind of speedup is hiding. Start here.