Claude Mythos 5: The Model Anthropic Decided Not to Give Everyone

TL;DR — Claude Mythos 5 is, by Anthropic’s account, its most capable model: state-of-the-art at cybersecurity, biology, and healthcare research. It can design drug candidates and propose novel science. And precisely because of that, the company decided most people will never get to use it directly.

We are used to AI news that follows one script: a model gets more powerful, then it gets released. Claude Mythos 5 breaks the script at the second step. Anthropic built a system capable enough to make experts nervous, and then, deliberately, did not hand it to the public.

That refusal is the most interesting thing any AI lab has done in a while.

What Mythos 5 can do

The capabilities are genuinely striking. Anthropic describes Mythos 5 as state-of-the-art at cybersecurity, biology research, and healthcare. In testing, the model generated viable drug-design candidates for 9 of 14 protein targets it was given. On open-ended science, researchers preferred its novel molecular-biology hypotheses to the alternatives in roughly 80% of head-to-head comparisons.

Read those two sentences again. A general-purpose model is now producing original scientific hypotheses that working scientists rate as better than the competition, and proposing drug candidates against real targets. That’s not autocomplete. That’s a research collaborator.

A scientist working in a modern biology research laboratory — Photo by Dmitriy Suponnikov on Unsplash

The same power, pointed the other way

The uncomfortable truth about a model that’s brilliant at biology and cybersecurity is that those skills don’t come with a conscience attached. The capability that designs a drug can probe for a pathogen. The capability that finds a software vulnerability to fix it can find one to exploit it. This is the dual-use problem, and Mythos 5 is the sharpest version of it the industry has shipped.

Anthropic’s response was to split the model in two. The unrestricted system is Mythos 5. The public, fenced version is Claude Fable 5, which routes cybersecurity and biology queries to the older, less capable Opus 4.8 instead of answering with the full model. Same brain; very different access.

Access by invitation only

Getting near Mythos 5 directly means clearing a bar. Access is limited to a small group of vetted partners, through programs Anthropic refers to as Project Glasswing, and to enrolled biology researchers. Use comes with a 30-day data-retention policy so the company can monitor for misuse. Pricing matches Fable 5 at $10 per million input tokens and $50 per million output tokens, but the gate here isn’t money. It’s trust.

It’s a posture that lines up with the threats our cybersecurity desk keeps flagging: the most valuable raw material for an attack is increasingly just capability that’s already been built, and the question is who gets to point it.

A secure data center server room lit in cool blue light — Photo by Tyler on Unsplash

A real test of the safety pitch

For years, AI labs have talked about safety in the abstract: red teams, alignment research, responsible scaling policies. Mythos 5 turns the talk into an actual, visible trade-off. Anthropic had a more capable model it could have shipped, would have generated headlines and revenue by shipping, and chose to withhold the unrestricted version because it judged the risk too high.

You can be skeptical, and you should be. “We built something too dangerous to release” is also, conveniently, the best marketing a frontier lab could write. The restriction is self-imposed and self-policed, and a 30-day retention log is not the same as oversight. But it’s a concrete decision with real commercial cost, which is more than the field usually offers.

What this means

Mythos 5 marks a line the industry has been walking toward: the point where a lab’s most capable model is too consequential to release the way the last one was. Whether Anthropic’s two-tier answer, gate the powerful version, ship a fenced one, becomes the norm or an outlier depends on what its rivals do next, and on whether the fence around Fable 5 actually holds.

Either way, the interesting question in AI has quietly changed. For a long time it was “how capable can these models get?” With Mythos 5, it’s becoming “who decides who gets to use them?”