Claude Mythos 5,
and why you cannot have it.
Mythos 5 shipped on the same day as Fable 5, with the same specifications and the same price, and almost nobody reading this can access it. That gap is the most informative thing about the release.
Mythos 5 is the same underlying model as Fable 5 with certain safeguards lifted, intended for authorised users such as cyberdefenders and infrastructure providers. It is being deployed initially through Project Glasswing, in collaboration with the US government, with a trusted access programme for biology researchers planned and broader access flagged as later. It is invitation only.
The capability and the guardrails are separable
The two models are not different sizes or different training runs. They are the same thing with a different safety configuration, which makes something explicit that is usually left implicit: what you can do with a frontier model is set by a deployment decision as much as by the model itself. Fable 5 routes some requests elsewhere and declines others. Mythos 5 does less of that, for a small number of organisations.
Access is a policy question now, not a payment one
For most of the last few years, the frontier was available to anyone with a card. This is a different arrangement: the most capable configuration goes to specific institutions for specific purposes, starting with cyberdefence and government work. Whatever you make of that, it is a change in how capability reaches the market, and it is worth noticing while it is still novel.
What it means for your planning
Build on what you can actually get. Fable 5 is the ceiling for general commercial work, and designing a workflow around capability you have read about but cannot buy is a good way to waste a quarter. If your organisation genuinely sits in one of the named categories, the route in is an application, not a purchase order.
The gap between frontier and available is now visible
Anthropic has been open that Fable 5 is the version made safe for general use, which tells you the two are not identical in what they will do. That was probably true of earlier releases as well; the difference is that it now has a name and a separate model identifier. Useful context the next time somebody quotes a benchmark at you without saying which configuration produced it.
The part worth sitting with
A safeguard that can be lifted for one customer is a policy, not a property. That is not a criticism, and honestly it is preferable to the alternative of pretending the safeguards are physics. It does mean the sensible question about any AI vendor is no longer only what the model can do, but who decides what it will do, and on what basis.
Boffin Hub trains people on tools they can actually use, and Mythos 5 is not one of them. It is here because the two-tier release is the clearest signal yet about how frontier capability is going to be distributed, and that shapes what your team should plan around for the next year. Our position on the trust side of this is set out on our ethics page.
Things people
usually ask.
Access is limited to approved organisations, starting with Project Glasswing partners, with a trusted access programme for biology researchers planned. If you are not in one of those categories, the practical answer today is no.
It is the same underlying model. The difference is in the safeguards applied at deployment rather than in raw capability, which means the two behave the same on ordinary work and differ at the edges Anthropic restricts.
Not directly. It matters for planning rather than practice: assume the model you can buy is the model you will be building on, and keep your workflows portable enough to move when the line shifts again.