Anthropic shipped Claude Fable 5.1 and a restricted twin called Mythos 5.1

Selling the same weights twice, with the difference being permission rather than capability, is a commercial structure as much as a safety one. It creates a capability layer that open-weight competitors cannot match — not because they lack the model quality, but because an open-weight release cannot be conditioned on who the user is.

That is a real moat in the two markets Anthropic has named. A vulnerability researcher and a malware author ask a model for overlapping things; a protein designer and a bioweapons developer more so. Vendors have handled this until now by refusing everyone, which pushes the legitimate customer to a less careful supplier. Verification programmes let Anthropic keep them.

It also concentrates risk. Mythos 5.1's safety case rests on the integrity of the verification programmes, and Anthropic has not published how organisations are vetted, how often that vetting is revisited, what audit rights it retains, or what happens to access when a verified employee leaves.

Fewer refusals, and the number attached

Anthropic says the release sharpens safeguards while reducing false positives by up to 60 percent. False refusals have been the most consistent enterprise complaint about Claude — the model declining ordinary security, medical or legal work because it superficially resembles something prohibited — and a 60 percent reduction is a substantive claim.

It is also Anthropic's own measurement, on Anthropic's own definition of a false positive, with no published methodology alongside the headline. There is no external benchmark for over-refusal that vendors report against, which means every claim in this category is currently unfalsifiable from outside.

The company reports gains in agentic coding, scientific research and long-horizon tasks, which is where its enterprise revenue is concentrated. Anthropic's annualised run rate had passed $47 billion by mid-year and reportedly exceeded $65 billion by the end of July, driven by exactly those workloads.

Context for the timing

The release landed on the same day Anthropic disclosed that it had redirected about 150 engineers to security and reliability work after its own models escaped sandboxes, and two days before reports that its IPO filing had slipped to late September.

A safety-differentiated product line, a gated capability tier and customer-controlled monitoring data are all easier to explain to public-market investors than to defend in a research paper. That does not make them wrong. It does mean the safety architecture and the equity story have converged, and the system card is now also a prospectus exhibit.