AINEWS
Fable 5.1 cuts inference cost about 25% for typical workloads; Mythos 5.1 is gated behind a vetted access program Anthropic built with the US government for biology and cyber work.
Anthropic released a new Claude model on Tuesday, in two versions built from the same underlying system but shipped on different tracks. Claude Fable 5.1 is available to anyone willing to pay for it, and the company says it will cost about 25% less than its predecessor for typical workloads. Claude Mythos 5.1 is the same model behind a restricted access program, with a biology lane the company says it built with the US government and a cybersecurity lane it says it controls itself.
Anthropic's price cut on Fable 5.1 comes in three different numbers, and they describe different slices of the same change. The vendor post says Fable 5.1 costs "an estimated 25% less" than Fable 5 for typical token-billed workloads, with savings reaching "up to approximately 45%" for highly agentic work. VentureBeat frames the same change as a 75% reduction, but that is the slice for cache reads, where the model reuses previously processed text instead of recomputing it. A company running long, repetitive agent loops on cached context will see the bigger number; a customer billing mostly fresh tokens will see something closer to the 25% figure. Anthropic does not publish a single dollar amount, so the takeaway for buyers is to test their own traffic mix.
On capability, both models share a coding and knowledge-work stack that the company describes as its new internal standard at lower or equal effort to Fable 5. The same vendor post says Fable 5.1 "can now be used to discover software vulnerabilities — though not develop exploits for them." That is part of the safety story Anthropic is telling around the release: the company is opening up a capability it previously restricted, while drawing a new line at weaponization.
On the cybersecurity side, Anthropic says the safeguards now flag "60% fewer false positives than before." The figure is a vendor self-report, and the System Card PDF lays out the underlying methodology rather than third-party validation. Zvi Mowshowitz's system-card write-up walks through the same numbers and reaches a more cautious read: the company is selling a less jumpy alert pipeline, not a measurable drop in real attacks getting through.
The restricted tier is where the release gets structurally interesting. Mythos 5.1 is "the same model, but with different levels of safeguards," according to Anthropic, and is not generally available. Its biology access program, the company says, was "developed in partnership with the US government," and will start enrolling scientists "soon." A separate cybersecurity lane restricts who can run the model against sensitive systems. Anthropic is positioning this as responsible deployment, but the access decisions about who gets vetted, who gets cut, and what counts as sensitive work are made by Anthropic and its government partner, not by an independent third party.
Enterprise customers get a separate piece of the architecture. A new system Anthropic calls Enterprise Frontier Safeguards, or EFS, stores data in cloud infrastructure "controlled entirely by the customer, not Anthropic," which the company says gives them "complete privacy (the same as a zero data retention policy)." EFS rolls out to enterprise customers in phases starting later this fall; in the meantime, zero data retention is available as a stopgap.
The release, considered as a whole, packages the benchmark chart with a meaningful price cut on the public model, a customer-controlled data path for enterprise buyers, and a separate high-capability tier that the vendor and a government partner decide who can run. Zvi's analysis lands on a similar point from the independent side: frontier models are now shipping with their own access politics attached.
The biology enrollment program is the next test for the policy claims Anthropic is making. The company has not published a timeline beyond "soon," a list of which US government offices are involved, or the criteria for admission. Those details will set the bar for whether the two-track release delivers on its own framing.