Anthropic is broadening frontier AI by refusing to sell every buyer the same access. Claude Fable 5 can reject work researchers call innocuous; Claude Mythos 5 reserves a different operating envelope for trusted organizations.

A uniform safety layer assumed a uniform market

For roughly two years, commercialization followed a common design: build a frontier model, place one safety policy around it, then distribute it as broadly as the guardrails permitted. The model supplied capability; the policy constrained behavior. Everyone, in principle, approached the same boundary.

That design answered the early question: how can one powerful system be made broadly available? As the capability became useful across more consequential settings, one boundary had to perform two incompatible jobs: remain permissive enough for legitimate work while remaining restrictive enough for the widest range of unknown users and uses.

The problem was not a safety team choosing the wrong threshold. No single threshold preserves identical utility across unequal deployment contexts. The public and a trusted organization do not present the same operating conditions.

Claude Fable 5 and Claude Mythos 5 replace that assumption with conditional model access. Fable is the public-safe tier. Mythos is the trusted-organization tier. Anthropic has encoded the distinction in product design, pricing, access rules and technical controls, moving safety from a common wrapper around capability into the terms under which capability is sold.

The guardrail has become part of the product

The split changes what the customer is buying. Under the earlier structure, the commercial object was primarily the model’s capability, with safety experienced as a shared constraint. Under the Fable-Mythos structure, the commercial object also includes the deployment envelope: which controls apply, which access conditions must be met and which capability envelope can be used.

The split also reveals a new purpose for safety controls. Under a uniform policy, they defined the outer boundary of broad distribution. Expressed through separate products and access classes, they also segment the market.

That does not make the partition cosmetic. Product names can be marketing; prices, account rules and technical controls are architecture. Safety is no longer merely a constraint on commercialization. It is one of the mechanisms by which Anthropic organizes commercialization.

The public tier pays a utility tax

The partition resolves one contradiction by exposing another. Cybersecurity researchers say Fable’s guardrails reject tasks they characterize as innocuous, including reading blog posts and reviewing code. Those refusals show that the public-safe boundary can suppress legitimate utility as well as misuse.

That is not a minor implementation complaint. It is the structural cost of asking one public tier to absorb uncertainty about identity, intent and deployment context. When the system cannot distinguish a benign request from a dangerous one with sufficient confidence, restriction becomes the balancing mechanism. The safer boundary is not free; some of its cost appears as work the product declines to do.

Anthropic’s red-team testing found no universal jailbreaks, but that result is narrower than durability. It neither proves that the safeguards resist every form of abuse nor erases the false positives reported by researchers. Refusal accuracy and jailbreak resistance are different properties.

The split-market design does not eliminate this trade-off. It assigns it. Public users receive a more restrictive envelope, while trusted organizations receive a different combination of capability, controls and access requirements. Risk has not disappeared; it has been redistributed across product tiers.

The utility collision forced a trust gate

The sequence changes the unit of governance. The decision moves upstream, from whether an individual prompt passes a filter to whether an organization qualifies for a different operating envelope. The access gate now carries part of the burden once assigned entirely to the refusal layer, making deployment accountability central.

The old strategy backfires when legitimate utility and misuse risk spread too far apart. Anthropic’s response is not to remove the boundary, but to multiply it.

Safety now has a price sheet

Frontier AI is entering neither a uniformly open phase nor a uniformly closed one. The same underlying capability is being commercialized broadly after being divided into public-safe and trusted-organization tiers. Safety controls now live in the catalog, the price, the account relationship and the technical deployment.

Anthropic can broaden the frontier precisely because it no longer opens it the same way to everyone. Fable’s refusal screen and Mythos’s access gate are two pieces of one commercialization system: two doors, with safety printed on the price sheet and trust checked at the handle.