An AI model considered dangerous enough to be kept behind limited-access programs has a version for ordinary paid users, called Claude Fable 5. Its maker, Anthropic, includes it at no extra cost on Pro, Max, Team and seat-based Enterprise plans through June 22.
Fable 5 and Mythos 5 share the same underlying model, from a tier above the company’s Opus models. What differs is access and safeguards. Mythos 5 stays limited to selected cyberdefenders, infrastructure providers and future trusted access programs, while the public version carries extra controls around sensitive areas.
If a request is flagged as related to cybersecurity, biology and chemistry, or attempts to copy the model, a separate model, Claude Opus 4.8, answers instead. Anthropic says this switch happens in fewer than 5% of sessions, though it also says the filters err on the side of caution and may catch harmless requests.
Away from those sensitive areas, Anthropic’s strongest public claim for Fable 5 is in software engineering. Stripe, the payments company, tested it on 50 million lines of Ruby code. The model completed a migration across all of that code in one day, work that would otherwise have taken a team more than two months by hand.
The product page describes the model as built for long-running projects. It runs inside agent tools such as Claude Code or Claude Managed Agents, where AI carries out tasks on its own. There it can plan across stages, hand off smaller tasks, write tests and check its own work. At Lovable, an app-building company, the technology chief said apps that took a hundred prompts a year earlier could be built from one. That is one customer’s account, not an independent test.
Many AI projects still need a person to track the whole job, including what failed and which files changed. Fable 5 is pitched as better at holding that larger picture in view. AI models read and write text in small chunks called tokens, and this one is said to stay focused across millions of them.
Cost is the constraint. Once the free period ends, paid users will need usage credits, unless Anthropic has the capacity to extend it. Both versions are listed at $10 per million tokens the model reads and $50 per million it writes.
