Claude Fable 5 Explained: Pricing, Capability Filters, and Data Retention Changes

Anthropic announced Claude Mythos in April 2026, but held back a general release because its offensive cybersecurity capabilities were considered too strong. About two months later, on June 9 (June 10 in Japan time), a Mythos-class model finally reached the market.

This launch was unusual because the same underlying model was released in two different forms. The full-capability version, "Mythos 5," is restricted to a narrow set of trusted use cases such as security work. The public-facing release is "Fable 5," which applies capability filters. Only the latter is broadly accessible.

Yasuno argues that Fable 5 ranks first across many major benchmarks, and not just by slim margins. In several areas, he says, it opens up a visibly larger lead over competing models such as ChatGPT and Gemini.

As a concrete example, he points to the migration of Stripe’s 50-million-line Ruby codebase. Work that would have taken more than two months manually was reportedly completed by Fable 5 in a single day. He also highlights demos ranging from automating the game Factorio, to designing printable models in browser-based 3D CAD, to a fluid simulator synced to EDM beats, as signs of unusually broad applicability.

Compared with Opus 4.8, token pricing is roughly double: $10 per million input tokens and $50 per million output tokens. In yen terms, that works out to around 1,600 yen for input and 8,000 yen for output.

Because leaked Mythos preview pricing had suggested $25 input and $125 output, Yasuno says he expected it to come in even higher. Even so, doubling versus the previous generation is enough that finance teams at companies scaling AI use will need to revisit next year’s token budgets. The segment frames this as a broader planning problem: how to set the right balance between AI and human labor while assuming the models will keep improving over the next year.

Fable 5 includes capability filters across four domains: cyberattacks, biology, chemistry, and AI development. Cyber was already the core concern. Biology and chemistry were added because of possible misuse involving dangerous substances or DNA/RNA synthesis. AI development restrictions were introduced to block distillation by rival labs.

The distillation response is the most novel part. If Anthropic decides another company is using its model to collect input-output data for training a different model, it can silently degrade performance. Because the company does not normally disclose whether the downgrade is happening, a firm attempting distillation may not easily notice that the model has been weakened. Yasuno describes this as a sign that competition among AI companies has entered a new phase.

For now, Fable 5 is available as part of subscription plans, but after June 23 it will move to usage-based billing. Yasuno recommends using the remaining two weeks as a de facto evaluation period, especially to surface issues hidden inside existing codebases all at once.

There is also an operational caveat: the data retention policy changes. Fable 5 retains data for 30 days for safety verification. That alters the assumptions for organizations that had been operating enterprise plans on a no-retention basis. IT departments will want to review how that fits with their internal data policies.