Anthropic ships Claude Fable 5, with a locked-down Mythos 5 for cyberdefenders

Anthropic is putting a Mythos-class model on the open API for the first time. Claude Fable 5 lands at $10 per million input tokens and $50 per million output tokens — Anthropic says that's less than half the price of Claude Mythos Preview, the limited-release model that preceded it.

The pitch to builders is long-horizon agentic work. Anthropic claims Fable 5 "can work autonomously for longer than any previous Claude models," and the customer quotes are uniformly about multi-step coding and analysis rather than chat. Stripe says Fable 5 ran a codebase-wide migration on a 50-million-line Ruby codebase in a day; Cognition says it's the top scorer on its FrontierCode evaluation "even at medium effort," implying real token-efficiency gains. GitHub frames it as a step toward handing "increasingly ambitious work to agents."

A few things worth flagging for anyone planning to wire this in:

• The safeguard routing. Anthropic says some queries will silently be answered by Opus 4.8 instead of Fable 5 when classifiers flag them. They've "tuned these safeguards conservatively" and admit they "sometimes catch harmless requests," though they say it triggers in "less than 5% of sessions" on average. For production agents, that's a non-trivial behavioral inconsistency to plan around — your eval suite and your live traffic may not get the same model.
• Memory and long context. Anthropic highlights persistent file-based memory as a force multiplier, citing a Slay the Spire test where giving Fable 5 notes improved performance three times more than it did for Opus 4.8. The implication for agent designers: scaffolding that worked for prior Claudes may be under-using this one.
• Vision with less scaffolding. The Pokémon FireRed run — completed from raw screenshots with no map or game-state helpers — is the load-bearing demo here. If it generalizes, harnesses for browser, desktop and game-style agents get thinner.
• Mythos 5 is the same underlying model. Anthropic is explicit about this: Mythos 5 is Fable 5 with some safeguards lifted, deployed initially through Project Glasswing with the US government as an upgrade to Mythos Preview, with a broader "trusted access program" promised. The capability ceiling you can buy on the open API is, by design, lower than the one Anthropic itself is shipping to cyberdefenders.

Alignment numbers come from Anthropic's own automated assessment, which puts Mythos 5's misaligned-behavior rate "similar to that of Opus 4.8." Take that as a self-report pending the system card and outside replication.

The through-line: if you're building long-running agents, this is the first time Anthropic is putting its top-tier capability tier on the price list — with a silent fallback model attached.