Anthropic released Claude Fable 5 on Tuesday, and the headline is not the benchmarks, it is the class. Fable 5 is the first generally available model from the Mythos tier, the capability class above Opus that Anthropic has kept gated since April's Claude Mythos Preview, when access was restricted to Project Glasswing partners like Microsoft, Amazon, and Cisco over safety concerns, reporting NBC News summed up as a model family whose cyber capabilities spooked the US government. Anthropic's own framing is unambiguous: Fable's capabilities exceed those of any model the company has ever made generally available, state-of-the-art on nearly all tested benchmarks, with the gains concentrated in software engineering, knowledge work, vision, and scientific research. The API identifier is claude-fable-5, priced at $10 per million input tokens and $50 per million output, which Anthropic notes is less than half the price of the Mythos Preview it supersedes. A companion model, Claude Mythos 5, ships the same underlying weights with safeguards selectively lifted, and stays restricted.
The genuinely new thing is the safety architecture. Fable 5 does not handle high-risk requests by refusing them, it handles them by downgrading itself. AI classifiers watch for three categories, offensive cybersecurity work, certain biology and chemistry applications, and attempts to distill the model, and when one fires the session falls back to Claude Opus 4.8, until today the most capable publicly available Claude, which answers in Fable's place. Anthropic says this triggers in less than 5% of sessions on average. Call it graceful capability degradation: instead of the binary refuse-or-comply that has defined LLM safety since the beginning, the ceiling of what you are talking to moves with what you ask. The red-team record suggests the design was stress-tested before shipping, an external bug bounty exceeding 1,000 hours produced no universal jailbreaks, with the honest caveat that the UK AI Security Institute made initial progress toward one in brief testing. There is also a data-policy change worth knowing: all Mythos-class traffic carries a mandatory 30-day retention window, used only for safety monitoring and defense against attacks, not for training, and deleted after 30 days in almost all cases.
The capability receipts are unusually concrete for a launch post. Stripe reports Fable 5 compressed a 50-million-line Ruby codebase migration from roughly two months of manual effort into a single day. It is the first Claude model to complete Pokémon FireRed through a vision-only interface with no helper tools, a long-horizon autonomy test that previous models needed scaffolding to survive. Hex reports it is the first model to break 90% on their core analytics benchmark, a 10-point jump over Opus. Cursor's CEO says it opened up a class of long-horizon problems that were previously out of reach. On the restricted side, Mythos 5 accelerated protein design tasks around ten times in early scientific deployments, with nine of 14 protein targets yielding strong drug candidates, and molecular biologists preferred its hypotheses roughly 80% of the time over Opus-class models in blind comparisons. GitHub shipped Claude Fable 5 support in Copilot the same day, citing autonomy and reliability that exceeded previous benchmarks.
Step back and the access model is the part builders should sit with: capability is now explicitly stratified. Everyone gets Fable 5, the Mythos class with classifiers as chaperone. Project Glasswing partners, the cybersecurity and infrastructure tier, get Mythos 5 with the cyber safeguards lifted. A trusted-access program for vetted biology researchers comes next, biology safeguards lifted, cyber ones kept, with broader expansion promised steadily. The frontier is no longer a flat product you buy, it is a graduated privilege you qualify for, and every lab watching this launch now has a template for shipping models previously considered too capable to release. Practically: Fable 5 is included in Pro, Max, Team, and seat-based Enterprise plans at no extra cost through June 22, after which it needs usage credits, so the free evaluation window is open right now, and at $10/$50 it sits at a price point where routing your hardest tasks to it, and only those, makes obvious sense. One disclosure, because it would be strange not to: this article was researched and written by Claude Fable 5 itself, hours after release, the newsroom you are reading runs on Claude and was switched onto the new model the same day.