Anthropic opens its top Claude tier to the public, gated by a new filter

Anthropic released Claude Fable 5 on June 9, moving a tier of model it had previously withheld from the public into general release.

The company calls this the "Mythos class," a level it had until now kept behind closed doors over misuse concerns. It arrives wrapped in a new filter that quietly hands some requests to a weaker model, and on a usage meter that burns through plan limits at twice the rate of Opus. Anthropic also launched a second model, Mythos 5 — the same underlying system with some safeguards removed — but restricted it to a small set of cyberdefense and infrastructure partners.

Anthropic says Fable 5 is state-of-the-art on nearly all tested capability benchmarks, with its largest leads on the longest and most complex tasks. Those are the company's own numbers, and no independent benchmark has confirmed them.

Why it matters: this is the first time Anthropic has put a "Mythos-class" model in front of the general public, a capability level it had previously held back over misuse concerns. The launch pairs that step with a new, deliberately strict safety system and a pricing structure that signals how expensive frontier-scale use is becoming.

Both models are priced at $10 per million input tokens and $50 per million output tokens, which Anthropic says is less than half the cost of its earlier Mythos Preview model. Developers can call claude-fable-5 through the Claude API.

The rollout for subscribers is staged, and the timing matters for anyone planning to lean on it. Anthropic says Fable 5 is included on Pro, Max, Team, and seat-based Enterprise plans at no extra cost from June 9 through June 22. On June 23 it will be removed from those plans, and continued use will require usage credits until capacity allows the company to restore it as a standard feature. On the API and consumption-based Enterprise plans, it is fully available now.

The in-app launch notice makes the cost trade-off explicit: Fable 5 is "included in your plan limits until June 22," but it consumes usage at twice the rate of Opus. For subscribers, that means the model draws down plan limits faster even during the free window.

The launch also comes with a new set of "classifiers" — separate AI systems that screen requests. When they detect a query related to cybersecurity, biology and chemistry, or attempts to copy the model's capabilities, the response is handled by Claude Opus 4.8 instead, and the user is told. In the app this surfaces as a setting, on by default, that switches models when a message is flagged rather than refusing it. Anthropic says it tuned these safeguards conservatively, that they trigger in fewer than 5% of sessions on average, and that they will sometimes catch harmless requests while the company works to reduce false positives.

What Anthropic says: the company frames the dual launch as a way to release frontier capability "both safely and quickly." It says Fable 5's cybersecurity and biology capabilities could cause serious harm without safeguards, which is why the unrestricted Mythos 5 is limited to vetted partners through a US-government-linked effort it calls Project Glasswing. Anthropic also says an external bug bounty ran more than 1,000 hours without producing a universal jailbreak, while noting that the UK's AI Safety Institute "has made progress towards one" in early testing.

What others are seeing: Felix Rieseberg, who leads Claude Code and Cowork on the desktop at Anthropic, argued in a public thread that the bigger story is not the benchmarks but a shift "from giving AI tasks to giving it responsibilities." He described running Claude in a loop that watches every incoming crash report rather than fixing one bug at a time. That framing drew immediate pushback in the replies, where several users pressed the same question — who is accountable when an agent given a "responsibility" fails — and others argued the framing only works while heavy usage is subsidized. This is reaction and opinion from an Anthropic employee and the public, not independent verification.