Claude Mythos 5 hits the top of every major benchmark, but only Glasswing partners and the US government can use it. Everyone else gets Fable 5, and even that is going credits only after June 22.
Anthropic's most capable model is now available, and the people who can actually use it are not the public. The company jointly launched Claude Fable 5 and Claude Mythos 5 on June 9, and the two products sit on the same underlying model. The difference is who is allowed to touch it. Mythos 5 goes to Project Glasswing partners: the US government, cyber defenders, and a small set of critical infrastructure providers. Fable 5 is the consumer-facing sibling, and it is the only version any ordinary paying user will ever see. Even that version is going away from flat-rate plans in two weeks.
The structural claim is the actual news. Anthropic shipped the highest-benchmarked model in its catalog and routed most paying users to a classifier-gated sibling. After June 22, that sibling becomes credits-only on the paid plans, and the real frontier model stays behind Glasswing.
The benchmarks are real, even when the framing is Anthropic's. The company says Mythos 5 has "the strongest cybersecurity capabilities of any model in the world." Stripe, the payments company, told Anthropic it compressed months of engineering into days on a 50-million-line Ruby migration using Fable 5. Cognition reported the highest score on its FrontierCode benchmark. Hebbia said Fable 5 topped its Finance Benchmark. IMC, a trading firm, said Fable 5 "aced" trading analysis. These are vendor-curated quotes from the Anthropic launch post, not independent benchmarks, and they should be read with that in mind. The capability story lines up with the benchmark story, and the capability story is also the one Anthropic wants told.
Fable 5 is the same base model, with a safety-classifier layer in front. That layer routes cyber, biology and chemistry, and distillation queries to Claude Opus 4.8 instead of serving them from Fable 5. Anthropic says the classifier fires in less than 5% of sessions on average. For most users on most tasks, Fable 5 will behave like the underlying model. The classifier exists to satisfy US government partners and a few large enterprise customers, not the average user. It is a structural feature, and the structural feature is also the gate.
Pricing is the second lever. Both models list at $10 per million input tokens and $50 per million output tokens, less than half what Claude Mythos Preview cost. The headline price is not the price most users will pay, though. The Pro and Max style plans, which until now have included flat-rate or generous bundled access, are being pulled from Fable 5 access after June 22. After that date, Fable 5 on those plans is credits-only and pay-as-you-go. The two-week window is the entire offer.
For a working developer or researcher, the decision point is mechanical. If a project depends on Fable 5's specific capability profile and the cost of credits-only access is acceptable, the project is fine. If a project depends on the bundled, flat-rate model of the past, the project needs to recalibrate. The free and lower tiers continue to work, with their existing limits. Fable 5 is not disappearing from Claude entirely. It is disappearing from the plans that made it cheap.
The scientific claims around Mythos 5 are the part the public cannot verify, because the public cannot run Mythos 5. Anthropic says its internal protein design experts accelerated drug design by roughly 10x, that nine of 14 protein targets yielded strong candidates, and that scientists preferred Mythos 5's molecular biology hypotheses about 80% of the time compared with the Opus class. One Mythos hypothesis on an E. coli protein was corroborated in an independent bioRxiv preprint, which is not the same thing as peer review. A Mythos 5 genomics model 100x smaller outperformed a recent model published in Science, according to Anthropic's launch post.
Those claims, if they hold up, are the most consequential part of the announcement. They suggest that the next tier of biological research is not arriving on consumer hardware through consumer products. It is arriving through partner relationships, with the safety-classifier layer stripped out for vetted users and bolted on for everyone else. The public version of the frontier is the version with the safety layer, and the public is, by design, not at the frontier.
The safety-classifier framing is also where the legitimate criticism lives. Anthropic describes the layer as a guardrail. Critics will describe it as a tiering mechanism dressed in safety language, and they are not entirely wrong. The classifier routes the queries that are most strategically interesting to enterprise and government customers to a model that is older, slower, and more expensive to run at scale. The classifier is also the difference between a public product and a partner product. Calling that a safety feature is not wrong. Calling that the only safety feature in the system would be a stretch.
Anthropic's automated alignment assessment, per the launch post, found Mythos 5's misaligned-behavior level to be low and similar to Opus 4.8. Fable 5 is expected to land in the same range. The company is not claiming Mythos 5 is unsafe. It is claiming Mythos 5 is different in kind from the version a normal user can access, and the difference is delivered through the classifier layer.
The pattern, taken together, is two-tier AI as a product surface. The frontier model is infrastructure for the US government, defense-adjacent cyber defenders, and a small set of Glasswing partners. The general-release version is a sibling with the sharpest edges filed down, served through a classifier. The public pays a price premium on the sibling, in the form of credits-only access, and gets a model that is, on most tasks, almost as good. On the tasks the classifier catches, the public gets Opus 4.8.
What to watch next. The June 22 cutoff is the first measurable test. If Pro and Max subscribers churn at a high rate when Fable 5 moves to credits-only, Anthropic will face a real pricing question. If usage holds, the two-tier model is the new template. The Glasswing trusted access program that Anthropic promises "later" in the launch post is the second test. If the broader trusted access tier arrives within a quarter, the gate widens. If it does not, the gate is the product.
For now, the working user has two weeks of affordable Fable 5 access on the plans that made it cheap. After that, the math is different. The frontier, in the strict sense, is somewhere they cannot follow.