If you are trying to work out what Anthropic actually launched with Claude Fable 5 and Claude Mythos 5, the short answer is this: one new model family, split into two products. The names suggest a clean fork. The details do not. Anthropic says Fable 5 and Mythos 5 are the same underlying model, but they ship with different safeguard settings and different access rules.
That distinction matters because the June 9, 2026 release is as much about containment as capability. Fable 5 is the version Anthropic says is safe for broader use. Mythos 5 is the restricted-access version for a small set of vetted partners working in cybersecurity and biology. If you only read the headlines, you could miss the real point of the launch.
The big story is not two new brains. It is one model, carefully fenced into two very different products.
What Claude Fable 5 and Claude Mythos 5 actually are
Anthropic describes Fable 5 as the general-release product and Mythos 5 as the tightly controlled one. That does not mean Mythos is a separate foundation model. According to the company, both sit on the same base system. The difference is how that system is exposed, monitored and constrained.
That framing is unusually direct. AI companies often present new releases as clean product upgrades. Here, Anthropic is spelling out that access policy is part of the product itself. In plain terms, Claude Fable 5 is the version built for wider adoption, while Claude Mythos 5 is reserved for a narrow slice of users whose work involves higher-risk domains.
Why Anthropic split the release in two
The split reflects a straightforward tension. Anthropic says the new model family is stronger in coding, knowledge work, vision and computer use, and that Fable 5 can sustain long-running agentic work for days with planning, delegation and self-checking. The launch article also says it can outperform earlier Claude models across software engineering, memory, scientific research and long-horizon tasks.
Those are serious capability claims, but Anthropic pairs them with equally serious limits. When Fable 5 or Mythos 5 flags a cybersecurity or biology request, the query is routed to Opus 4.8 instead of the newer model. That fallback is one of the most revealing details in the whole launch. It shows Anthropic is willing to ship the stronger system while still keeping hard boundaries around where it will and will not trust it.
If a request crosses the wrong line, the newer model steps aside and Opus 4.8 takes over.
How the safeguards work, and where they may annoy users
Anthropic says its safeguards are conservative enough that some benign requests may be flagged. That is worth noting because it cuts against the usual product-marketing instinct to promise seamless behaviour. The company also says these classifiers trigger on average in less than 5 per cent of sessions, which suggests most users will never notice them, but the ones who do may well hit frustrating false positives.
The research behind that safety layer matters here. Anthropic’s next-generation constitutional classifiers use a two-stage detection approach and are presented as a lower-cost, lower-refusal successor to the first generation. In other words, the company is trying to reduce both expense and collateral damage while still screening risky queries. Whether that balance holds up in practice is a separate question, but the design goal is clear enough.
Pricing, retention and the trade-off hiding in the small print
Both products are priced the same: $10 per million input tokens and $50 per million output tokens. That makes the access split a governance question, not a pricing one. You are not paying extra for Mythos 5 because it is more expensive to meter in public. You need permission because Anthropic does not want broad access to that configuration.
The other detail worth pausing on is retention. Anthropic requires 30-day data retention for safety monitoring on both Fable and Mythos access. For some teams, that will be a non-issue. For others, especially those dealing with sensitive internal work, it will be a real procurement consideration rather than a footnote.
The sharpest commercial catch is not the token price. It is the mandatory 30-day retention window.
Where Project Glasswing fits into the launch
Mythos 5 did not appear out of nowhere. Anthropic says Project Glasswing is its defensive cybersecurity initiative with major industry partners, and that Mythos Preview was used there before the June 2026 release to help secure critical software. That gives Mythos a fairly specific origin story. It was not introduced as a mass-market chatbot with sharper teeth. It was tested in a defence-first setting before being offered, in restricted form, to vetted partners.
That context also helps explain why Anthropic says Fable 5 and Mythos 5 are aligned at a level similar to Opus 4.8 in its automated assessment. The company is trying to argue that improved capability does not automatically mean worse behaviour, even while it keeps the riskiest use cases under much tighter control.
What to take away from Claude Fable 5 and Claude Mythos 5
The cleanest reading of this launch is that Anthropic wants the upside of a stronger model without pretending every user should get the same version of it. Claude Fable 5 is the public-facing answer. Claude Mythos 5 is the guarded one. Same underlying model, different rules, same price, strict monitoring, and a visible safety backstop in Opus 4.8. That makes this release less flashy than some rivals, but more honest about the trade-offs.