Anthropic launched Claude Fable 5 and Claude Mythos 5 on June 9, 2026: two products that, according to Anthropic, are based on the same underlying model and differ only in their safeguards. Fable 5 is intended for general use with active safety classifiers, while Mythos 5 is reserved for a restricted circle of cyber defenders via Project Glasswing, with those same safeguards lifted in certain areas. The containment layer, which Anthropic describes in its Anthropic engineering post as a configuration overlay applied at inference time, distinct from the weights, thus becomes the commercial boundary between the two offerings.
The novelty, then, is not simply the launch of a more powerful model. It lies in the explicit separation between model weights, the configuration applied at inference, and the access regime. Fable 5 and Mythos 5 become two distinct products not necessarily because they represent two separately trained models, but because they do not expose the same usage space. This distinction is strategic: it allows Anthropic to make so-called “Mythos-class” capabilities available while reserving certain uses for verified actors, notably within the framework of Project Glasswing.
According to Anthropic’s announcement, Fable 5 is priced at $10/M tokens input and $50/M output via the Anthropic API; according to LLMReference (updated on June 9, 2026), the Mythos Preview price was $25/M tokens input.
A price cut that fits into a sector-wide compression of inference costs
DeepSeek has been offering a pricing structure significantly below that of American frontier models for more than a year, while OpenAI’s GPT-5 is listed at $1.25/M input and $10/M output, or 8 and 5 times cheaper than Fable 5 on base pricing, respectively. The Mythos Preview price was $25/M input according to LLMReference, or 2.5 times the input price of Fable 5. The explanation based on safeguard configuration—Anthropic charges for the containment work rather than for the weights—therefore coexists with a competitive explanation; the June 9 announcement does not settle the issue, and the official wording remains silent on the breakdown.
Documented real-world performance, but sensitive results left out of the review
On the FrontierCode benchmark published by Cognition AI on June 8, 2026, Fable 5 rose to the top of the ranking according to Cognition AI, although the exact score was not included in the launch release; according to Cognition AI, Claude Opus 4.8 had led FrontierCode with a peak of 13.5% before Fable 5 arrived, and more than 50% of SWE-bench outputs remained non-mergeable. Cognition develops Devin and therefore operates in the coding-agent segment where Anthropic is positioning itself, something the launch release does not mention. On the customer side, Stripe reported during early testing that a migration on a 50-million-line Ruby codebase was completed by Fable 5 in one day, compared with more than two months estimated for an entire team. In the sensitive domain, Anthropic says that 9 of 14 protein targets in an internal drug-design study produced strong candidates, but these results have not been peer-reviewed. The company also specifies that its automated alignment evaluation places Mythos 5 at a level of misaligned behavior similar to that of Opus 4.8, meaning low but not zero.
Glasswing, classifiers, and the AI Act timeline: what European buyers read in the announcement
The most cited result Anthropic uses to support the Mythos class predates the launch of Mythos 5: according to the Glasswing expansion announcement of June 2, 2026, the initial 50 partners—including AWS, Apple, Cisco, CrowdStrike, Google, Microsoft, NVIDIA, JPMorganChase, and the Linux Foundation—had identified more than 10,000 high- or critical-severity security flaws by scanning their codebases via Mythos Preview. These flaws are therefore attributable to the Preview access opened in April 2026, not to Mythos 5, whose Glasswing deployment is only just beginning. On the Fable 5 side, Anthropic says the safety classifiers trigger in fewer than 5% of sessions on average and then redirect the request to Opus 4.8; the company acknowledges a deliberately conservative setting capable of intercepting harmless requests. For a European buyer, the regulatory timeline eases integration pressure: according to the provisional political agreement reached on May 7, 2026 on the AI Act Omnibus, the high-risk obligations in Annex III are postponed to December 2, 2027. The safeguard configuration, as billed by Anthropic, becomes an object sold independently of the model; the classifiers’ false-positive rate—currently reported as below 5% of sessions, measured by Anthropic alone—is not published by the severity level of the intercepted request.