Fable 5, Anthropic's Mythos Model for Everyone: Free Only Until June 22
Anthropic has made available to the public Fable 5, the first model from the Mythos-class family to reach general distribution. It is the same underlying model as Claude Mythos 5, announced simultaneously but reserved for a select circle: what distinguishes the two names are the safety barriers present in Fable and removed in certain areas in Mythos.
In the first official update of Project Glasswing, the company had anticipated its intention to bring the Mythos-class models to the public once adequate safety barriers were developed, and in the following weeks traces of the model had appeared in the Claude Code interface. In April, the position was the opposite: Mythos Preview, the first model from the family, had been presented as too dangerous for an open release and distributed only to partners of Project Glasswing, the controlled access program for defensive cybersecurity.
On a capabilities level, Anthropic places Fable 5 above Opus 4.8, so far its most advanced general-use model. The company claims state-of-the-art performance across nearly all tested benchmarks, with advantage growing as tasks become longer and more complex. Among the data reported by early access customers, Stripe indicated a migration on a Ruby codebase of 50 million lines completed in a day, compared to over two weeks estimated for a whole team. Fable 5 also successfully completed Pokémon FireRed with vision-only harnessing, where previous models struggled even with additional support tools.
How the New Barriers Work
The novelty that makes public release possible is the fallback mechanism. When Fable 5's classifiers intercept a request that falls within cybersecurity, biology, and chemistry, or attempts to distill the model, the response is not rejected but automatically handled by Claude Opus 4.8, a less capable but still robust model. The user is informed every time this happens. According to the company's initial data, the rerouting activates in less than 5% of sessions: in the remaining 95%, Fable 5 operates at its full capacities, effectively equivalent to those of Mythos 5.
Anthropic admits to having tuned the filters cautiously, to the point that they sometimes intercept harmless requests, and states the goal of reducing false positives after launch. On the robustness front, an external bug bounty of over 1,000 hours resulted in no universal jailbreak formulation that completely bypasses barriers; external red-teaming organizations have not found them on long-duration agentic tasks, although the UK AI Security Institute made partial progress towards one in an early test window. Meanwhile, the defensive logic of Glasswing has produced concrete results: in the first month of the program, about 50 partners identified over 10,000 high or critical vulnerabilities, as we reported a few weeks ago.
Pricing, Access, and Data
Fable 5 and Mythos 5 are offered at $10 per million input tokens and $50 per million output, less than half the price of Mythos Preview. For developers, the model is available immediately via API and on consumption Enterprise plans. On subscription plans (Pro, Max, Team, and Enterprise per workstation), Fable 5 is included at no additional cost until June 22; from the 23rd, its use will require usage credits, pending sufficient available capacity to reintegrate it as a standard component.
Mythos 5 remains confined to Project Glasswing partners, who from today can upgrade from Mythos Preview with the cyber barriers removed, and shortly to some selected biology researchers. The company intends to gradually expand access through a more systematic trusted access program, with periodic additions of new partners as we discussed in the ENISA case. Finally, data management changes: for Mythos-class models, a mandatory retention period of 30 days applies on all traffic, both first and third party, which Anthropic declares it does not use to train new models and which is aimed at defending against complex attacks and reducing false positives.