YOU GUYS — Anthropic just dropped Claude Fable 5, their first "Mythos-class" model that's supposed to crush previous Opus frontiers. But here's the twist: Mythos is only available for a small group of vetted cyberdefenders through Project Glasswing after its monthlong preview phase ended today. The publicly accessible Fable uses the same base but redirects banned topics — cybersecurity, biology, chemistry — back to Claude Opus 4.8 with an in-app warning; Anthropic admits they've tuned it "stricter than ideal" and expects fewer than five percent false positive rejections of harmless queries. Their defense is serious: over 1000 hours of red-team testing through a bug bounty program, and no universal jailbreak was found for Fable. They specifically worried Mythos about "agentic hacking"—executing multi-part cyberattacks — after internal evaluations showed it's significantly more capable than previous models at such attacks, even if the UK AI Security Institute found Mythos Preview performed comparably to GPT-5.5 on Capture the Flag challenges.
And the guardrails get specific: a bioweapon classifier that previously only flagged bioweb queries now blankets all chemistry and biology questions in Fable because "even benign" phrasing could help dangerous research — they want zero ambiguity here. For those who actually need unrestricted access, they're expanding Project Glasswing with US government consultation AND launched a new trusted program for life sciences organizations to lift the bio/chem blocks while keeping cyber ones locked down. The price tag is where it gets real: API and Enterprise users pay $10 per million input tokens and $50 per million output — that's 67% to 100% higher than OpenAI's GPT-5.5, which is already at the high end of what anyone wants to pay for a frontier model! Existing subscribers keep Fable access only through June 22 before they need usage credits while Anthropic builds capacity to bring it back into standard plans — so if you want it, grab it now.
Source: https://arstechnica.com/ai/2026/06/anthropic-says-these-topics-are-too-dangerous-to-let-its-fable-5-model-talk-about/
And the guardrails get specific: a bioweapon classifier that previously only flagged bioweb queries now blankets all chemistry and biology questions in Fable because "even benign" phrasing could help dangerous research — they want zero ambiguity here. For those who actually need unrestricted access, they're expanding Project Glasswing with US government consultation AND launched a new trusted program for life sciences organizations to lift the bio/chem blocks while keeping cyber ones locked down. The price tag is where it gets real: API and Enterprise users pay $10 per million input tokens and $50 per million output — that's 67% to 100% higher than OpenAI's GPT-5.5, which is already at the high end of what anyone wants to pay for a frontier model! Existing subscribers keep Fable access only through June 22 before they need usage credits while Anthropic builds capacity to bring it back into standard plans — so if you want it, grab it now.
Source: https://arstechnica.com/ai/2026/06/anthropic-says-these-topics-are-too-dangerous-to-let-its-fable-5-model-talk-about/