Anthropic on Tuesday publicly released Claude Fable 5, the first model from its Mythos tier that anyone with a paid account can actually use. According to the company, Fable 5's capabilities exceed those of any model it has ever made generally available.
The release comes two months after Anthropic unveiled Claude Mythos Preview and promptly locked it away from the public, citing its alarming ability to find and chain zero-day exploits across every major operating system and browser. That model was restricted to a handful of partners through a gated cybersecurity initiative called Project Glasswing, which includes Amazon Web Services, Apple, Microsoft, Google, Cisco, and JPMorgan Chase.
Same Model, Different Guardrails
Fable 5 is built on the same underlying architecture as Mythos. The difference lies in a layer of safeguards that redirect queries in sensitive domains to Claude Opus 4.8, Anthropic's previous flagship model. Those domains include cybersecurity, biology, chemistry, and what the company calls model distillation, where bad actors attempt to extract Claude's capabilities to train competing systems.
Anthropic says it tuned these safeguards conservatively. The classifiers will occasionally catch harmless requests, though internal testing shows they trigger in fewer than 5% of sessions. The company has promised to reduce false positives as it refines the system post-launch.
To test whether users could bypass these restrictions, Anthropic ran both internal and external red team campaigns. An external bug bounty program produced no universal jailbreaks in over 1,000 hours of testing. External red-teaming organizations also failed to find universal bypasses. The company acknowledges that jailbreaking risks remain and expects motivated adversaries to try.
Performance and Pricing
Fable 5 leads Anthropic's benchmark charts across software engineering, knowledge work, vision, and scientific research. The company says the model's advantage grows on longer, more complex tasks. On some benchmarks, Fable 5 scored more than 10% higher than Claude Opus 4.8.
Third-party testing reinforces these claims. Analytics company Hex reported that Fable 5 was the first model to score above 90% on its core analytics benchmark. Stripe, during early access, said the model compressed months of engineering work into days by completing a large-scale codebase migration in a single day.
Both Fable 5 and Mythos 5 are priced at $10 per million input tokens and $50 per million output tokens. That's less than half the cost of Claude Mythos Preview, though still double the price of Opus 4.8. Enterprise token costs have become a recurring concern in the industry; Uber recently disclosed that it burned through its annual AI budget in four months.
Mythos 5 Stays Gated
Alongside the public Fable 5 release, Anthropic deployed Claude Mythos 5 to existing Project Glasswing partners. This version lifts the safeguards in certain areas and offers what the company describes as the strongest cybersecurity capabilities of any model in the world. Anthropic plans to expand access through a broader trusted-access program but has not disclosed a timeline.
A Strategic Launch
The timing is deliberate. Anthropic confidentially filed its IPO prospectus with the SEC last week. The company says its revenue run rate has reached $47 billion, up from roughly $10 billion in annual revenue last year. Its most recent funding round valued the company at $965 billion, surpassing OpenAI's $852 billion valuation from late March.
The release also follows Anthropic's public plea for major AI labs to establish coordinated mechanisms for slowing frontier development if needed. The company has warned that AI systems may soon achieve recursive self-improvement, autonomously enhancing themselves without human intervention.
For now, Fable 5 is available through the Claude API and included on Pro, Max, Team, and Enterprise plans until June 22. After that, usage credits may be required until Anthropic scales capacity. The company also introduced a mandatory 30-day data retention policy for all traffic on Mythos-class models, citing the need for safety monitoring. Anthropic says this data will not be used for training.
