Two months ago, Anthropic unveiled Claude Mythos Preview and immediately put limits on who could use it. Among those with limited access to the AI model via a programme called Project Glasswing were a small group of cybersecurity organisations and critical infrastructure providers.
The company argued that Mythos-class models had become powerful enough to create significant risks if misused, particularly in areas such as cybersecurity and advanced biological research. But it did make a promise to release models with Mythos-level capabilities more broadly if it could develop safeguards strong enough to prevent abuse.
Now, it seems the Claude creator may have found a workaround because on Tuesday, Anthropic launched Claude Fable 5, a new model it describes as Mythos-class and its most capable generally available AI system to date.
Anthropic says Fable 5 outperforms previous Claude models across software engineering, knowledge work, vision, scientific research, and long-running tasks. But unlike a typical AI launch, Anthropic is openly acknowledging the trade-offs involved. "Releasing a model this capable comes with risks," the company said in its announcement.
The catch is that Fable 5 comes with new safeguards
While Claude Fable 5 and Claude Mythos 5 are the same underlying model, Fable automatically falls back to Claude Opus 4.8 when users enter certain cybersecurity, biology, chemistry, or model-distillation requests. Anthropic says the safeguards are designed to prevent misuse while still allowing broader access to Mythos-level capabilities.
Perhaps even more interesting is that the restrictions are relatively limited. According to Anthropic, more than 95% of Fable 5 sessions never trigger a fallback, meaning most users interact with the model without encountering the safeguards.
Speaking to CNBC, Anthropic's head of product management for research, Dianne Penn, said the company sees the launch as part of a broader "race to the top," balancing more capable AI systems with stronger safety guardrails so the technology delivers "asymmetrically more benefits than harm."
Claude Mythos 5 is still not available to the general public
But the company is still keeping Claude Mythos 5 restricted to Project Glasswing partners and plans to gradually expand access through trusted-access programmes for cybersecurity organisations and life-science researchers.
Anthropic says the model represents a major capability jump. During testing, Stripe used Fable 5 to complete a migration across a 50-million-line Ruby codebase in a single day, a task that would otherwise have taken a team of engineers more than two months to complete manually.
The launch comes days after Anthropic confidentially filed for an IPO and as competition among leading AI companies like OpenAI continues to intensify.
Both Claude Fable 5 and Claude Mythos 5 cost $10 per million input tokens and $50 per million output tokens.
More than a new model launch, Fable 5 is Anthropic's attempt to answer a bigger question facing the AI industry: how do you make powerful systems widely available without opening the door to misuse? For now, the company's answer is to release the model, but keep some of its most sensitive capabilities on a shorter leash.
Damilare Odedina
