Anthropic on Tuesday released Claude Fable 5, a public, guardrailed version of its vaunted “Mythos” AI model.
The AI startup said Claude Fable 5 is more capable than any other model it has released to the public, showing outstanding performance in software engineering, knowledge work, and vision tasks, among other areas.
The model is superior to earlier iterations in performing longer, more complex tasks. The analytics company Hex said that Fable 5 was the first model to break 90% on its core benchmark of complex, long-running analytical tasks–a 10-point jump over earlier Opus models. Claude Fable 5 scored 80.3% on the SWE-Bench-Pro, which measures AI models’ agentic coding skill, compared to (OpenAI) GPT-5.5’s score of 58.6%, and (Google Deepmind) Gemini 3.1 Pro’s score of 54.2%. The testers said Fable showed “strong judgment and attention to nuance.”
Anthropic released the original Mythos Preview model in April, but only to a select set of defensive cybersecurity pros and overseers of critical infrastructure. Mythos, the company said, had proved to be very good at finding and exploiting security weaknesses in commercial software.
The new Fable 5 model comes with a few limitations, Anthropic says. It won’t give responses to questions in the areas of cybersecurity and biology, for example, instead routing such questions to the less-capable Claude Opus 4.8 model. This happened in only about five percent of sessions, however, Anthropic says.
Fable 5 was subjected to attempts by both internal and external teams to trick it into answering questions about prohibited subject areas. None of these attempts were successful.
As for the cybersecurity researchers who already have access to Mythos Preview, they will now get access to a new model called Claude Mythos 5. More people and organizations will get access through “a more systematic trusted-access program,” Anthropic says.
This public release of a Mythos model comes just days after Anthropic urged major AI laboratories to implement a coordinated global pause on development of AI, warning of growing risk that rapid advancements in AI capabilities could outrun human control mechanisms.
Anthropic’s rival OpenAI doesn’t disagree. In a new blog post Monday, CEO Sam Altman (with Jakub Pachocki) wrote that the industry needs an international organization that can “make it possible for the world to take coordinated action, including slowing frontier development when needed, so societal resilience, safety, and alignment can keep pace.”
AI company leaders have called for more aggressive government regulation on frontier model development, but the AI industry lobby and the Trump Administration have worked hard to derail binding AI safety laws that require more than voluntary compliance from AI labs.
However, the potential security threat of Mythos reportedly fueled disagreement among key White House staffers about the strength of the government’s oversight role. It also helped trigger the recent AI executive order, which asks AI labs to give the government a month of advance testing before they release certain new frontier models.
Claude Fable 5 is available to developers via Anthropic’s API.
