Anthropic is making its first Mythos-tier model available to the general public. On Tuesday, the AI lab announced it was rolling out Fable 5—the first widespread release of an AI model from a significantly more powerful class of model—as a general release, and Claude Mythos 5 to vetted partners who already have access to the Claude Mythos Preview.
It’s a considerable step for the lab, which initially deemed Mythos-class models too dangerous to release owing to their increased ability to find cybersecurity weaknesses. It also comes just over a week after the company confidentially filed for IPO paperwork.
Fortune first reported the existence of the new and more powerful model in March after a data leak revealed its existence. In April, Anthropic officially announced Mythos, calling it a “step change” in capabilities. It opted to tightly control its rollout through an initiative called Project Glasswing. The project was aimed at giving cybersecurity professionals access to the models’ cyber skills so they could strengthen their defenses against new threats posed by increasingly advanced AI models.
Now, Anthropic says it is confident new guardrails are enough to ensure these dangerous skills don’t fall into the wrong hands.
“The reason why we’re releasing Fable 5 now is very much due to us feeling more confident with our safety guardrails in place,” Dianne Na Penn, Anthropic’s head of product management, research and labs, told Fortune.
The company said that responses in specific high-risk areas, such as biology and cybersecurity, will be blocked and will, in most cases, be answered by Claude Opus 4.8, Anthropic’s less powerful model, released earlier this year. The fear is that giving users free access to Mythos-level intelligence could enable bad actors to carry out more advanced cyberattacks or develop bioweapons more easily. The company also said it had extensively red-teamed its classifiers (machine learning algorithms) to test their robustness against jailbreaks.
Anthropic says the new model capabilities “exceed those of every model we’ve previously made generally available,” and demonstrate exceptional performance in areas such as coding, knowledge work, and vision.
Penn said that Fable 5 is particularly strong at what is known as “long-horizon memory management”—an AI model’s ability to keep track of what it’s doing over a long, complex task.
Where previous models sometimes “lose the thread” partway through complex, long tasks, the new model is much better at remembering what it’s doing over long stretches, she said. Penn added that the lab has seen improvements in self‑verification and instruction following: The model is more deliberate about checking its own work, validating assumptions, and course‑correcting before producing a final answer. These changes mean higher overall quality across both coding and non‑coding tasks, she noted.
“We think there are some use cases where it’s a replacement [for knowledge work], whether it’s coding or others, but we also think that it’s really expanding the use cases,” Penn said. “Now you could use a model like Fable 5 to run very long-horizon projects, potentially overnight … and that might include things like being able to review your whole code base for improvements.
“We recommend that customers give their most challenging work to Fable 5,” she added.
The lab said that early data shows at least 95% of Fable sessions run entirely on Fable’s own responses, rather than being routed back to Opus. However, Anthropic has faced backlash from some users in the past, who claim that the company’s safety filters occasionally “over-block” benign requests or issue false refusals.
Penn acknowledged the guardrails might not be perfect the first time around.
“We recognize that there might be some benign requests that end up being blocked initially,” she said. “We’re working actively on making those safeguards improvements post-launch, but we wanted to make the model accessible generally in a safe manner as soon as we could.”
Fable 5 and Mythos 5 are being offered at $10 per million input tokens and $50 per million output tokens—twice the price of the standard version of the next most advanced model, Opus 4.8.
