Two months after first warning that it had built an AI model too powerful for a broad roll-out, Anthropic on Monday, June 9, publicly released Claude Fable 5, its latest AI model with Mythos-level capabilities.
The AI startup also unveiled another AI model called Claude Mythos 5, which it says has greater capabilities than the Mythos Preview model released in April this year.
Anthropic is currently only releasing Claude Mythos 5 to a limited set of industry partners, most of whom are already part of Project Glasswing, the company’s joint industry initiative to find and fix critical software vulnerabilities using AI. Select biology researchers will also be given access to Claude Mythos 5. The company further said it is collaborating with the US government on the rollout of the AI model.
Both Claude Fable 5 and Claude Mythos 5, reportedly use the same underlying model. The key difference between the two AI models is that Claude Fable 5, which is being publicly released, has guardrails in place that will block the model from answering many user questions related to cybersecurity, biology, and chemistry.
Anthropic had initially moved to limit the release of Claude Mythos over concerns that the model’s capabilities could be exploited by bad actors to develop hacking tools that could catch defenders off guard. The company’s claims of Mythos’ ability to find and exploit hidden flaws in software that are used by banks, power grids, and other critical infrastructure set off alarm bells globally, with governments and organisations around the world scrambling to prepare for how attackers might eventually use AI to carry out cyber attacks.
Now, with the public release of a Mythos-class AI model, Anthropic’s idea of giving researchers and policymakers a headstart to prepare for advanced AI systems, will be put to the test. The two month-interval also underscores Anthropic’s business tension of wanting to release a Mythos-class AI model for general use before the tech industry has resolved the cybersecurity concerns of these models.
Anthropic rival OpenAI also privately launched a model with advanced cybersecurity capabilities in April. Both AI giants have also confidentially filed for IPOs and are looking to go public later this year.
“We’re working as quickly as we can to safely release Mythos-level capabilities in general access. To do so, we’ll need highly robust safeguards that prevent the model’s cyber capabilities from being misused—safeguards that we (and, to our knowledge, all other AI developers) have yet to develop,” Anthropic wrote in an update about Project Glasswing last week.
“We’re trying to make improvements in a way that’s beneficial, even if we don’t have the perfect [solution] for every use case to start. Out of all the different approaches, this emerged as the most viable and the best one. We just ended up feeling like this was the best product choice for users to get the maximum value out of Fable 5,” Anthropic’s head of product management, Diane Penn, was quoted as saying by Wired. Penn added that testing and user input has helped hone its strategy.
Guardrails for Claude Fable 5
Anthropic’s strategy of making Claude Fable 5 safe for widespread use involves its older AI technology. When Claude users input queries related to topics that could be perceived as too risky, Anthropic said that these requests will instead be rerouted to an older AI model, Claude Opus 4.8, which was released last month and is designed to avoid the security risks of Mythos.
These guardrails have been designed to block responses related to cybersecurity, biology and other vulnerable areas. Users will be informed whenever this occurs. While they are meant to avoid hackers using Fable to attack computer networks, these measures may also make businesses and cybersecurity experts struggle to defend networks using the new model.
Notably, Anthropic said that if a user is suspected of trying to conduct distillation, the technique of training a smaller AI model off a larger AI model’s responses, on Claude Fable 5, those requests will also be rerouted to Claude Opus 4.8 without informing the user about it.
The protective mechanism is built to err on the side of caution, which means that some user queries may be routed to the less capable AI model even if they are benign, according to Anthropic’s Diane Penn. In the past, Anthropic has expressed concerns that competitors, including open-weight model developers, could also eventually build AI models with Mythos-level capabilities.
The company said that it carried out more than 1,000 hours of red-teaming, a process where researchers attempt to breach AI products as though they are bad actors. The testers found no universal jailbreaks for the model. However, it remains to be seen how resistant Claude Fable 5’s safeguards are in the wild. Several research studies in the past have demonstrated that guardrails fitted on AI systems are not always reliable.
Claude Fable 5 also comes with major changes to Anthropic’s data retention policy. As part of its revised policy, Anthropic said it will retain all traffic on Mythos-class models (including prompts submitted to, and outputs generated by the models) for a period of 30 days.
The company said that it is retaining this data to help defend against complex and novel attacks (including new jailbreaks and attacks that operate across many requests) as well as help identify and reduce false positives. This data will not be used to train new Claude AI models, or for any non-safety-related purpose, Anthropic added.
Benchmark performance
In line with the company’s naming of AI models after various literary forms such as its existing Haiku, Sonnet, and Opus models, Claude Fable 5 is said to be adept at generating code and solving complex mathematical problems as well as tasks that require visual understanding.
It beat all other publicly available AI models in overall performance based on benchmark evaluations carried out by Vals AI, a company that tracks the performance of the latest AI models. Fable 5 also recorded a 5 per cent improvement in overall performance over Claude Opus 4.8.
Fable 5 and Mythos 5 can work autonomously for longer than any previous Claude models, according to Anthropic. Fable 5 is also more token-efficient than previous Claude models, scoring the highest among other frontier models on Cognition’s FrontierCode evaluation, which involves passing difficult coding tasks while meeting the standards of high-quality production codebases.
“You can give it a lot more ambitious tasks than what you’re used to, the model “gets it” and it will just go, and it’s never felt this tempting to stop looking at the code at all (but don’t do this in prod!). The model still has quirks that people will run into and the safeguards are configured to be a little too trigger happy for launch, which can hopefully be tuned over time,” Andrej Karpathy, famed AI researcher, OpenAI co-founder and member of Anthropic’s technical staff, said in a post on X.
However, the model is twice as expensive as Opus 4.8, costing developers $10 per million input tokens while Claude Mythos 5 is expected to cost developers $50 per million output tokens. Claude Fable 5 was also outperformed by other models when evaluated on tasks involving healthcare and tax evaluation, as per The New York Times. While Claude Mythos 5 is restricted to Glasswing partners, Anthropic has hinted that it might expand its trusted access programme.
On the other hand, Fable 5 is available for use by paid subscribers of Claude Pro, Max, Team, and seat-based Enterprise plans at no extra cost. Developers can also use the new model via the Claude API. Anticipating strong demand for Fable 5, Anthropic said that using it after June 22 will require usage credits with plans to make it a standard offering after building sufficient capacity.
Dario Amodei, chief executive and co-founder of Anthropic, speaks during a conversation with Andrew Ross Sorkin at the DealBook Summit in New York, Dec. 3, 2025. (Karsten Moran/The New York Times) 
Share
