Anthropic Releases New ‘Mythos-Class’ Model to General Public With Guardrails
Queries about dangerous topics such as cybersecurity or bioweapons will be steered to older Opus model
- Anthropic is releasing Claude Fable 5, a next-generation AI model based on Mythos, to the general public with safety guardrails.
- Fable 5 will revert to the older Opus 4.8 for sensitive queries related to cybersecurity or biological research, removing dangerous capabilities.
- The full Mythos model, on which Fable 5 is based, prompted the Trump administration to re-evaluate its AI policy.
Anthropic is releasing a next-generation “Mythos-class” model to the general public with guardrails that remove dangerous capabilities related to areas such as cybersecurity and biological research.
Called Claude Fable 5, the large language model will mostly let users query Mythos, which the company previously deemed was too dangerous for general release. However, if users ask Fable about sensitive issues such as a bioweapon or exploiting a software bug, it will kick them back to the older Opus 4.8 version of the Claude chatbot.
Fable 5 will cost more than Opus 4.8, but it will also do a better job of remembering things. That will make it better at completing large, complex tasks with fewer instructions, said Dianne Penn, Anthropic’s head of product management, research and labs.
“We wanted to be able to provide this level of intelligence for general users in a safe manner,” she said.
Hackers are likely to try to trick Mythos into answering these questions, despite Anthropic’s controls—a process known as jailbreaking—but the company says it has done extensive testing to make that harder to pull off.
Since it was announced in April, Mythos has caused the Trump administration to re-evaluate its artificial-intelligence policy and spooked cybersecurity experts, many of whom say that Mythos and other advanced AI models are now producing an avalanche of bug reports.
Anthropic has held off on sharing its full Mythos model with the world. Presently, only about 150 organizations have access to this technology, including technology companies such as Verizon and Microsoft. They are rushing to security-test their systems and products before hackers gain access to similar technologies.
Anthropic is also upgrading the model available to these early access partners, from Mythos Preview to Mythos 5. The company plans to broaden access to this model over time through “a more systematic trusted-access program,” the company said.
When priced per-token, Fable 5 will cost twice as much as Opus 4.8, but because it will use fewer tokens for certain types of tasks, it might be more cost efficient for certain types of jobs, Penn said.