Anthropic unveiled Claude Fable 5 and Mythos 5, its long-awaited “Mythos-class” AI models.
The company said the models include safeguards that limit some frontier AI research tasks.
Critics said the restrictions blur the line between AI safety measures and competitive strategy.
Anthropic’s latest AI release is igniting controversy over what its new models will and won’t do.
On Tuesday, Anthropic unveiled Claude Fable 5 and Mythos 5, its long-awaited “Mythos-class” models.
Alongside the launch, the company disclosed two unusual safeguards: the models may secretly provide degraded assistance when they suspect users are working on frontier AI research, and certain requests are automatically routed to less capable models.
In a system card, Anthropic said the measures are designed to reduce the risk that powerful AI systems help users develop competing frontier models or accelerate dangerous capabilities.
Critics, however, said Anthropic’s safeguards could disadvantage researchers, concentrate power among leading AI labs, and degrade responses without users’ knowledge.
Here’s what smart voices across AI, cybersecurity, and policy are saying about Anthropic’s latest move:
“I do very much take Anthropic at their word that they are trying hard to de-risk what they see as the two risky features of Mythos,” Kasten told Business Insider.
Still, Kasten said releasing the model carries risk because “it’s always a bit of a cat-and-mouse game between attacker and defender.”
Kasten added that Anthropic likely sees itself in a race with rival AI labs that may soon have similarly powerful models.
“For them, I think there’s a little bit of a, how much of their lead in the race do they think they can afford to burn?” he said.
Davi Ottenheimer, VP of Trust and Digital Ethics at Inrupt
Davi Ottenheimer, a prominent digital safety expert, questioned whether Mythos was ever as dangerous as Anthropic suggested.
“Anthropic marketing said in April that Mythos was too dangerous to be in public hands,” Ottenheimer told Business Insider. “And yet today they are selling it, unchanged, to the public.”
Mythos was withheld from public release in April after Anthropic warned that the model was so capable at finding cybersecurity flaws that it posed safety risks.
“They’re using security as a marketing trick,” Ottenheimer added.
Roman Stanek, founder of GoodData
Roman Stanek said that AI capability isn’t the real problem in cybersecurity.
“The thing with AI and cybersecurity is that the vulnerabilities we’re worried about AI exploiting, well we’ve known about most of them for 20 years,” Stanek said in a note sent to Business Insider. “We just never fixed them.”
“Nobody wanted to pay a human engineer to fix it,” he added. “They’re not going to pay an AI to fix it either.”
Elie Bakouch, research engineer at Prime Intellect
Elie Bakouch criticized Anthropic’s decision to intentionally limit Mythos’ performance on certain AI development tasks.
“Mythos will be bad ON PURPOSE on AI ‘frontier LLM research’ tasks,” Bakouch wrote on X on Tuesday. “This is very very sad for the research community.”
Bakouch also called it “crazy” that the intervention would not be visible to users.
Jeremy Howard, cofounder of AnswerDotAI
Jeremy Howard said that Anthropic’s safeguards could increase concentration in the AI industry.
“Anthropic has chosen the opposite of the safe path,” Howard wrote in an X post on Wednesday. “They are allowing themselves, the current top lab, to use their top model for frontier AI research.”
Howard said the result is that “the AI frontier advances, & power imbalance increases.”
Patrick Moorhead, founder and CEO of Moor Insights & Strategy
Patrick Moorhead.
Errich Petersen/SXSW Conference & Festivals via Getty Images
Patrick Moorhead said his first experience with Fable 5 left him underwhelmed.
In a post on X on Wednesday, Moorhead said the model refused to help with earnings-analysis talking points and a board presentation because Fable 5 deemed the tasks “too dangerous.”
“Is this the model we’re all frightened of and makes Anthropic worth $1T?” he wrote. “Okedokee.”
Gergely Orosz, author of ‘The Pragmatic Engineer’ newsletter
Gergely Orosz said Anthropic’s safeguards could end up affecting people who aren’t actually building competing AI models.
“Anthropic assumes SemiAnalysis is developing a competing LLM and so it dumbs down their model for them,” Orosz wrote on X on Wednesday.
“Anthropic trying to limit competition limits many others,” he added.
Deedy Das, partner at Menlo Ventures
Deedy Das focused on the model’s capabilities rather than its safeguards.
“Claude Fable 5 is by far the most ridiculous model that makes me genuinely afraid for the future of software engineering,” Das wrote in an X post on Wednesday.
Das pointed to examples including migrating a 50 million-line codebase, generating advanced 3D graphics, and outperforming rival models on optimization tasks.
He also said that Fable 5 is about the same price as OpenAI’s GPT-5.5 and six times cheaper than GPT-5.5 Pro.
