Anthropic has shoved its Mythos-class AI into public view, but wrapped it in enough guardrails to annoy the grown-ups.
Claude Fable 5 is the public version of the same Mythos base model that Anthropic previously kept behind the velvet rope so that the bad guys could not get their paws on it.
The original Claude Mythos 5 is still limited to approved partners and government programmes working on cybersecurity and other critical infrastructure. For those users, Anthropic loosens some of the stricter controls so experts can run probe-and-defend systems without the model hiding under the desk.
For everyone else, Claude Fable 5 is the friendly public face of Mythos-class capability, which means power with a compulsory safety nanny telling you what you can’t do. Fable 5 blocks or reroutes sensitive material involving cybersecurity and biological threats.
Anthropic says this keeps most of Mythos’s useful tricks available for everyday work, such as coding, research, analysis and content creation.
The idea is to let punters use the clever bits without handing them a recipe book for digital or biological mischief. Fable 5 now sits above Claude Opus 4.8 in Anthropic’s line-up and becomes its most capable model for ordinary users.
Early benchmarks suggest it beats Opus 4.8 by a decent margin in coding, reasoning, scientific tasks and other brainy chores. Anthropic claims that Fable 5 can work on complex projects for days, making it useful for software engineering, finance, legal work, and long-running analysis.
That sort of claim will please the AI agent crowd, who keep promising that tomorrow’s software will quietly toil away while humans drink tea. Anthropic is pitching Claude Fable 5 as proof that high-end AI can reach regular users without exposing the sketchiest behaviour from the original Mythos release.
That is the corporate line, at least. A model powerful enough to need a public-safe costume still sounds like something regulators will want to poke with a stick.
However, it appears Anthropic has managed to annoy AI researchers by making Claude Fable 5 safer, dimmer and harder to trust. A wave of AI experts reacted with frustration after the company was accused of gatekeeping its shiny new model to kneecap possible rivals.
Critics said the setup limited outside researchers’ ability to assess and use Fable to its full potential. Anthropic responded by saying it would make those safeguard notifications visible.
“A hidden safeguard is harder to probe and work around,” the company said in a statement. “We made the wrong tradeoff and we apologise for not getting the balance right.”
Many users complained that Fable was blocking them from discussing apparently harmless topics such as mathematics, biology and chemistry. Some said it even blocked attempts to analyse Fable’s own publicly released system information.
One user posted a screenshot showing Fable refusing to answer a basic cellular anatomy question: “Tell me about mitochondria.”
Jackson Laboratory immunologist and cancer researcher Derya Unutmaz said he quickly loaded up Fable to try it out. Instead, Unutmaz found it refused to answer any query. He suspects that is because his chat history includes biological topics.
“I can’t even say the word ‘cancer’—or the word ‘hello. To me that’s a nightmare scenario, that one company can decide what questions we are allowed to ask and what is a security issue or not,” Unutmaz said.
Regarding scientific research, Anthropic said it was initially “necessary to be overly conservative” and “block most queries tied to biology work” to avoid malicious actors conducting “highly risky” research.
The company said it is now working to reduce unnecessary obstructions. Anthropic now plans to make Mythos-class models such as Fable “available without these safeguards to the broader biology and life sciences community so these capabilities can be used to accelerate biomedical research and drug discovery,” a spokesman said Wednesday.
