After deeming its AI model Mythos too “powerful” to release, AI giant Anthropic has unexpectedly unleashed a model of the same class to the public, minus the powerful cyber capabilities.
Anthropic’s Claude Fable 5 model was released for general consumption this week – amid concerns the company’s generative AI could be used to rapidly develop exploits in the world’s most sensitive software.
With Mythos undergoing limited previews with select private sector and government organisations, Anthropic led its announcement by affirming Fable 5 was of the same, highly capable class.
“Fable 5’s capabilities exceed those of any model we’ve ever made generally available,” said Anthropic.
To ensure bad actors don’t run amok with the purportedly unprecedented AI model, Anthropic promised that Fable 5 had been “made safe for general use” through a series of “safeguards” that divert queries on some topics to its existing Claude Opus 4.8 model instead.
Anthropic said this would include queries about cybersecurity, as well as biology and chemistry.
“Releasing a model this capable comes with risks.
“Without safeguards, Fable 5’s capabilities in areas like cybersecurity could be misused to cause serious damage.”
What can it do?
Anthropic promoted Fable 5 as a powerful tool for software engineers – citing an example from payments company Stripe where the model performed a codebase-wide migration in one day where it otherwise would have taken a “whole team” more than two months.
The company further lauded Fable’s ability to perform complex, “vision-based” tasks.
Not only could Fable play through retro Pokémon titles using only raw videogame screenshots, but the company claimed it could also rebuild a web app’s source code from images alone.
Claude's Fable 5 had no problem playing Pokemon. Photo: YouTube
Memory and long-context were also improved from previous models, while complex analytical tasks such as chart and table interpretation were handled with leading results.
“The capabilities of models like Fable 5 and Mythos 5 have the potential to do profound good for the world,” Anthropic wrote.
Fable 5 has launched at double the cost of Opus 4.8: totalling US$10 per million input tokens, and US$50 per million output tokens.
Tokens, which effectively measure usage for AI models, can be consumed at a wildly different pace depending on use-case and approach, though AI consultancy Iternal estimates a high-use developer can burn between six and 40 million tokens a month.
Should the cybersecurity industry be worried?
Much of the discourse surrounding Fable has been focused on whether Anthropic’s safety assurances can be trusted — especially considering AI chatbots’ history of being tooled to behave outside their intended functionality.
Some have complained the safeguards hinder even the most innocuous of cybersecurity tasks, while others warn they won’t be enough to stop dedicated hackers.
“If you’re reassured that Anthropic has only shipped a ‘safe’ version of Mythos, don't be,” said Charles Guillemet, chief technology officer at digital security company Ledger.
“Large language models safeguards have repeatedly shown [they] don't survive contact with even the laziest adversary.
“Ask politely a few times, frame it as your son's science-fair project.
“The model will cheerfully produce an exploit to break into a hospital network.”
During early testing under his security firm Dvuln’s evaluation framework, Jamieson O’Reilly said attempts to have Fable conduct its ‘Mythos-class’ exploitation capabilities were successfully resisted.
“On real, multi-step, authorised exploitation work, Fable 5 hands every turn over to Claude Opus 4.8 instead of doing the tasks itself, exactly as Anthropic's announcement says it should,” O’Reilly told Information Age.
“The day-one cyber safeguard holds under sustained, multi-turn agentic pressure, which is the harder case beyond the single-turn 'plan a cyberattack' prompts most tests rely on.
“But to be clear, this is [an] early signal, not a scientific result.”
It may be ‘safe’, but its data practices have scared Microsoft
Shortly after Fable’s release, sources told The Verge Microsoft was already limiting its use of the new model for employees because of Anthropic’s newly introduced data retention requirements.
Though Anthropic lauded Fable’s new memory capabilities, Information Age understands the new model is not subject to the ‘zero data retention’ rules which its other Claude models have operated under.
Anthropic explained its new safety measures meant the company would retain prompts and outputs for Mythos-class models over a 30-day period before deletion, while prompts that violate the company’s usage policy may be stored for up to two years.
Citing customer data and confidentiality concerns, Microsoft reportedly told employees its legal teams are evaluating the data retention requirements.
Information Age has asked Microsoft whether it expects to clear the new model for internal use but did not receive a response prior to publication.
The release of Fable 5 comes after Anthropic and rival AI giant OpenAI filed IPO plans this month, joining Elon Musk’s SpaceX in a race to the public stock market.
