Anthropic has cleared its top-tier Mythos technology for general release through Claude Fable 5, betting that fresh safeguards can contain a capability it previously deemed too dangerous to share.
Subscribers on Pro, Max, Team and Enterprise plans get Fable 5 free through 22 June, after which Anthropic will pull it and require usage credits until capacity expands, while business users face a mandatory 30-day data retention requirement that the firm says supports safety monitoring rather than model training.
The pitch, stripped of the marketing, seems straightforward enough: a model powerful enough to worry its own makers, now available to almost anyone, provided it never notices you asking the wrong question.
Anthropic originally launched Mythos as a preview restricted to a handful of partners, citing concerns about the model's potential to do damage in the wrong hands, a caution rooted in the system's unusual talent for finding and exploiting software flaws.
What changed is not the underlying technology but the wrapping around it, because Anthropic said Fable 5's broad release became possible thanks to new safeguards that block responses in specific high-risk areas, including cybersecurity and biology.
When the model detects a sensitive query, it quietly diverts the request to Claude Opus 4.8, the company's next-most-capable system, rather than answering itself.
Anthropic describes Fable 5 as state-of-the-art on nearly all tested measures, with its lead widening on the longest and most complex tasks.
During early testing, Stripe reported that the model compressed months of engineering into days, performing a codebase-wide migration of a 50-million-line Ruby codebase in a single day.
The company is candid about why a model this useful needed restraining first.
"Releasing a model this capable comes with risks," Anthropic said.
"Without safeguards, Fable 5's capabilities in areas like cybersecurity could be misused to cause serious damage.
"We've therefore launched the model with safeguards that mean queries on some topics will instead receive a response from our next-most-capable model, Claude Opus 4.8.
"To release the model both safely and quickly, we've tuned these safeguards conservatively - they'll sometimes catch harmless requests, though they trigger, on average, in less than 5% of sessions."
Three areas draw the automatic fallback to Opus 4.8 - cybersecurity, biology and chemistry, and attempts to distil the model's abilities to train rival systems - while the unrestricted version, Claude Mythos 5, stays confined to Project Glasswing partners working alongside the U.S. government.
The launch also sits awkwardly against the company's own messaging, arriving days after the Dario Amodei-led firm urged major global AI labs to agree a coordinated brake on frontier development, warning that systems are advancing fast enough to approach recursive self-improvement.
Commercial terms reward scale while hedging on capacity, with both models priced at US$10 per million input tokens and US$50 per million output tokens, less than half the cost of the earlier Mythos Preview.
