Anthropic announced the release of Claude Fable 5 on Tuesday, unveiling the first Mythos‑family model that is not restricted to a narrow set of cybersecurity experts. The company described the new system as its "most capable model yet," emphasizing gains in software‑engineering assistance, knowledge‑work productivity and visual‑recognition abilities. In internal testing, Fable 5 even defeated the classic Pokémon FireRed game using a vision‑only configuration, a feat earlier models could not achieve.
Safety was a central theme of the launch. Anthropic reported that Fable 5 rejected every single‑turn request that attempted to plan a cyberattack, develop exploits or evade defenses, even when users employed any of 30 publicly known jailbreak techniques. Those protections differentiate it from the more permissive Claude Mythos 5, which remains limited to trusted cybersecurity and software professionals through the Project Glasswing program. To further curb misuse, the model currently redirects most biology and chemistry queries to Anthropic’s Opus 4.8, a precaution driven by concerns over AI‑enabled bioterror.
The rollout follows a cautious, staged approach. Fable 5 is available now to Claude subscribers on Pro, Max, Team and enterprise plans, but only until June 22. After that date, the model will be removed from the standard subscription menu. Existing users can continue accessing Fable 5 by consuming special usage credits, which act as a pay‑as‑you‑go buffer once the allotted quota is reached. Enterprises that already include credits in their contracts may see little disruption, while others could incur additional fees.
Pricing reflects the model’s higher compute footprint. The Claude API now charges $10 per million input tokens and $50 per million output tokens for Fable 5, roughly double the cost of Anthropic’s Opus line. A pop‑up in the Claude interface warns users that the new model consumes twice the usage compared with Opus, signaling capacity constraints that prompted the limited release window.
Anthropic also introduced a new data‑retention policy for business customers. Subscribers must agree to a 30‑day retention period, during which usage data will be employed solely to defend against future cyberattacks and AI misuse—not to train future models. The company said it is consulting with the U.S. government on the broader release of Claude Mythos 5, aligning with recent White House calls for AI firms to submit new models for review before deployment.
Developers eager to experiment can integrate Fable 5 via the Claude API immediately, though the higher token costs and compute demands may limit large‑scale adoption. Anthropic expects strong demand and has signaled that the initial two‑week window is designed to balance user interest with system stability.
Overall, Claude Fable 5 represents a significant step for Anthropic, marrying advanced capabilities with robust safety measures. While the model’s price and limited availability may temper enthusiasm among cost‑sensitive users, its performance gains and strict guardrails position it as a compelling option for enterprises seeking powerful, responsibly managed AI tools.
Cet article a été rédigé avec l'assistance de l'IA.
News Factory APP - actualités agentiques pour booster votre SEO et AEO.