Anthropic launches Claude Fable 5 public mythos AI model

Yesterday 09:11

By: Dakir Madiha

Anthropic launches Claude Fable 5 public mythos AI model

Anthropic has introduced Claude Fable 5, making a Mythos class artificial intelligence model available to the public for the first time. The release follows a period in which the company restricted its most advanced systems to a limited group of cybersecurity specialists. The move signals a controlled expansion of high capability AI access under tighter operational safeguards.

Claude Fable 5 is built on the same underlying architecture as Claude Mythos 5, a more restricted version used under an internal approval program. The key distinction lies in its safety framework. When internal systems detect prompts related to cybersecurity, chemistry, biology, or model distillation, the model automatically redirects queries to a lower capability fallback system. The company reports that more than 95 percent of interactions proceed without triggering these safeguards, indicating limited interference for general use cases.

Early testing suggests strong performance across software engineering, knowledge work, and visual reasoning tasks. Industry partners reported significant productivity gains in complex engineering environments. One example involved large scale codebase migration work completed in a fraction of the expected time, with millions of lines of code processed in rapid cycles. The model also demonstrated advanced capability in visual problem solving, including completing structured gameplay challenges that previously proved difficult for earlier systems. Benchmark results in financial reasoning placed it ahead of competing models in internal evaluations.

The release also introduces strict data retention rules. All traffic associated with Mythos class systems is stored for 30 days across both native and third party interfaces. The company states that retained data is not used for training and is deleted after the retention period in most cases. Officials acknowledged that the safeguards may produce occasional false positives, where harmless prompts trigger protective classification layers. External security testing has not identified universal jailbreak methods, although early research groups have reported partial progress in controlled evaluations.

Pricing for Claude Fable 5 and Claude Mythos 5 is set at 10 dollars per million input tokens and 50 dollars per million output tokens. Subscribers on higher tier plans receive temporary access without additional charges before transitioning to usage credits. A separate access track remains reserved for specialized scientific research in biology and chemistry, while broader availability of the most restricted model continues to expand through government coordinated programs.