nthropic has finally released the world's scariest language model, if you believe its warnings.
On Tuesday, the AI firm launched two "Mythos-class" models, Fable 5, which it claims is safe for general use, and Mythos 5, which has safeguards lifted in some areas. The company said that both models are state-of-the-art on nearly all tested AI benchmarks and can work autonomously for longer periods, exceeding the capabilities of any model Anthropic has ever made generally available.
Here are the details:
Fable 5 provides exceptional performance in domains such as software engineering, knowledge work, vision and scientific research, and is capable of handling longer and more complex tasks than previous iterations. However, given that this version of the model is available to the general public, queries on certain topics will be sent toClaude Opus 4.8, Anthropic's next most capable model, released at the end of May.Mythos 5, meanwhile, doesn't have such safeguards. This model is available only to a small group of cyberdefenders and infrastructure providers and will still be deployed throughProject Glasswing. Anthropic said that Mythos 5 has the strongest cybersecurity capabilities of any model in the world, and that it intends to expand access through a "broader trust program."- Anthropic is offering both models at $10 per million input tokens and $50 per million output tokens, half the cost of the Mythos preview. Until June 22, Fable 5 will be included on Claude Pro, Max, Team, and seat-based Enterprise plans at no extra cost. After that, using it after that will require usage credits.
Several early testers of Fable, including Cursor, GitHub, and Lovable, broadly confirm that the model is the most capable in the Anthropic lineup they've tested. Figma Director of Developer Product Matt Colyer said that the model is a "clear step forward in agentic coding and prototyping."
Additionally, Anthropic announced changes to its data retention policy for business customers. For Fable, Mythos and all future releases with similar or higher capabilities, Anthropic will require 30-day retention for all usage in both first- and third-party services. Though Anthropic says it will not train the models on this data, using it only to defend against "complex and novel attacks," it's a stark shift from the company's previous Zero Data Retention policy.
The release follows the company's announcement of a $65 billion funding round at a $965 billion valuation, which it said would allow it to significantly expand its compute capacity to scale Claude models.
Our Deeper View #
It's an interesting time for Anthropic to release its most capable model yet, given that just last week, its researchers called for an industry-wide slowdown on model development amid concerns that recursive self-improvement could arrive sooner than expected. However, it's inevitable that the larger these companies get, the more beholden they become to shareholder interests. Given that the company is preparing for an IPO at the same time as rival OpenAI, it may have believed it had no choice but to release Mythos, even in a somewhat stunted capacity as Fable. And now, backed by a pile of cash that allows it to actually scale these compute-hungry Mythos-class models, Anthropic has the resources to release them.