Anthropic: Claude Fable 5, Enhanced Power and Security

Anthropic recently launched Claude Fable 5, the first model from the Mythos family made available to the public. This model stands out due to its increased power and the integration of new safeguards, which has resulted in a doubling of the price compared to previous versions.

The launch of Claude Fable 5 took place on June 9, after Anthropic deemed the safeguards robust enough to allow for public distribution. According to The Verge, these safeguards block responses in areas deemed high-risk, although the effectiveness of these measures remains a topic of debate.

Improved Performance on Long Tasks

The Fable 5 model particularly excels in autonomous software engineering, advanced reasoning, and visual analysis. On the SWE-Bench Pro benchmark, it achieves a score of 80.3%, surpassing Claude Opus 4.8 (69.2%), GPT-5.5 (58.6%), and Gemini 3.1 Pro (54.2%). However, this advantage is mainly evident in long and complex tasks, with a smaller gap on shorter tasks.

Claude Mythos 5: Limited Access

In parallel, Anthropic also introduced Claude Mythos 5, which shares the same underlying model as Fable 5 but with less stringent safeguards. Currently, access is limited to organizations participating in the Glasswing program, with an extension planned through a "trusted access program."

Both models are priced at $10 per million tokens in input and $50 in output, which is double the rates of Claude Opus 4.8.

Safeguards for Double-Edged Capabilities

The safeguards of Fable 5 block responses related to biology, chemistry, model distillation, and cybersecurity. When a sensitive request is detected, the model hands off to Claude Opus 4.8. Anthropic claims that 95% of sessions among early users did not trigger any filters, and no jailbreaks were successful during over 1,000 hours of internal testing.

Published figures on Mythos Preview show that the model is capable of transforming a software patch into an operational exploit within hours, compared to several weeks previously. Out of 21 vulnerabilities in the Windows kernel tested, Mythos caused a blue screen in 18 cases.

A Two-Tiered Business Model

Anthropic offers a two-tiered access model: Fable 5 for the general public and Mythos 5 for a select clientele. CEO Dario Amodei emphasizes safety as a selling point, but the doubling of prices could influence the perception of this "safely designed" model.