Anthropic Launches Claude Fable 5: Public AI Under Control

Anthropic has announced the availability of Claude Fable 5, a public version of its AI model Mythos, marking a significant step in opening its technologies to the general public. This launch, which took place on Tuesday, is accompanied by rigorous security measures to prevent potential abuse.

Claude Fable 5 is designed to excel in areas such as software engineering, knowledge management, and vision. However, in high-risk sectors like cybersecurity, biology, chemistry, and distillation, the model is programmed to block certain responses, reverting to Claude Opus 4.8.

Initially launched in preview in April, Mythos was reserved for a select few partners, particularly for cybersecurity reasons. Last week, Anthropic expanded this access to hundreds of organizations across 15 countries, particularly targeting those involved in critical infrastructure.

Access to Claude Fable 5 is now available via Anthropic's API and consumption-based enterprise plans. Until June 22, this model will be included at no additional cost in Pro, Max, Team, and enterprise plans. Starting June 23, access will require usage credits, with the goal of reintegrating it as a standard subscription feature as soon as possible.

Anthropic has also introduced a new version of Mythos, named Mythos 5, for organizations already approved to use this advanced model.

The launch of Fable 5 comes as Anthropic prepares to enter public markets, alongside other AI giants like OpenAI and SpaceX. The company has recently called for global coordination to curb the development of advanced AI, warning that these systems could soon reach recursive self-improvement.

To secure the model, Anthropic has subjected its classifiers to rigorous testing, including an external bug bounty program and collaborations with red-teaming organizations. No universal bypass has been found after more than 1,000 hours of testing.

Furthermore, Anthropic now imposes a 30-day data retention policy on all traffic related to Fable 5 and Mythos 5, even for companies with non-retention agreements. Anthropic clarified that it would not use the data for training but solely to defend against complex and novel attacks, including new bypasses, and to identify and reduce false positives.

Despite these restrictions, Claude Fable 5 has demonstrated impressive performance. According to analytics firm Hex, it scored 90% on a complex analytics benchmark. Vibe Base44 praised its capabilities in creating complete applications, while Genspark noted its superior performance in user interface design and game coding.

Anthropic claims that instances where Fable must refer to Opus 4.8 are rare, with early data showing that at least 95% of Fable sessions operate entirely on the model's own responses.

The usage cost for Fable 5 and Mythos 5 is set at $10 per million input tokens and $50 per million output tokens, which is double the price of Opus 4.8. This pricing could limit its widespread adoption, as many companies are already concerned about the rising costs of AI.

Anthropic anticipates strong demand for Fable 5, although it is difficult to predict. Some companies, like Rakuten, believe the benefits of Fable 5 justify its cost, highlighting its ability to reflect and validate its own work for highly autonomous operations.