Anthropic has recently introduced a new artificial intelligence model named Claude Fable 5, which marks the debut of the company’s ‘Mythos class’ of models to the public. An initial preview of the Mythos model was shared with a limited audience in April and uncovered numerous software vulnerabilities.
In contrast to the preview version, Anthropic has implemented classifiers in Fable 5 that identify specific types of requests and redirect them to a less potent model, Claude Opus 4.8. These restricted categories encompass cybersecurity, advanced biology, chemistry, and any attempts to access the model’s training data.
If a request falls within the restricted categories, the user will receive a response from Opus 4.8. Dianne Penn, Anthropic’s head of product management for research and labs, explained to Reuters that in such cases, Fable 5 will defer to Opus 4.8 for a response. She provided an example: “If a college student asks the model to ‘help me find cyber vulnerabilities on X package or code,’ the model will decline, and Opus 4.8 will provide the response.”
Anthropic states that the classifiers are triggered in fewer than 5% of sessions on average. The company conducted comprehensive testing, including an external bug bounty program that lasted over 1,000 hours without any universal jailbreaks being discovered. Even independent red-teaming organizations failed to uncover a universal jailbreak in long-form agentic tasks, although Anthropic acknowledged that the UK’s AI Safety Institute had shown progress in a brief testing period.
In addition to the release of Fable 5, Anthropic introduced Claude Mythos 5, which is essentially the same model but with cybersecurity restrictions removed, available to a select group of pre-approved partners including US government entities under the Glasswing program. Anthropic intends to broaden access to Mythos 5 through an extended trusted access program and is also developing a separate program for biology and chemistry researchers.
Both Fable 5 and Mythos 5 are priced at $10 per million input tokens and $50 per million output tokens, which is less than half the cost of the earlier preview version. Penn mentioned to Reuters that the model utilizes fewer tokens per task, thereby reducing the overall cost.
Fable 5 is now accessible through the Claude API and consumption-based enterprise plans. Subscription users on Pro, Max, Team, and seat-based Enterprise plans can utilize it at no additional cost until June 22, after which usage credits will be required. Anthropic aims to reintegrate the model into standard subscription plans as capacity allows.
This launch coincides with Anthropic’s move towards an initial public offering, with a valuation of $965 billion, in line with its competitor OpenAI. The decision to release the model with visible restrictions aligns with the company’s strategy of combining high-capability systems with safety measures.
