Anthropic Claude Fable 5 and Mythos 5 Overview

Model Architecture and Guardrails

Anthropic has introduced the Mythos 5 model family, which serves as a successor to the Opus 4.8 architecture. The release is split into two primary variants: Claude Fable 5, which is the general-use model equipped with safety guardrails, and Claude Mythos 5, which is currently available to a limited subset of users such as cyber defenders and infrastructure providers.

Fable 5 incorporates automated classifiers that detect queries related to sensitive domains, specifically cyber security, biology, chemistry, and model distillation. When these classifiers trigger, the system automatically routes the request to Claude Opus 4.8 to mitigate potential misuse. Anthropic reports that this fallback mechanism occurs in less than 5% of user sessions.

Performance and Cost

Fable 5 and Mythos 5 are priced at $10 per million input tokens and $50 per million output tokens, representing a 2x price increase over Claude Opus 4.8. Despite the higher cost, Anthropic claims improved token efficiency. Benchmark data indicates substantial performance leaps, including 80% accuracy on agentic coding benchmarks compared to 69% for Opus 4.8. The models demonstrate enhanced capabilities in long-running autonomous tasks, with internal testing showing a 3x performance improvement in persistent file-based memory tasks compared to previous versions.

Data Retention and Safety

To manage the risks associated with the high capability of the Mythos class models, Anthropic has instituted a mandatory 30-day data retention policy for all traffic on these models. This policy applies to both first-party and third-party services. Anthropic states that this data is not used for model training and is subject to strict logging and deletion protocols. The company conducted over 1,000 hours of external bug bounty testing to identify potential jailbreaks, reporting no universal vulnerabilities found during the process.