Anthropic's Fable 5 and Mythos 5: A New Frontier in AI Capability
Matthew Bermango watch the original →
the gist
Anthropic has released Fable 5 and its un-safeguarded counterpart, Mythos 5, a 10-trillion parameter model class that demonstrates state-of-the-art performance in complex agentic coding, long-horizon tasks, and high-density reasoning.
The Fable and Mythos Model Class
Anthropic has introduced a new generation of models, Fable 5 and Mythos 5. Fable 5 is designed for general use with safety guardrails, while Mythos 5 is an un-safeguarded version intended for the security community to harden software and identify vulnerabilities. Both models are built on a massive 10-trillion parameter architecture, marking a significant leap in reasoning, coding, and agentic capabilities.
Performance and Agentic Behavior
In testing, Fable 5 demonstrates a unique "exploratory" behavior. Unlike previous models that provide quick, shallow answers, Fable 5 approaches tasks by analyzing entire codebases and considering every possible edge case. It excels at long-horizon tasks, showing no degradation in performance over extended periods. A notable feature is its "Ultra Code" workflow, which utilizes a planning agent to delegate sub-tasks to hundreds of parallel agents. While powerful, this approach is compute-intensive and can lead to high token consumption.
Information Density and Efficiency
Fable 5 exhibits an extremely high level of information density. Its outputs are verbose and technically dense, effectively increasing the model's intelligence per token. This efficiency allows the model to accomplish more within a given compute window. However, this density creates a "readability gap" for humans, leading to speculation that future AI models might develop hyper-dense, non-human-readable languages for internal communication to maximize efficiency.
Operational Quirks and User Experience
Despite its capabilities, the model is notably "chatty" and prone to excessive clarifying questions. Users often find themselves in a loop of confirming specs, agentic approaches, and summaries before the model executes a task. Additionally, the model is slow, likely due to its massive parameter count and the depth of its internal reasoning processes. Users are encouraged to use model routing—assigning the most complex tasks to Fable 5 while relying on smaller models like Haiku or Sonnet for routine work—to manage costs.
Strategic Release and Privacy
Anthropic’s delayed release of Mythos 5 appears to be a strategic move to maintain a competitive lead in model development. The company has implemented a new 30-day data retention policy for these models, explicitly stating that data will not be used for training, but rather to defend against novel jailbreaks and cross-request attacks. Anthropic also notes that if a user attempts to use the model for distilling or training competing models, the system will automatically fall back to the older Opus 4.8 model.