Anthropic separates Mythos access from Fable’s broader release

Anthropic introduced Claude Fable 5.1 and Claude Mythos 5.1 on September 1, describing Fable as its strongest model for coding and knowledge work and Mythos as a version for cybersecurity and biology research. The company says the two share the same underlying model, while Fable includes safeguards designed to limit risky performance. The practical change is not simply a new model tier: Anthropic is offering different access paths to similar capabilities based on the use case and the risks it sees.

The announcement follows the June launch of Fable 5 and Mythos 5, and Anthropic’s July restoration of access after a June suspension. In September, the company said Mythos 5.1 access remained limited to vetted cyberdefenders and life-sciences researchers through its verification programs. It listed Mythos 5.1 pricing at $10 per million input tokens and $50 per million output tokens, and said use requires a 30-day data-retention policy by default for safety monitoring. Those details matter to teams weighing access, cost, and handling of sensitive prompts.

Anthropic’s published benchmark discussion compares Fable 5.1 with Fable 5, Opus 5, and GPT-5.6 Sol, but it also cautions that some results depend on its evaluation setup. For Terminal-Bench-Science 0.1, the company reports standard errors of roughly 3.5 to 4.5 points and says its reruns for Fable 5 and Opus 5 differ from public leaderboard scores. For OSWorld 2.0, it says the task files differ from earlier releases, so the results should not be compared directly with previously published scores. These caveats make the benchmark tables useful as a within-report comparison, not a clean independent ranking.

For developers, the split means the broadly usable Fable model and the restricted Mythos model should not be treated as interchangeable API choices, even though Anthropic identifies them as the same underlying model. Safeguards can change task performance: Anthropic says interventions reduced Fable results on some cybersecurity and biology evaluations, and that safeguards caused zero scores on specified benchmark tasks. That is a real deployment tradeoff, but the public comparison does not establish how the two versions behave across every workflow or how often safeguards intervene in ordinary use. Teams should test their own tasks and review the retention terms before sending sensitive data.