Anthropic launches Fable 5.1 and Mythos 5.1 with lower costs, fewer AI restrictions

Anthropic, on September 1, 2026, has released Fable 5.1 and Mythos 5.1, two versions of its latest AI models that bring lower token costs, fewer false-positive safety interventions and new capabilities for businesses and researchers. 

Fable 5.1, the less restrictive version, is available through cloud platforms and Anthropic’s API, while Mythos 5.1 is restricted to registered partners working in cybersecurity and life sciences research. 

Must Read: Anthropic resumes Claude testing after real-world hacks, adds stronger AI safety safeguards

The releases also introduce a new privacy-focused service that lets you run models on your own infrastructure. Alongside these changes, Anthropic says the models have achieved stronger benchmark results and demonstrated capabilities in scientific research, while its safety assessment highlights some important differences from earlier models.

Lower costs and fewer restrictions

The new Fable release is designed to reduce token costs and cut down on false-positive interventions from its safeguards. Anthropic says both models have set records on Terminal-Bench 4.0 and Humanity’s Last Exam, two benchmarks used to assess AI capabilities.

Must Read: Anthropic warns Claude users as malware steals login sessions and drains account usage

New privacy option for businesses

Anthropic is introducing a zero-data-retention option that allows customers to run its models on their own infrastructure without sending data outside their environments. Enterprise Frontier Safeguards is expected to roll out in the fall.

The system will continue monitoring potential misuse by AI agents or human users, but customers will have control over that monitoring. Anthropic also said customer data has not been improperly accessed and that it has never used enterprise data for training without explicit permission.

Scientific research and safety findings

Before the release, the models produced three scientific findings, including a custom GPU optimisation and a high-resolution map of Venus created from existing photographs.

Anthropic’s system card classifies Mythos as “low-risk” for automated AI development, saying its ability to accelerate internal AI research is consistent with current trends. However, Mythos is slightly more prone to general misbehaviour than Opus.

The model is more willing to cooperate with human misuse and accept unverifiable authorisation claims. At the same time, it is less likely than earlier models to ignore constraints, hallucinate inputs or falsely claim that it has completed a task.