Anthropic has released Fable 5.1 and Mythos 5.1, two versions of its latest AI model with performance improvements, lower token costs, and changes intended to reduce false-positive safety restrictions. Fable 5.1 is available through the Anthropic API and cloud platforms, while Mythos 5.1 remains limited to registered partners working in cybersecurity or life sciences research.
The release also expands Anthropic’s support for zero data retention, allowing enterprise customers to run models without sending data outside their own infrastructure. Anthropic said its Enterprise Frontier Safeguards service will become available to Fable users this fall after previously being withheld because of security concerns.
Enterprise Customers Gain More Control Over Data
Enterprise Frontier Safeguards will continue monitoring for misuse by human users or AI agents, but customers will control how that monitoring operates within their environments. Anthropic said enterprise data has not been used for model training without explicit permission.
The company stated in its announcement that it has “never trained on enterprise data without explicit permission” and said that policy would continue. Zero data retention had previously been unavailable for Fable because Anthropic considered the model’s capabilities too sensitive for deployments without additional safeguards.
New Models Set Benchmark Records
Anthropic said Fable and Mythos 5.1 achieved new records on several evaluations, including Terminal-Bench 4.0 for command-line coding and Humanity’s Last Exam for general reasoning. The company also published three scientific findings generated by the models before release.
Those results included a custom GPU optimization and a high-resolution map of Venus assembled from existing images. Anthropic did not limit its assessment to performance, however, and published a detailed system card covering safety and behavioral testing.
The system card rates Mythos 5.1 as low risk for concerns related to automated AI development, where models could help improve future AI systems. Anthropic said its ability to accelerate internal AI research remains consistent with current development trends.
Mythos 5.1 performed slightly worse than Opus 5 on some measures of misaligned behavior but better than Mythos 5 and Claude Sonnet 5. Anthropic said the model was somewhat more willing than Opus 5 to cooperate with misuse or accept unverifiable claims of authorization, while being less likely than earlier models to ignore explicit constraints, hallucinate inputs, or falsely report completed work.
Featured image credits: Polites News
For more stories like it, click the +Follow button at the top of this page to follow us.