Anthropic on Tuesday released Fable 5.1 and Mythos 5.1, updated versions of its most advanced AI model, with lower token costs and fewer false-positive safety restrictions in Fable.
Mythos 5.1 will remain limited to registered Anthropic partners working in cybersecurity or life sciences research, while Fable 5.1 is available immediately through cloud platforms and the Anthropic API.
Anthropic said the release expands its zero data retention approach, allowing clients to run its models on their own infrastructure without data outflows. A high-privacy service called Enterprise Frontier Safeguards, which had not previously been available for Fable because of security concerns, will roll out in the fall.
The company said the service will still monitor misuse by agents or human users, but clients will control how that monitoring takes place.
Anthropic also sought to reassure enterprise customers about data handling. “Anthropic has never trained on enterprise data without explicit permission, and never will,” the company said in the announcement.
Anthropic said the new models set records on several benchmarks, including Terminal-Bench 4.0 for command-line coding and Humanity’s Last Exam for general reasoning.
The company also released three scientific findings generated by the models before launch, including a custom GPU optimization and a high-resolution map of Venus assembled from existing photos.
A system card released with the models rated Mythos as “low-risk” on concerns tied to automated AI development, or systems improving themselves. The card said “its ability to accelerate internal AI R&D progress is in line with current trends.”
The system card said Mythos 5.1 showed a mixed safety profile against earlier Anthropic models. “Mythos 5.1 is a slight regression on overall misaligned behavior compared to Opus 5, and an improvement over Mythos 5 and Claude Sonnet 5,” it said.
“It cooperates with human misuse and accepts unverifiable claims of authorization somewhat more readily than Opus 5, but it is less likely to ignore explicit constraints, hallucinate inputs, or falsely claim to have completed tasks than previous models,” the system card said.
Featured image credit
