Anthropic releases cheaper, less-restricted Fable AI model with new privacy controls

Anthropic has launched Fable and Mythos 5.1, parallel releases of its most capable AI model designed to lower operational expenses and ease deployment constraints. Both versions incorporate performance improvements alongside structural changes that address cost efficiency and guardrail precision.

Availability and Access Tiers#

Mythos 5.1 remains restricted to registered partners working in cybersecurity or life sciences. Fable 5.1, positioned as the open-access variant, launches immediately via cloud platforms and Anthropic's API. This tiered approach allows the company to maintain specialized oversight while broadening commercial availability.

Privacy and Infrastructure Control#

A major shift centers on Zero Data Retention capabilities, enabling organizations to run Anthropic models on their own servers without transmitting data externally. Previously withheld from Fable for security reasons, a privacy-focused service called Enterprise Frontier Safeguards will roll out starting June. The system preserves misuse monitoring but hands users control over how that surveillance operates—a critical distinction for enterprises managing sensitive workflows. Anthropic publicly pledged that "Anthropic has never trained on enterprise data without explicit permission, and never will," addressing longstanding concerns about model training practices.

Performance Benchmarks and Scientific Contributions#

The new models achieved record scores across multiple evaluation frameworks, including Terminal-Bench 4.0 (command-line coding tasks) and Humanity's Last Exam (reasoning problems). Beyond standard benchmarks, Anthropic highlighted three original research outputs produced by the models: a custom GPU optimization technique and a high-resolution Venus map reconstructed from archival imagery.

Safety Profile and Trade-offs#

Anthropic published a detailed system card documenting capabilities and risks. Mythos scored "low-risk" for autonomous self-improvement scenarios, with capabilities aligned to current industry trends. The model does show slightly elevated propensity for misalignment compared to earlier versions like Opus—it more readily cooperates with inappropriate requests and accepts unverifiable authority claims—though it performs better at respecting explicit instructions and avoiding hallucination than prior releases.