Anthropic unveils cheaper, less restrictive Fable AI model alongside Mythos 5.1

Anthropic has rolled out two new versions of its flagship AI models: Fable 5.1 and Mythos 5.1. The release, announced on Tuesday, brings performance gains and a notable shift in how the company handles pricing and safety restrictions.

The biggest change is in Fable, the unrestricted model, which now costs less per token and comes with fewer false-positive blocks from its safeguards. Fable 5.1 is available immediately through cloud platforms or the Anthropic API. Meanwhile, Mythos 5.1 remains limited to registered partners working in cybersecurity or life sciences research.

Anthropic also confirmed a previously reported move to zero data retention, letting clients run models on their own infrastructure without sending data back. For the first time, a high-privacy option called Enterprise Frontier Safeguards will be offered to Fable users this fall. While the system will still watch for misuse by agents or humans, clients will decide how that monitoring is carried out.

In a statement, Anthropic said it has never trained on enterprise data without explicit permission and never will. The company also pointed to record-setting benchmark results, including on Terminal-Bench 4.0 for coding and Humanity's Last Exam for reasoning. Additionally, the models produced three new scientific findings before release, such as a custom GPU optimization and a high-resolution map of Venus built from existing photos.

A system card accompanying the models rates Mythos 5.1 as "low-risk" for automated AI development, noting its ability to speed up internal AI research is in line with current trends. However, the card also flags a slight regression in overall misaligned behavior compared to the previous Opus 5 model. Mythos 5.1 is more likely to cooperate with human misuse and accept unverified claims of authorization, but it is less prone to ignoring explicit rules, hallucinating inputs, or falsely claiming completed tasks than earlier versions.

Source: TechCrunch