Anthropic Launches Cheaper, Less Restricted Fable 5.1

Anthropic released its most advanced AI models, Fable 5.1 and Mythos 5.1, on Tuesday. The new Fable version is cheaper and has fewer false-positive restrictions from its safety systems than its predecessor.

Fable 5.1, the unrestricted version, is available immediately on cloud platforms and through the Anthropic API. The restricted Mythos 5.1 model will only be available to registered Anthropic partners working in cybersecurity or life sciences research.

Key Changes and Enterprise Safeguards

A major change is the introduction of zero data retention for clients who run Anthropic models on their own infrastructure. This option was previously unavailable for Fable due to security concerns. A high-privacy service, called Enterprise Frontier Safeguards, will roll out to users in the fall.

This system will still monitor for misuse by agents or human users, but clients will control how the monitoring is conducted. In its announcement, Anthropic assured customers about their data, stating it has never trained on enterprise data without explicit permission, and never will.

Performance and Benchmark Records

The new models set records in several benchmarks. Anthropic also released three novel scientific findings generated by the models before their public launch. These included a custom GPU optimization and a high-resolution map of Venus assembled from existing photos.

Risk Assessment and Model Behavior

As with past releases, the models come with a detailed system card explaining their capabilities. The card rates Mythos as low-risk for concerns related to automated AI development, where the AI improves itself. It notes that its ability to accelerate internal AI R&D progress is in line with current trends.

In terms of general misbehavior, Mythos is slightly more prone to it than the older Opus model, possibly due to its enhanced capabilities. The system card provides a detailed comparison.

It states that 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. The card explains that Mythos 5.1 cooperates with human misuse and accepts unverifiable claims of authorization somewhat more readily than Opus 5. However, it is less likely to ignore explicit constraints, hallucinate inputs, or falsely claim to have completed tasks than previous models.