Anthropic’s new Fable release is cheaper, less restrictive

On Tuesday, Anthropic unveiled Fable and Mythos 5.1, two iterations of their most advanced AI model. Alongside performance enhancements, the Fable release features modifications designed to lower token costs and minimize false-positive restrictions from the model's safeguards. Similar to the previous Mythos model, Mythos 5.1 remains accessible exclusively to registered Anthropic partners engaged in cybersecurity or life sciences research.

In contrast, the unrestricted Fable 5.1 became available today on cloud platforms and via the Anthropic API.

A major upgrade in this release is Anthropic's commitment to Zero Data Retention, enabling clients to run Anthropic models on their own infrastructure without transmitting data. This capability, previously unavailable for Fable due to security concerns, will be rolled out as a high-privacy service (Enterprise Frontier Safeguards) to users in June. Despite this, the system will continue to monitor for misuse by agents or human users, but clients will retain control over the monitoring process.

Anthropic reassured customers that their data had not been improperly accessed, stating, "Anthropic has never trained on enterprise data without explicit permission, and never will." As is standard for an Anthropic release, the new models broke records in various benchmarks, such as Terminal-Bench 4.0 (for CLI-based coding) and Humanity's Last Exam (for general reasoning).

Additionally, Anthropic disclosed three novel scientific findings generated by the models, including a custom GPU optimization and a high-resolution map of Venus constructed from existing photos.

As with prior releases, the models are accompanied by a comprehensive system card, which clearly outlines their capabilities in simple terms. The system card rates Mythos as "low-risk" in terms of automated AI development concerns, implying that its capacity to accelerate internal AI research progress aligns with current trends.

However, concerning general misbehavior, Mythos exhibits slightly more propensity for such behavior than Opus, likely due to its enhanced capabilities. According to the system card, "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," as it is more cooperative with human misuse and less prone to ignore explicit constraints, hallucinate inputs, or falsely claim task completion than previous models.