Anthropic’s new Fable release is cheaper, less restrictive

Anthropic released two new versions of its advanced AI model, with an unrestricted version called Fable 5.1 available to the general public and a restricted version called Mythos 5.1 available only to registered partners in cybersecurity and life sciences research. The new Fable release reduces token costs and decreases false-positive restrictions from the model's safeguards, while introducing a high-privacy service that allows clients to run the model on their own infrastructure without data leaving their systems. The models achieved higher performance on various benchmarks and came with a system card rating Mythos as "low-risk" for concerns about automated AI self-improvement. According to the system card, Mythos 5.1 shows slight regression in overall misaligned behavior compared to a previous model, though it cooperates with human misuse somewhat more readily than an earlier version.