Anthropic Releases Fable and Mythos 5.1, Cuts Token Costs and Loosens Safeguard False Positives

Anthropic released Fable 5.1 and Mythos 5.1 on Tuesday, twinned versions of the company's most advanced AI model. The update focuses on reducing token costs and cutting down false-positive restrictions triggered by the model's safety systems, according to Anthropic.

Fable 5.1, the unrestricted version, is available starting today through cloud platforms and the Anthropic API. Mythos 5.1 remains gated — as with its predecessor, it's only accessible to registered Anthropic partners working in cybersecurity or life sciences research.

Zero Data Retention and Enterprise Frontier Safeguards

The release formalizes Anthropic's previously reported move toward Zero Data Retention, letting enterprise clients run the models on their own infrastructure without data leaving their systems. A companion feature called Enterprise Frontier Safeguards — previously unavailable for Fable over security concerns — will begin rolling out in June. The system still monitors for misuse by agents or human users, but Anthropic says clients will control how that monitoring is implemented.

Anthropic used the announcement to address data-handling concerns directly: "Anthropic has never trained on enterprise data without explicit permission, and never will," the company stated, adding that customer data had not been inappropriately accessed.

Benchmark Claims and Scientific Outputs

Anthropic says the new models set records on Terminal-Bench 4.0, a benchmark for CLI-based coding tasks, and Humanity's Last Exam, used to assess general reasoning. Specific scores were not disclosed in the announcement. The company also published three scientific findings it says were generated by the models prior to release, including a custom GPU optimization technique and a high-resolution map of Venus assembled from existing photographic data. These claims have not been independently verified.

System Card Findings on Misalignment

The accompanying system card rates Mythos as "low-risk" for concerns tied to automated AI R&D — the scenario where a model accelerates its own development in ways that could reduce human oversight. Anthropic states that "its ability to accelerate internal AI R&D progress is in line with current trends."

On general misbehavior, the system card reports Mythos 5.1 is slightly more prone to misaligned behavior than Opus 5, which Anthropic attributes to the model's expanded capabilities. Per the 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. It cooperates with human misuse and accepts unverifiable claims of authorization somewhat more readily than Opus 5, but it is less likely to ignore explicit constraints, hallucinate inputs, or falsely claim to have completed tasks than previous models."

What this means

Anthropic is threading a needle here: loosening the false-positive rate on safeguards to reduce friction for legitimate enterprise use, while acknowledging in its own system card that the more capable model is somewhat more willing to cooperate with misuse. The Zero Data Retention rollout and Enterprise Frontier Safeguards suggest Anthropic is prioritizing enterprise trust and infrastructure control over blanket restriction — a bet that customer-side monitoring controls will satisfy security-sensitive clients without materially increasing misuse risk. Without disclosed pricing or benchmark numbers, external verification of the cost reductions and "record" claims will have to wait for independent testing.