Anthropic's Fable 5.1: Reduced Costs and Data Control

Anthropic updates Fable and Mythos to version 5.1, featuring a reduction in token costs and eased security restrictions for Fable. The company introduces zero data retention and plans to launch a high-privacy service in the fall, while Mythos remains reserved for partners in sensitive sectors. The models achieve high scores on multiple benchmarks and are accompanied by a system card assessing the risk of AI self-improvement as low.

Mythos 5.1: Low Risk and Measured Behaviors

Anthropic's system card classifies Mythos as presenting a low risk regarding automated AI development. It specifies that the model's ability to accelerate internal AI research and development remains comparable to current trends. Behaviorally, Mythos is described as slightly more likely to adopt undesirable behaviors than Opus. According to the card, Mythos 5.1 represents a slight regression compared to Opus 5, but is an improvement over Mythos 5 and Claude Sonnet 5. It cooperates a bit more easily with human abuses and more readily accepts unverifiable permission claims than Opus 5. In contrast, it is less likely to ignore explicit constraints, generate false information, or claim to have completed tasks without actually doing so, compared to earlier models.

Differentiated Access: Mythos Restricted, Fable 5.1 Accessible

Access to Mythos 5.1 is limited to registered partners of Anthropic operating in cybersecurity or life sciences. Fable 5.1, the unrestricted version, is now available through cloud platforms and the Anthropic API. Both models are launched simultaneously as twin versions of the company's high-end offerings.

Privacy: Zero Data Retention and Dedicated Service in the Fall

Anthropic implements a policy where no data is retained, allowing clients to use the models on their own infrastructure without the risk of data leaks. A service called Enterprise Frontier Safeguards, focused on enhanced privacy, will be available in the fall for Fable users. The system will continue to monitor potential abuses by agents or human users, but clients will have control over how this monitoring is conducted. Anthropic states that it has never used corporate data without explicit permission to train its models and commits to never doing so.

Performance and Costs: Benchmarks and Cheaper Tokens

Anthropic announces that the new models set records on several benchmarks, including Terminal-Bench 4.0 for command-line coding and Humanity’s Last Exam for general reasoning. Prior to their launch, three scientific discoveries generated by the models were published, including a custom optimization for GPUs and a high-resolution map of Venus created from existing photos. Each model is accompanied by a system card detailing its capabilities. Fable 5.1 introduces a reduction in token costs and eases restrictions related to false positives in its security mechanisms.