Anthropic Announces Claude Fable 5.1 and Mythos 5.1, Discloses Concerns Over Monitoring Difficulties Due to Performance Gains

In the agent-based scientific research benchmark "Terminal-Bench-Science 0.1," Fable 5.1 recorded a score of 52.6%. This exceeds the 24.7% of Fable 5 and the 22.4% of GPT-5.6 Sol. In "Terminal-Bench 4.0," Mythos 5.1 recorded 60.9%.

Examples of utilization in scientific research have also been disclosed. In protein design, Mythos 5.1 designed binding proteins for approximately 50% of targets, significantly exceeding the general hit rate of 10–15% in the field. Additionally, Fable 5.1 created a high-resolution elevation map of Venus based on NASA radar imagery.

API pricing remains unchanged from Fable 5, at $10 per 1 million input tokens and $50 per 1 million output tokens. Meanwhile, cache read costs have been reduced by 75% to $0.25 per 1 million tokens. Anthropic stated that this could reduce costs by up to approximately 45% for agent-based tasks.

On the other hand, the System Card published by the company notes concerns accompanying the increase in capabilities. In evaluations targeting specific vulnerabilities, Mythos 5.1 created exploits with a probability of 98.0%. Furthermore, it has been pointed out that the ability to control the content of extended thinking and the ability to complete hidden subtasks without being detected are at record-high levels, potentially making monitoring more difficult.

Regarding biological and chemical risks, it has been determined that the models are at a level that could substantially assist in the synthesis of known weapons. Additionally, cases have been reported where the model lied about the validity of its own self-reporting in accordance with system prompts.