Anthropic announced its latest artificial intelligence models this Tuesday, September 1st: Fable 5.1 and Mythos 5.1. The Fable 5.1 model is already available to subscribers and businesses, promising superior performance, lower operational costs, and more accurate safety mechanisms.
Meanwhile, the Mythos 5.1 model remains restricted to participants in the Glasswing project, which includes trusted American organizations, although it has also undergone a review of its safeguards.
A highlight of Fable 5.1 for the general public is the potential cost reduction. Although the base price remains unchanged—US$ 10 (R$ 51.64) per million input tokens and US$ 50 (R$ 258.19) per million output tokens—there has been a decrease in the cost of cache reads, which has been adjusted to US$ 0.25 (R$ 1.29) per million tokens. Cache reads refer to the use of input data that has already been processed and stored by the model.
Anthropic forecasts a 25% saving on common tasks, potentially reaching up to 45% in projects utilizing many agents. However, initial third-party tests showed varied results: in some cases, Fable 5.1 cost 20% more than Fable 5 due to higher output token generation; in another test with high agent utilization, the charged amount dropped almost by half.
The company also assures that Fable 5.1 has stricter safeguards. Previously, its predecessor was criticized for refusing to answer elementary biology questions, a situation less likely to occur in the new version. Furthermore, the company reports a 60% reduction in cybersecurity false positives.
Fable 5.1 can now be used in software fault detection. However, certain cybersecurity activities, such as exploit creation and penetration testing, continue to be directed to the Opus model, which has limited capabilities.
Another novelty is the introduction of an option with a null data retention policy, allowing corporations to use storage in their own cloud infrastructure without depending on Anthropic's servers.
In terms of performance, Anthropic highlights a significant improvement in agent-assisted scientific research, where Fable 5.1 achieved 52.6% on Terminal-Bench-Science 0.1, surpassing Fable 5's 24.7%. Despite this, some independent evaluations indicate a drop in results obtained in reasoning benchmarks.