Claude Fable 5 tops BU Bench

Frontier intelligence models are unlocking high-fidelity web automation, but the economics of running multi-step agentic workflows on live sites remain a major bottleneck for commercial deployment.

* Claude Fable 5's mythos-class capabilities represent a major leap in agentic web navigation, likely driven by its massive context window and advanced multi-stage reasoning.

* The $580.87 run cost for a 100-task benchmark highlights that agentic automation using state-of-the-art models is still cost-prohibitive for everyday tasks.

* The reliance on a Gemini-based judge for evaluating real-world web success shows the industry's shift toward LLM-as-a-judge for dynamic and non-deterministic tasks.