gpt-4.1 比 claude sonnet 4.6 便宜，直到缓存机制加入讨论
huggingface 内部基准测试显示，仅优化价格会导致模型路由失败，因为缓存机制使 claude sonnet 4.6 在代理任务中的有效成本仅为 gpt-4.1 的一半。
huggingface tested model routing on 417 tasks using the appworld test challenge and found that claude sonnet 4.6 was nearly twice as cost-efficient as gpt-4.1 despite having higher base token prices and longer reasoning trajectories. the unexpected cost difference was driven by sonnet's aggressive cache-read pricing, which drastically lowered effective input costs during high-reuse agent workloads. the report concludes that effective routing systems must move beyond simple classification and optimize for the interaction between model pricing, workload characteristics, and infrastructure factors like latency and compliance.