How many of your agent's calls actually need a frontier model?

За 30 секунд

Что произошло

NVIDIA's Switchyard router sends only 7% of agent calls to expensive frontier models, cutting costs 74% while retaining 93% accuracy.

Почему это важно

You build multi-step agents that make many LLM calls and want to reduce spending without losing capability on hard tasks.

На что обратить внимание

Routing widens cost variance by 67% across runs and adds latency. The judge model itself costs 21% of routed spend, making the math only work if your models differ sharply in price.