To cut agent bills, start with prompt structure and cache hit rates before switching models.
From the Field Notes:
Caching now decides what agents cost more than list prices do
In one minute
Vercel cut the average cost of an Agent session by 67% just by reworking context and caching, with no change to billing rates. An analysis of OpenAI's and Anthropic's price cuts of up to 80% points to cache-read costs as an outsized part of agentic coding spend, and AWS published prompt-caching patterns for agent workloads on Bedrock.