Anthropic cuts request payload from about 800 KB to 1.8 KB, and why that doesn't clash with NVIDIA's long-context advice
从 800 KB 到 1.8 KB,账单却一分不减:为什么英伟达劝你砍掉长上下文,Anthropic 却帮 100 万 Token 修高速路
Anthropic's Message Threads drops the median long-conversation request payload from about 800 KB to 1.8 KB by cutting repeated uploads, validation and tokenization. It does not reduce token consumption or the bill.
Why it matters: Anthropic trims gateway latency with incremental transfer while Nvidia trims decode cost with context compression, two routes for two different long-conversation costs.