Unverified50% confidenceBenchmarkExact time
在1.78百万token的长会话实测中，因为97%的token被缓存，发送token最多的配置反而最便宜
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
8/20/2026
First Seen
Valid until: 11/18/2026
Sources
Related Claims
UnverifiedIn a typical 30-minute programming session, terminal command output can consume approximately 118,000 tokens, of which less than 20% is useful information.63% similarUnverifiedDeepSeek-Reasonix can reduce long-session costs to one-fifth (saving approximately 80%) through an intelligent caching mechanism.63% similarUnverifiedAgentMemory consumes only approximately 1,900 tokens per session, saving 92% compared to manually pasting context.63% similarUnverifiedInput tokens now account for more than 90% of total non-cached tokens, and the cost share of input context has risen from 47.5% to nearly 70%.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/778560API
curl https://kongchang.com/api/v1/knowledge/claims/778560MCP
get_claim(id=778560)