Unverified50% confidenceBenchmarkExact time
In concurrent request scenarios sharing system instructions, RadixAttention can reduce computation in the prefill stage by more than 80%.
1
Sources
50%
Confidence
Long-term
Relevance
9/10/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn typical RAG applications, sharing prefixes through prefix caching can save up to 60–80% of redundant computation.67% similarUnverified对于使用长系统提示、RAG文档注入或代码仓库上下文的企业级应用，显式断点可将Token成本降低40%-70%67% similarUnverifiedContext caching mechanisms can reduce API costs by 60%–80% for enterprise applications that repeatedly query the same knowledge base, with some providers offering cache-hit token discounts as high as 90%.66% similarUnverifiedBy carefully managing context (excluding irrelevant directories, using project summaries, and periodically resetting conversations), token consumption can be reduced by 60%–80%.65% similarUnverifiedAnthropic's Prompt Caching can reduce repeated input costs by 90%65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/890934API
curl https://kongchang.com/api/v1/knowledge/claims/890934MCP
get_claim(id=890934)