Unverified50% confidenceBenchmarkExact time
在博主样本中Astra约94%的输入、Sonnet约79%的输入来自缓存
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
9/16/2026
First Seen
Valid until: 9/30/2026
Sources
Related Entities
Related Claims
Unverified当Sonnet的上下文突破5万到10万token范围时，它在上下文窗口中查找重复词的准确率会暴跌近50%66% similarUnverified发帖者指出编码任务中缓存命中率通常高达90%以上62% similarUnverified投机解码的草稿接收率每提升10%，整体吞吐量可提升15%-25%59% similarVerified输出Token价格通常为输入的4-6倍，源于预填充与解码两阶段的算力差异，解码阶段GPU利用率可能仅为预填充的10-20%59% similarUnverified在试卷批改案例中，Cowork 通过一句指令生成了含平均分约72%、最低30分、最高100分的分析及HTML报告59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/924288API
curl https://kongchang.com/api/v1/knowledge/claims/924288MCP
get_claim(id=924288)