Unverified50% confidenceFactExact time
Prompt caching is isolated per model rather than per session; measured in real coding sessions, when switching models mid-session, in 82% of cases cache reads fell below 10% of the previous value, and a warm 107000-token prefix became 110000 tokens of uncached input at roughly ten times the cached price.
1
Sources
50%
Confidence
Long-term
Relevance
10/19/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedOverly long system prompts bring hidden overhead: for example, when the system prompt has 500 tokens while the actual input has only 50 tokens, it means 91% of the cost goes to boilerplate text; auditing and trimming the system prompt by 40% can reduce costs.72% similarUnverifiedAfter a long interruption (cache expiry), compaction is actually cheaper than simply continuing the conversation, because each subsequent turn then reads only about 15k tokens instead of the entire context.72% similarUnverifiedPrompt Caching在提示词前缀完全相同且超过约1024 token阈值时命中，缓存token通常按原价的25%甚至更低计费71% similarUnverifiedThe actual extent of cache savings depends on the hit rate; if 80% of input hits the cache, the actual input cost is approximately 0.8 × cache price + 0.2 × regular price.70% similarVerifiedThe cache read mechanism retains the KV cache for recurring long prompt prefixes, charging at an extremely low rate upon a cache hit; the leaked cache read price for Opus 5.5 is approximately 1/20 of the input price.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1066276API
curl https://kongchang.com/api/v1/knowledge/claims/1066276MCP
get_claim(id=1066276)