Unverified50% confidenceBenchmarkExact time
Across the four models, total calls dropped from 12000 to 3588, tokens dropped 70.1%, and cost dropped 69.6%; about 70% of requests were answered by the cache, of which exact matching contributed about 39% and the semantic layer contributed an additional about 31%.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/19/2026
First Seen
Valid until: 1/17/2027
Sources
Related Entities
Related Claims
UnverifiedReducing model parameters from 70B to 7B lowers both the VRAM required for inference and the per-request computation by approximately one order of magnitude.75% similarUnverifiedIn a simulation of 2000 requests, the small model (1 unit cost/request) had an accuracy of 74.5% and a cost of 1.34 per correct answer; the large model (20 unit cost/request) had an accuracy of 95.8% and a cost of 20.88 per correct answer.74% similarUnverifiedIn the hybrid mode at τ=0.4, 27 of the 400 cases are escalated (6.8%), billed Tokens drop from 1,222,211 to 82,499, and the saving including plugin overhead is 86.1%.74% similarUnverifiedIn 11 of the 12 models, between 87.5% and 100% of capitulation cases stopped at the exact wrong number provided by the user, meaning capitulation typically involves switching sides entirely.74% similarUnverifiedAt the 0.81 threshold, the proportion of cached answers that "fit the question" is 5.8 to 10.7 percentage points lower than fresh answers; converted to all responses, the overall quality loss is about 2 to 3 percentage points; Haiku 5.5 has the largest gap (10.7 percentage points), with a cached-answer pass rate of only 67.9%.74% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1065406API
curl https://kongchang.com/api/v1/knowledge/claims/1065406MCP
get_claim(id=1065406)