How we count savings
Every savings number Kertus AI reports is measured across four layers, reported net of what your provider already gives you, and reconciled to raw usage records.
Provider-native cache discounts
Never claimedDiscounts your provider already gives you for repeated prompt prefixes. These are passed through and credited to you in the dashboard; they are never counted as savings we produced.
Semantic-cache avoided cost
ClaimedA request answered from the semantic cache never reaches the provider, so 100% of its provider cost is avoided. The cache stores masked text only, encrypted, with a TTL.
Model-routing arbitrage
ClaimedA task routed from a frontier model to a cheaper model per your routing rules saves the price difference between the two.
Prompt compression
Measured token delta · liveLive since July 2026: conservative whitespace and duplicate removal, measured as the estimated token delta and priced at the request's input rate. Off by default; enabled per project.
Reported savings = L2 + L3 + L4, net of L1
A worked example
A request answered from semantic cache avoids 100% of its provider cost (L2). A summarisation request routed from a frontier model to a mini-model saves the price difference (L3). Provider-native prefix-cache discounts (L1) are shown in your dashboard but never claimed as Kertus savings. Reported savings = L2 + L3 + L4.
The methodology is versioned: the active version is returned with every savings API response. The dashboard reconciles every reported number to raw usage records, so finance can audit the claim, not just read it.