A deep dive into how subscription tiers, volume rebates, and semantic caching stack on top of list price. With real examples.
Three separate mechanisms reduce what you pay, and they stack multiplicatively rather than additively. Understanding the order they apply in is the difference between an estimate that is close and one that is off by fifteen points.
Layer 1: subscription tier
Your plan sets a base discount off list price, applied per token at request time. Free pays list, Lite pays 10% under, Pro 20%, Max 28%. This is a flat multiplier — it does not depend on how much you use.
Layer 2: volume rebate
Once monthly spend crosses a threshold, every token above it drops another tier. Rebates are calculated on pooled spend across all models and capabilities, which is why consolidating onto one gateway matters more than negotiating any single vendor down.
Layer 3: semantic caching
Cache hits are not discounted tokens — they are tokens you never buy. This layer applies last and is the only one that can push effective cost below any vendor's floor price.
A worked example
List price for the month $10,000
Pro tier (-20%) $ 8,000
Volume rebate (-9% effective) $ 7,280
Semantic cache (31% hits) $ 5,023
Effective discount 49.8%
Order matters. Apply caching first and the same inputs look like a 44% saving, because the rebate then lands on a smaller base and drops a tier.