LLM CFO

Committed use, cache rates and true-ups

Updated: September 27, 2026

By the LLM CFO team

AI model spend has moved from a line item to a procurement category. At that scale the mechanics that govern a large bill are not the same ones that govern a small one, and the negotiation is almost never about the headline rate.

Term 1: the floor, and who owns the risk

A committed-use agreement typically trades 10–35% off list for a 12-month floor. That is a bet on your own forecast, and the failure mode is specific: commit to a floor, then optimise onto a cheaper model, and end up paying committed rates for tokens you no longer need. Three controls make the bet survivable.

Term 2: the cache rate

Cached input is billed as a separate rate and, on cache-heavy agent workloads, is the majority of input cost. Reported flagship rates put cache reads at $0.20 per million against $4.00 for fresh input. Negotiate the cache rate explicitly — it is a separate line in the rate card and a separate line on the invoice.

Term 3: the line items the rate card missed

Server-side tools are billed independently of tokens. Web search, code execution and similar server tools produce a non-token cost that no headline rate negotiation touched, and on an agent built around web search it can be substantial. Ask for the tool-unit rates and the monthly tool ceiling, separately from the token commitment.

Term 4: the pass-through clause

This is the clause that matters most. When two labs cut prices on the same afternoon, a rate card locked six months earlier becomes the expensive option. Structure the agreement so provider price revisions pass through, or trigger a review window. Without it, every price cut by a competitor raises your cost until renewal.

Bottom line

Committed floor, cache rate, tool ceiling, pass-through clause. Those four terms decide more of the bill than which model you chose, and they are the terms to negotiate before the renewal date, not after.

Related

← Back to llmcfo.com