What makes FinOps for AI different from cloud FinOps?#
The unit. Cloud FinOps meters instance-hours, which are provisioned in advance and consumed steadily. AI spend is metered in tokens, counted separately for what goes into the model and what comes out, with output priced higher than input. Consumption is driven by how the work is shaped rather than by how much capacity is running, so the same team can vary its bill by an order of magnitude without changing anything it has provisioned.
Everything loaded into a session is charged again on every turn it stays in context, including tool definitions and retrieved files. That makes context the first place to look, and it is one of the reasons context engineering is a cost discipline as well as a quality one.
FinOps for AI: why does the bill rise when token prices fall?#
Because the price per token and the number of tokens per task move in opposite directions. The Stanford AI Index 2025 reports that the cost of querying a model at GPT-3.5 level fell more than 280-fold between November 2022 and October 2024. Researchers at MIT FutureTech found the same pattern across models: the price of a fixed level of performance falls by roughly five to ten times a year, while the cost of reaching the frontier rises by three to eighteen times a year. Most engineering teams move to the newest model soon after it ships, so the falling price is one they seldom pay.
Reasoning models add to this. Epoch AI estimates that they produce around eight times more tokens than non-reasoning models on comparable work, and extended thinking settings can multiply output by ten to twenty times. Output is the expensive side of the meter, so a model that is cheaper per token can still produce a larger invoice per task. Agents multiply it again, because one request becomes dozens of calls and each call carries the context with it.
That makes unit price the wrong number to forecast with. Forecast tokens per completed task and the model mix behind them, and review both every month. The FinOps Foundation revised its definition of FinOps in March 2026 to cover technology spend beyond cloud, and its 2026 Framework treats AI as one of five technology categories. Billing data has not caught up: FOCUS 1.4, ratified in June 2026, has no columns for tokens or models, which are scoped for version 1.5. Until then, teams have to normalise each provider's usage exports themselves.
What should FinOps for AI measure spend against?#
Value, not volume. Tokens consumed, agent runs and pull requests are diagnostics: they tell you what the system did, not whether it was worth doing. The measure that connects spend to outcome is human attention per outcome, which counts the human effort one useful result consumes — reviewing, briefing, catching up and fixing — alongside the machine cost of producing it.
An organisation can cut its token bill and raise its total cost, if the cheaper output needs more review. Reading the two together is what stops that trade being made by accident.