Most enterprises already run FinOps for cloud infrastructure: spend is attributed to owners, reviewed on a cadence, and judged against the value it produces. LLM spend belongs inside that same practice, with one twist that makes it more tractable, not less. Because Claude bills per token and reports token usage on every response, cost attribution can be more precise than for almost any other cloud service, down to the individual request, if you set up the plumbing.
Attribution: every token has an owner
The foundation is being able to slice spend by team, application, feature, and environment. Cloud-native tools get you partway: accounts, projects, subscriptions, tags, and labels attribute the invoice to organizational units, the same way they do for compute and storage. The finer grain comes from your own telemetry: log input and output token counts and the model ID for every request, stamped with feature and tenant metadata, and multiply by your prices in a nightly job. Teams that route all LLM traffic through an internal gateway get consistent stamping for free, which is a major reason the gateway pattern pays for itself. The test of adequate attribution is simple: when spend jumps, can you name the feature and the owning team within the hour?
Unit economics: cost per business outcome
Raw spend rising is not automatically bad news; it may mean the product is succeeding. The number that carries meaning is cost per unit of work: per ticket resolved, per document processed, per report generated. As a worked example, assume a support assistant averages 2,000 input tokens and 500 output tokens per ticket on Claude Sonnet 5 at list price ($3.00 per million input, $15.00 per million output). That is (2,000 × $3.00 + 500 × $15.00) ÷ 1,000,000 = $0.0135 per ticket, roughly a cent and a half. Set that against what a ticket costs to handle by other means and the conversation changes from "the AI bill grew" to "we handled 40% more tickets at just over a cent each." Unit costs also make optimization legible: prompt caching, model tiering, and prompt trimming each show up as a measurable drop in cost per unit, which is how you demonstrate that engineering effort paid off.
Showback before chargeback
Once attribution works, make it visible. Showback (publishing each team's spend without moving money) is the right starting point: it creates cost awareness and peer accountability without the accounting friction of internal billing, and it surfaces surprises while they are still cheap. Publish a simple monthly view per team: total spend, spend by feature, unit costs and their trend, and model-tier mix. Most organizations find showback alone changes behavior, because engineers who can see that their feature costs five times the neighboring one per unit will investigate without being told. Graduate to chargeback only if your organization already does it for other cloud spend; the mechanics are identical once attribution is trustworthy.
The operating rhythm
FinOps is a loop, not a dashboard. A workable cadence for LLM spend: daily automated anomaly checks on token volume per feature (machines watch this, not people); a monthly review where each owning team looks at unit costs, tier mix, and cache effectiveness, and picks at most one optimization to pursue; and a quarterly view for leadership connecting spend to business outcomes and informing commitment decisions. Keep the optimization backlog honest by pricing the engineering time too: a change that saves less than it costs to build and maintain is not an optimization. And when committed-use discounts enter the picture, keep measuring at list-price-equivalent internally, so discount negotiations and engineering efficiency remain two separately visible wins rather than one blurred number.
Where to go next
The alerting plumbing underneath this practice is covered in Setting Up Cost Alerts and Budgets, and the value side of the ledger in Measuring ROI on LLM Projects. For the cheapest levers your teams will find, see Prompt Caching and Model Tiering.