Guide
What Is FinOps for AI? What to Measure and Where to Start
How FinOps for AI differs from cloud FinOps, the metrics that matter, a worked example at real prices, and the first steps to control AI spend.
- Published
- Read time
- 5 min
On this page
FinOps for AI applies the FinOps Framework to AI spend: tokens, GPUs, provisioned capacity, and tools such as Claude Code and Cursor, so each team knows what its AI costs and what it returns. The method matches cloud FinOps, because cost is still price times quantity. The billing unit, the allocation data, and the pace of price changes differ. Start by bringing your bills together, giving each team its own key or project, and tracking one unit cost per product.
We read the FinOps Foundation's material in November 2026. If you came here to pick a tool, go straight to the best FinOps tools for AI.
What is FinOps for AI?
The FinOps Foundation describes FinOps for AI as managing the cost and value of AI. The similar name misleads: AI for FinOps is a separate subject, using AI to do FinOps work, which we cover in what agentic FinOps is.
The work became routine in two years. In the State of FinOps 2026, 98% of FinOps teams say they manage AI spend, up from 31% in 2024, and AI cost management ranks as the skill those teams most need to build.
How is FinOps for AI different from cloud FinOps?
Tags, allocation, and commitment discounts still apply, and part of your AI spend already lands on the cloud bill. These points differ:
| Point | Cloud | AI |
|---|---|---|
| Billing unit | Instance hours, GB | Tokens, API calls, provisioned throughput units, GPUs |
| Who spends | Mostly engineering | Also product, marketing, sales, and leadership |
| Allocation | Tags on resources | Keys, projects, and workspaces, because new SKUs and direct bills often arrive untagged |
| Prices | Change slowly | New models and frequent repricing |
| Anomalies | A forgotten resource left running | A runaway job, an agent loop, a prompt change that doubles token use |
| Optimization | The smallest resource that holds the load | The cheapest model that meets the quality bar |
AI column adapted from the FinOps Foundation's FinOps for AI Overview, read in November 2026. Cloud column: our own summary.
What should you measure in FinOps for AI?
Total spend won't tell you whether AI earns its cost. These metrics, all from the FinOps Foundation's list, answer different questions:
| Metric | Formula | What it answers |
|---|---|---|
| Cost per million tokens | Token cost / (tokens / 1 million) | What you pay per unit, comparable across models |
| Cost per inference | Inference cost / requests | What each answer costs |
| Cost per successful outcome | Cost of all runs / runs that meet the quality bar | What an answer that worked costs |
| Cache hit rate | Cache reads / (reads + writes + uncached input) | How much of the prompt you avoided paying for in full |
| Model fit | Requests on overqualified models / total requests | How much traffic uses a model too expensive for the task |
| Resource utilization efficiency | Used capacity / provisioned capacity | How much of your provisioned capacity gets used |
| Anomaly cost percentage | Anomalous spend / total AI spend | How much escapes the plan |
Source: FinOps Foundation, FinOps for AI Overview, read in November 2026.
A worked example with real prices
Say a support assistant handles 1 million requests a month, each with 1,500 input tokens and 300 output tokens. On Anthropic's price list, Claude Sonnet 5.5 costs $2 per million input tokens and $10 per million output tokens, and Claude Haiku 5.5 costs $0.10 and $0.50. The resolution rates below are hypothetical: yours come from your own evaluations.
| Setup | Cost per month | Per request | Resolved | Per resolved |
|---|---|---|---|---|
| Sonnet 5.5 only (90% resolve) | $6,000 | $0.0060 | 900,000 | $0.0067 |
| Haiku 5.5 only (70% resolve) | $300 | $0.0003 | 700,000 | $0.0004 |
| Haiku first, Sonnet for the other 30% (90% of those resolve) | $2,100 | $0.0021 | 970,000 | $0.0022 |
Prices per million tokens, standard tier, with no caching and no batch. Haiku 5.5 prices apply to prompts up to 100,000 tokens.
On cost per request, Haiku wins by a wide margin and leaves 300,000 requests unresolved. On cost per resolved request, Haiku alone is still cheapest, and the third row shows what that figure hides: it resolves more than either model alone and costs about a third of Sonnet. That's why you read an outcome metric beside the volume resolved, and never the price per token alone. For each provider's prices, read what an LLM API costs, and to cut the bill, LLM cost optimization.
How do you split AI costs by team?
Give each team its own billing unit: a project at OpenAI, a workspace at Anthropic, a tagged inference profile on Bedrock. Then keep one list of team names and point each unit at a name. How to track AI costs by team has the steps for each provider.
With virtual tags, LevelFour builds that map without touching your resources: team=Payments can claim an AWS account, an OpenAI project, and an Anthropic workspace.
Where do you start?
The FinOps Foundation advises building up in stages, crawl, walk, run. In practice, that can look like this:
- Crawl: bring your AI bills into one place, give each team a key or project, and review spend once a month.
- Walk: pick one unit cost per product, turn on budgets and anomaly alerts, and separate experiments from production.
- Run: evaluate quality, route between models, use caching and batch, and only then consider provisioned capacity.
Common mistakes
- Comparing models on price per token alone. The example above shows why the math needs quality in it.
- Sharing one key across teams. The bill then shows what you spent and hides who spent it.
- Buying provisioned capacity too early. It suits latency-sensitive production work, and it wastes money while usage still swings.
- Running agents with no spend limit. An agent loop is one of the anomalies the FinOps Foundation names.
- Forgetting coding tools. Claude Code and Cursor count as AI spend, and each tool bills its own way.
How LevelFour helps
LevelFour reads your Cloud, SaaS, and AI bills in one place, with OpenAI, Anthropic, Cursor, and more, splits spend by team with virtual tags, and ships the savings through to the bill. See what LevelFour can save you in 5 minutes, with read-only access.
Frequently asked questions
What is the best FinOps tool for AI?
LevelFour to bring Cloud, SaaS, and AI together and ship savings through to the bill, Vantage and Finout for multi-cloud visibility with AI cost, CloudZero for unit cost, and Zylo for SaaS and AI licenses and contracts. The right pick depends on where your spend sits. See the full comparison.
What is the difference between FinOps for AI and AI for FinOps?
FinOps for AI manages the cost and value of AI. AI for FinOps uses AI to do FinOps work, such as investigating an anomaly or preparing a fix.
What is the most important FinOps for AI metric?
Cost per successful outcome, because it divides spend by the runs that met your quality bar. Cost per token and per request can't show whether the answer worked.
How many FinOps teams manage AI spend?
98%, according to the State of FinOps 2026, up from 31% in 2024.
Try LevelFour free for 14 days
See what it saves on your cloud, SaaS, and AI bills. No credit card required.
Start free trial