AI cost management
Tracking, allocating, and controlling what you spend on model APIs, AI tools, and GPUs.
AI cost management is the work of tracking, allocating and controlling what your organization spends on AI: model APIs, managed platforms such as Amazon Bedrock, AI tools such as Claude Code and Cursor, and the GPUs that train and serve models. Providers bill tokens, API calls and provisioned throughput, units the FinOps Foundation contrasts with instance hours. OpenAI and Anthropic price per million tokens. Provisioned throughput on Amazon Bedrock and Microsoft Foundry bills by the hour, whether or not you send requests. The core tasks are daily visibility by model and team, spend limits on each project or workspace, and unit metrics such as cost per successful result. The FinOps Foundation warns of runaway usage without monitoring and limits, and its State of FinOps 2026 reports that 98% of respondents now manage AI spend. LevelFour reads OpenAI, Anthropic and other AI provider bills next to cloud and SaaS spend, and one virtual tag can claim an AWS account, an OpenAI project and an Anthropic workspace.
Frequently asked questions
What is the difference between AI cost management and FinOps for AI?
AI cost management covers the daily operations: tracking, allocating and limiting AI spend. FinOps for AI is the wider practice, which adds value measurement, forecasting and shared ownership across engineering, finance and product. A team that sets a spend limit on an OpenAI project is managing cost. A team that compares that project's cost per resolved ticket with the old process is doing FinOps for AI. The FinOps Foundation's State of FinOps 2026 calls FinOps for AI the top forward-looking priority for FinOps teams.
How do AI providers charge for usage?
Model APIs charge by the token, with separate rates for input, cached input and output. OpenAI's and Anthropic's price lists both quote rates per million tokens and offer lower batch rates. Reserved capacity follows another model: Amazon Bedrock Provisioned Throughput and Microsoft Foundry provisioned deployments bill by the hour for a fixed block of capacity. The FinOps Foundation warns that billed tokens can exceed what users type, since applications add prompts and history. Team plans for tools such as Claude add seats and usage credits on top.
How do you track AI costs by team?
Give each team its own billing unit at each provider, then read the provider's cost report by that unit. OpenAI's Costs API groups spend by project. Anthropic's Cost API groups spend by workspace, and its usage API filters by API key. On Amazon Bedrock, attach cost allocation tags to an application inference profile and call the model through it. Bring those reports together with your cloud bill, and retire shared keys, which hide who spent what.
How do you set spending limits on AI APIs?
Set limits at the provider, per project or workspace, and add alerts below them. OpenAI separates spend alerts, which notify you while API traffic continues, from hard limits, which fail requests with a 429 error, and a limit can cover an organization or one project. Anthropic's Console caps monthly spend per workspace and sets alerts. The FinOps Foundation also suggests approved model lists, quotas and rate limits, with lighter rules for pilots. Test what your application does on a 429, because users see that failure.
How do you reduce AI costs?
Start with model choice: route each task to the smallest model that handles it well. The FinOps Foundation tracks this as model fit, the share of requests sent to overqualified models. Reuse repeated prompts through caching: OpenAI enables prompt caching by default on supported models, and Anthropic's price list bills cache hits at a tenth of the base input price or less. Send work that can wait to batch, which Anthropic's docs discount by 50 percent. On provisioned capacity, watch utilization, since the Foundation notes that low use can cost more per token than on-demand.
Why are AI costs hard to forecast?
AI spend follows usage inside your application, and that usage moves fast. A longer prompt, a bigger model or an agent loop changes the bill with no new infrastructure. The FinOps Foundation says AI forecasts are less predictable than cloud forecasts, above all in the early months, and advises shorter windows and more frequent revisions. Its FinOps for AI topic adds that agentic AI makes cost non-deterministic. Forecast by driver, as requests times tokens per request times price, and revisit the model when a feature changes.
What should you measure in AI cost management?
Track spend by model, team and API key each day, then add unit metrics. The FinOps Foundation's FinOps for AI overview lists unit costs per request, per token and per successful result, along with GPU utilization, cache hit rate and anomaly share of spend. Anthropic's usage API reports uncached input, cached input, cache creation and output tokens, which gives you the cache hit rate. Tie each metric to a budget and an owner, so a rising number reaches someone who can change the model or the prompt.