AI spend audit: model costs by provider, model and project
The AI spend audit of Cureman shows what an organisation pays model providers and where the money goes. It reads the providers' usage reports through read-only connectors, prices every token against a model price list whose rows say whether a person has verified them, and breaks spend down by provider, model, project and day. For each model it also computes what the identical tokens would cost on a cheaper model of the same provider, as arithmetic and without a claim about quality.
A provider's bill for model usage is one total a month. The audit takes that total apart to the token: which model, which project, which day, and what the same tokens would have cost on another model of the same provider.
What it reads
The module reads the usage reports that model providers offer through their admin APIs, with read-only connectors for Anthropic and OpenAI. A report gives, per day and model, the input, output, cache-read and cache-write tokens and the number of requests, together with the project or workspace the provider recorded and whether the usage ran as a batch. Connectors only read, and a connector manifest is marked beta until it has been run against a live account.
For a provider without a connected report, the audit uses the requests metered by the AI gateway, where that module is booked, grouped by the tag the calling service sent. A provider is never counted from both sources.
The model price list
Every cost on these pages is arithmetic: token counts times a price per million tokens, separately for input, output, cache reads and cache writes. The prices stand in a model price list that is visible in the product, in US dollars, because that is how providers price. Each row carries a status. A price that shipped with the product or was drafted by research is unverified until a person has checked it against the provider's own list, and it is labelled as unverified wherever it is used.
Where the provider's report carries a cost of its own, that is the spend shown, and the page says when it differs from token counts times list price, for example because of a discount or batch pricing. A model that is missing from the list is shown as having no price, and nothing is computed for it. How the parts of a token price fit together is explained in LLM token costs explained.
What it shows
| View | What it answers |
|---|---|
| Spend by provider and day | How the total moved over the chosen period, compared with the period before where data exists |
| Spend by model | Which models carry the spend, with tokens, requests and the status of the price used |
| Same tokens, other model | What the identical token counts would cost at the list price of a cheaper model from the same provider |
| Spend by project | Which project or workspace, as the provider reports it, the usage belongs to |
| Cache-read share | How much of the input was served from the provider's prompt cache |
The comparison with another model is a repricing of the same tokens with list prices on both sides. It says nothing about whether the cheaper model could do the work.
What it finds
The audit raises two kinds of finding, and both size a question without answering it.
- Model mix. When a large share of spend goes to a provider's most expensive class of model, the finding shows what the same tokens would cost on the next cheaper one. That figure is a ceiling: it is shown with its calculation and is not counted as an identified saving, because a report does not say which requests were simple.
- Prompt caching. When requests carry a lot of input and almost none of it is read from cache, the finding says so. It states no amount, because a report does not show which prompts repeat.
Both findings carry low confidence on purpose. Measuring the answer is the job of the gateway in observe mode, which classifies each request and changes nothing.
From finding to a decision
Each finding has an action plan with one prepared step: booking the gateway module, so that traffic can be observed. The step waits in the list of decisions until a person approves it. Nothing is booked, switched or sent by the audit itself.
What it needs and what it measures
The module requires no other module. It needs a connection to at least one provider's usage report, or gateway traffic. It meters nothing for the monthly statement.
Related reading
Questions and answers
Does the audit need the gateway?
- No. The audit works from the usage reports a provider offers through its admin API and requires no other module. The gateway adds what a report cannot show, namely which individual requests were simple, which prompts repeat and who sent them. For a provider without a connected report, the audit uses the requests the gateway metered.
Where do the model prices come from?
- From a model price list in the product, in US dollars per million tokens, with separate prices for input, output, cache reads and cache writes. A price that shipped with the product or was drafted by research is unverified until a person has checked it against the provider's own price list, and it is labelled that way wherever it is used.
Does the audit say which model an application should use?
- No. It reprices the same token counts at the list price of a cheaper model from the same provider, which is arithmetic. Whether a cheaper model can do the work is not something a usage report can answer, so that difference is shown as a ceiling and is not counted as a saving.
Does the connector change anything at the provider?
- No. Connectors only read. The connector for a provider reads the usage report and nothing else is written or switched there. A connector manifest is marked beta until it has been run against a live account.