AI usage on your workspace dashboard (/app/<workspace>) shows what AI costs and uses, for every AI call in the workspace, down to each model. You need permission to view runs to see it.
What it shows
- Filters: Period (Today, Yesterday, Last 7, 30 or 90 days), Metric (Tokens, Cost, Calls), Source, and Group by (model, campaign, flow, key, source).
- Tiles: Estimated cost, Calls, Failed, Input, Cached (with its share of input), Output and Reasoning tokens.
- Per day: stacked bars by what you group on. The seven biggest keep their color, the rest are Other. Today and Yesterday chart the last 14 days.
- Tables: each model, source, key, flow and campaign (a flow called by another counts on the flow that started the run), with calls, failed, tokens, hit rate, average time, cost and change.
How to read the change
Each number compares with the same number of days before the period.
| Period | Compared with |
|---|---|
| Today | Nothing: yesterday's value is shown beside it, because the day isn't over |
| Yesterday | The day before |
| Last 7, 30 or 90 days | The days before them (these periods include today, so they read a little low until the day ends) |
New means nothing in the period before, and – means nothing in either.
What counts
| Source | The calls |
|---|---|
| Flow runs | AI nodes and AI agents in real runs, the no-repeats check, and knowledge a flow adds or searches |
| Test runs | The same in test runs: they cost real money too |
| Studio agent | The agent's replies, summaries and conversation names |
| Editor completions | Autocomplete in Studio's editors |
| Knowledge embeddings | Adding to and searching a knowledge base on its Dataset page |
| Post generation | Dataset › Posts › Generate with AI |
| Key tests | Testing an AI key |
Avg. time is how long an answered call took on average (failed calls aside). Reasoning tokens are part of Output, and Cached tokens are part of Input, so don't add them up.
How the cost is estimated
The price on your AI key comes first. A key with no price, or another model than the key's, uses Mooflair's price list. A model on neither shows tokens only (a custom URL, for example).
The dashboard adds both together: the Estimated cost tile says how much of it is estimated, and a dotted cost in a table shows it on hover. The cost is an estimate, not your provider's bill. Your provider's own billing page is the real number.
What it can't see
- AI called straight from an HTTP or Code node.
- A call you stopped: it may still be billed.
- Long-context, batch and off-peak prices. DeepSeek's estimate uses its peak price.
- Cached tokens from a service that doesn't report them. Anthropic's OpenAI-compatible API doesn't cache at all, so it shows 0.
FAQ
Why doesn't the cost match my provider's bill?
It's an estimate from list prices, and some prices depend on the time of day or the prompt's size. Set your own price on the AI key to change it, and use your provider's bill as the real number.
Why are there no numbers from before?
Counting started when this feature went live. Older calls can't be added, because their logs don't say which model they used.
Why is it higher than the AI column in Insights?
Insights counts AI nodes in flow runs at your key's price. AI usage also counts test runs, the Studio agent, embeddings and completions, and estimates a cost where a key has no price.
Do test runs count?
Yes, because they cost real money. Pick Flow runs under Source to see real runs only.