Use AI Usage to see which features use AI, how much their recorded requests cost, and whether spending is approaching your budget.
This guide covers sites that use their own AI provider API keys. Cost figures are estimates in US dollars, not a provider invoice. Check your provider account for the amount you are actually billed.
Open your usage report
Open Tools → Observability → AI Usage. Choose This month, Last 24h, Last 7d, Last 30d, or All to change the reporting period.
If AI setup is incomplete, follow the setup notice first. Some site configurations handle AI usage through Cedros Providers instead and do not show this tab. If the tab is missing or you cannot edit a budget, ask your administrator to check your site's setup and your feature access.
Reading the report does not send an AI request. You do not need to ask the assistant a test question to check existing usage.
Understand the headline figures
- Estimated cost totals the recorded requests that have a known cost in the selected period. A + means some costs are missing, so the displayed amount is a minimum. A dash means there is no priced usage to total; it does not mean the requests were free.
- Projected this month estimates a full month's spend from usage so far this month. It remains a monthly projection when you select another reporting period. A busy day early in the month can make the projection rise sharply.
- Failed calls includes failed and cancelled requests. Check the individual records to understand what happened; a failed request is not automatically a free request.
The report uses UTC for its date range and daily chart. Use the Cost and Tokens views to compare activity over time. Tokens measure the amount of text processed, not a fixed price: different models can charge different amounts for the same number of tokens.
The report covers usage recorded by this site. It is not a statement of everything charged to a provider account or to an external AI client.
Find what is using the budget
Under Where it went, start with Feature to see which parts of your site account for the usage. Switch to Model or Provider when you want to compare the services behind those features. Routine appears when the report contains routine usage.
Select a Feature, Model, or Provider row to open its individual calls. Use Failed to narrow that list and inspect the error information shown for a request.
The individual-call list can be incomplete on a busy site. It contains the report's most expensive calls, up to 250, while the headline figures and breakdown totals include all recorded calls in the selected period. Read any partial-history notice before treating the visible list as a complete record.
For an unexpected increase, compare a short period with This month, then check the largest feature and model rows. Look for repeated attempts or background activity before assuming a single assistant conversation caused the increase.
Set a spending budget
You need permission to change site settings to save a budget. Limits apply to AI features using the site's own API keys.
- Select Set a monthly budget, or Edit budget if one already exists.
- Under Monthly cap, all AI, choose an amount or enter a positive dollar amount in Or set your own. A blank field means no cap.
- Turn on Enforce spending limits. Saved caps do not block requests while enforcement is paused.
- If needed, set Cap on any single request. Open Advanced to set separate Per-provider caps.
- Review Block unpriced models. Turn it on if requests with no known price should be refused while limits are enforced. Leave it off only if you are comfortable with those requests being allowed without contributing a cost to the budget totals.
- Select Save limits and check the budget summary. Cancel discards the changes you made in the dialog.
Monthly caps use the current calendar month and reset on the first day of the next month. Selecting Last 7d or another report period does not change the budget's monthly period.
Cedros checks estimated costs before starting requests. These limits can refuse new work, but they do not guarantee that a provider's final bill will stay below the exact cap. Final output costs and work already in progress can differ from the estimate. Review any spending controls offered by your provider as well.
Choose alerts and review missing prices
In Spending limits, enable Alert me before I hit a cap, then open Advanced → Alert thresholds. Select 50% or 75% if you want a warning before reaching a cap; an alert at 100% is a notice that the threshold has already been reached. Use Manage notifications to review notification delivery, then save the limits.
If costs are missing, use Review model prices where offered, or open Advanced → Model prices in the budget dialog. Model prices are in USD per 1M tokens, with separate input and output amounts. Enter prices only after checking them with your provider; leave overrides blank to use the available catalog or default prices.
Adding a token price does not reconstruct unreported usage or make every older cost known. Continue to treat a total marked + as incomplete. Blocking unpriced models can also stop requests that do not have a supported price estimate, so review the affected services before enabling it.
When a limit stops a request
Read the limit message and check the budget summary. A site-wide monthly cap can affect multiple features; a provider cap applies to that provider, and a single-request cap can reject a request even when room remains in the monthly budget.
For a monthly cap, you can wait for the next month or have an authorized administrator review and adjust the limit. For a single-request cap, consider a smaller task or a different model before raising it. For an unpriced-model error, review its price or the blocking setting. Repeating the same request does not resolve the limit.
If the error is unrelated to a budget, inspect the failed call and check the provider setup. See Choosing AI providers and models. When asking for help, include the reporting period, feature, provider/model, and exact error. Keep API keys out of messages and screenshots.