AI spend
What your AI work costs, broken down by task, by agent, by person and by model. Workspace settings, then AI spend.
The work runs on your own model key, so this is a record of what you spent at your provider rather than a bill from Ithura. See Setting up AI.
The page is for workspace admins. Anyone else who reaches the URL is told so rather than shown an empty report.
There is no plan gate on reading it. Getting numbers into it needs an agent with an API token, and tokens need Pro.
What it shows
Pick a window of 7, 30 or 90 days. Days are counted in UTC, so a report does not shift under you depending on who opens it.
At the top, total tokens and total cost. Then a bar per day, and four breakdowns:
| Breakdown | The question it answers |
|---|---|
| By task | Which work is expensive |
| By agent | Which agent is spending it |
| By person | Who is directing the spend |
| By model | Whether the model choice is the cost |
Each list shows the top rows by cost, so a long tail does not bury the thing you need to see.
When the total is a lower bound
Not every provider reports a dollar figure. When some usage arrived with tokens but no cost, the page says the total is a lower bound and keeps the tokens, rather than quietly dropping them or inventing a price. A cost that is genuinely unknown reads n/a.
This is the same honesty as an interrupted run: see Stopping and steering a run.
Per-task spend
A task that cost something shows it on the task itself, with the split per agent on hover. A task with no AI usage shows nothing at all, so an ordinary task does not carry an empty cost line.
Reading a task's spend needs Member access and membership of its project. Guests do not see it.
How the numbers get there
An agent reports its own usage. When a run finishes, it sends the model, the token counts and the cost to the workspace's agent usage endpoint, using an API token bound to that agent.
Reports carry counts and money, never prompts or outputs. What the agent read and what it wrote are not part of a spend report, by design: a cost report is not a transcript.
A report can carry its own id, which makes it safe to retry. A repeated report is recognised and counted once rather than doubling the bill.
If you run your own agent, this is the one call to make for it to appear here. See Managing your agents and Connect a coding agent.
Troubleshooting
| What you see | What it means |
|---|---|
| "Only workspace admins can see AI spend." | The workspace report is admin-only. A task's own spend line needs only Member. |
| "No agent usage yet." | Nothing has reported. An agent run alone does not report cost; the agent has to send it. |
Cost says n/a but tokens are there | The provider did not report a price. The tokens are real. |
| "Some usage was reported without a cost" | Part of the window has tokens and no price, so the total is a floor rather than the figure. |
| "Only an agent API token can report usage" | The token used to report is not bound to an agent. A personal token cannot report spend. |
| "This token is not bound to an agent in this workspace" | The token belongs to an agent in another workspace. |
| A model reads "unknown" | The report arrived without a model name. |
| A task shows no spend line | Nothing was reported for it. |
Related
- Agent budgets: capping the spend rather than reading it.
- Managing your agents: registering an agent and its token.
- REST API: tokens and scopes.