Budgets and receipts
See what every turn cost, broken down by model, track spend by day, model and workspace, and set spending limits.
Cusp shows what your agents cost as they work, not at the end of the month. Every turn gets a receipt, the usage page adds those up over time, and budgets warn you before spending gets away from you.
Costs are calculated from each provider's published prices and the token counts the provider reports. Your provider's invoice is the final word; Cusp's numbers are meant to match it closely, not replace it.
Receipts
When a turn finishes, the transcript shows a summary line with how long it took, how many tools ran, the tokens produced and the cost. Click the cost to open the turn's receipt.
A receipt breaks the turn down by model, including the small model Cusp uses for titles and summaries, subagents, and any helper calls made during the turn:
- Input, output and cached tokens for each model.
- Cost for each model and in total.
- Cache saved: what prompt caching saved you on this turn, shown when it's at least $0.01.
- The outcome: completed, interrupted, failed, or Paused at budget.
The Receipt pane (in the pane toggles next to Changes, Terminal and Files) lists every turn of the session, newest first. Expand a turn for its per-model split.
How amounts are shown
| You see | Meaning |
|---|---|
$0.42 | The cost of the turn or model |
<$0.01 | Less than a cent |
local | A model on your machine (Ollama, LM Studio, vLLM, llama.cpp). It costs nothing per token |
— | Cusp has no price for this model. Totals leave it out, and the tooltip says so |
Spend also appears in the session header and in the title bar. Click the title-bar amount to open the usage page.
The usage page
Open Usage from the command palette, the title-bar spend, or the Receipt pane.
- Period: Today, This week (the default) or This month.
- Measure: cost or tokens.
- Totals: spent, cache saved, tokens, and unpriced tokens (usage on models without a price).
- Spend by day, plus tables By model, By workspace and Most expensive sessions.
- Export the report as CSV.
The usage page is built from the session history on this computer. Sessions on other devices aren't included.
Budgets
Set budgets in Settings → Budgets. Each budget has:
| Field | Options |
|---|---|
| Applies to | Each day, all projects · Each month, all projects · Each day, one workspace · Each month, one workspace · Each turn |
| Amount | A limit in US dollars |
| At the limit | Hard cap or Warn only |
| When reached (hard caps) | Pause and ask · Stop · Switch to a cheaper model |
The form previews the rule in plain language before you save, for example "Pause and ask when today's spend across all projects reaches $25." Each budget shows a bar with what's spent and when it resets, and can be switched off without deleting it.
Cusp warns you with a notification when spending crosses 50 % and 80 % of a budget, and again at 100 %.
Budgets in the app warn but don't stop work yet. When a budget reaches 100 %, Cusp tells you, but running sessions keep going: stop them or raise the cap yourself. Pausing a session at the cap, with the choice to raise it, finish the turn, switch to a cheaper model or stop, is coming soon.
Hard limits that work today
Two kinds of limit are enforced now:
Pass maxBudgetUsd (TypeScript) or max_budget_usd (Python) to a query. When the query's spend reaches the limit, Cusp interrupts the run and the result has the subtype error_max_budget_usd.
Available with the first public release.
The check runs as usage is reported, so a run can go slightly over the limit by the cost of the model call in progress. See Agent SDK.
Cloud spending
Cusp's managed gateway and its spending limits are coming soon. With your own provider keys, you pay your provider directly and Cusp adds nothing to your provider's bill.
Next
- Models and effort: pick cheaper models for routine work.
- Race: compare what different models cost on the same task.
Attention inbox
One list of everything your agents are waiting on — permissions, questions, plan approvals, failing checks and pull requests — with keyboard triage.
Review
Read what the agent changed in the Changes pane, revert or accept single hunks, send comments back to the agent, and run code reviews at four depths.