What an agent task costs

This page is about what the coding agent costs. For how credits work in general — the two currencies, where they come from and the order they are spent — see Credits, which is the canonical explanation.

What the agent spends

Model requests draw on AI credits. Tools the agent calls from the catalogue draw on general credits. A task can therefore be cheap in one and not the other, which is why the two balances are shown separately in the Account panel.

What makes one task cost more than another

Agent traffic has a shape that ordinary API calls do not: it resends a growing prefix on every turn. Input tokens, not output, are usually the larger half of the bill.

That is what the levers below all address.

Four things that lower it

Leave scout subagents on. Exploration moves to the cheap tier. It is most of an agent's token spend and almost none of its difficulty — finding a file is not the hard part. See Subagents.

Prefer a model that caches for long sessions. The prefix is resent every turn either way; with caching most of it is billed at a discount. On real traffic, agent requests run around 27% of input served from cache against 0% for one-shot calls. See Prompt caching.

Start a new task when the subject changes. A fresh context window is cheaper and sharper than a long one carrying unrelated history you are paying for on every turn.

Use Plan mode for large or unfamiliar changes. A wrong start costs more than the planning that would have avoided it.

What does not lower it

Auto-compact is not a cost control. It stops long tasks failing outright. Leave it on for that reason, not to save money.

Where to see what you spent

• Dashboard → Usage — spend by day, model and workspace.
• Dashboard → AI → Logs — individual requests: model, tokens, credits, cache share, latency, and which key served it.
• In the editor — turn on Settings → Chat → Show cost per request to see it inline.

Using your own key

With bring your own key you pay the vendor for tokens directly, and ToolsXpo charges a share of what the call would have cost instead of token credits. See How AI calls are billed for the general rule.

See also

• Credits — the two currencies, in full
• Prompt caching
• Agent and scout tiers