Run budgets

A run budget caps what one run costs. A run is a single piece of agent work: one chat turn, or one assignment execution. The budget bounds it along four dimensions at once — money, tokens, tool calls and wall-clock time — and stops the run when any of them is reached.

Budgets exist because the natural limits are the wrong shape. Max iterations counts tool-calling rounds, and a single round can fan out into a dozen parallel tool calls. A model change can multiply the cost of the same work without changing anything you configured. A budget measures what you actually care about instead.

Where a budget is configured

There are two places, from broadest to narrowest.

The platform, set by an administrator in the deployment's environment. This is the default every run inherits and, just as importantly, the ceiling nobody else can raise. See Agent run budget for the settings themselves.

One agent, on its Budget tab. This is where you say "this research agent may spend more than the platform default" or "this triage bot must never spend more than a few cents". It applies wherever that agent runs.

The agent's Budget tab

Open an agent and choose Budget. Each of the four limits is a single field, and an empty field means "inherit", not "zero". The placeholder tells you what it would inherit, so you can see the effective budget without saving anything.

That distinction is the whole point of leaving fields empty. A budget that only pins the cost keeps tracking the platform's token, tool-call and time limits — including any later change an administrator makes to them. Typing the inherited numbers in to "be explicit" freezes a copy that silently stops following the platform.

Use the Inherit link next to a field to hand a dimension back.

Who may raise a limit

Anyone who can edit an agent may lower any limit. Spending less is never a privilege.

Raising a limit above the platform budget requires an administrator. That is what the platform setting is for: it is the budget a user cannot exceed on their own. If you are not an admin, the tab says so up front, and a field you push above the ceiling is flagged with the platform's value before you attempt to save — you will not discover it as a rejected save.

If your agent genuinely needs a larger budget than the platform allows, ask an administrator to set it. Once they have, the value stays: you can keep editing everything else about the agent without the budget being questioned again, and without being able to raise it further yourself.

Three settings that need no admin

The final-answer reserve, the cheaper-model failsafe and the choice of fallback model are not gated, because none of them can increase what a run spends. The reserve only moves the in-run limit down, and the failsafe keeps the same cost ceiling — it just buys more work under it.

The final-answer reserve

A run that spends its entire budget on work has nothing left to pay for the call that turns that work into an answer, so it would return a raw digest of tool output instead of a written result.

The reserve prevents that: a slice of the cost and token ceilings is held back for the closing answer. Work is gated at the remainder; the final answer may spend the full budget. The platform default is 10%, which comfortably covers one closing call.

Finishing on a cheaper model instead of stopping

A cost ceiling of, say, $0.50 on an expensive model usually means "don't spend more than $0.50 at this quality" — not "abandon the task". The same $0.50 buys several times more work on a cheaper model.

Turn on Finish on a cheaper model and a run that reaches its cost ceiling switches to a cheaper model and carries on instead of stopping. The rest of the run, including its closing answer, is charged at that model's rates, so the ceiling still holds. The agent is told the switch happened, and the run's trace records it, so cost attribution stays readable per call.

Deliberately bounded:

Once per run. Never a ladder that limps along on progressively worse models.
Cost only. Tokens, tool calls and time are model-independent — a cheaper model buys no extra wall clock.

Fallback model chooses which model that is. Left on Platform fallback model, it uses the one configured in Settings > AI Models > Fallback Model — the same model the platform already falls back to when a provider fails, so there is one cheaper model to reason about rather than two.

The switch is refused — and the run stops at its ceiling instead — when it could not be made safely: when no pricing is known for the fallback (with no price there is no way to know the next call is actually cheaper), when the fallback's context window cannot hold the agent's system prompt, or when switching would silently drop a structured-output guarantee the run depends on. In each case stopping is the honest outcome; failing mid-run is not.

How the layers resolve

Nearest wins, dimension by dimension. For any agent step, each limit is taken from the first layer that pins it:

the agent's own Budget tab,
the platform.

Per-dimension is what makes partial budgets safe. A tab that pins only the cost does not unbound the token limit — the platform's still applies. Worked through:

	Cost	Tokens
Agent's Budget tab	$2.00	(not set)
Platform	$5.00	1,000,000
Effective	$2.00	1,000,000

A budget you did not type yourself

Who may raise a limit governs every way a budget can reach your account, not just a value you type into the tab. Which of two things happens depends on whether you chose the number:

you typed it — editing an agent: the save is refused, naming the dimension and the platform's value, so you know exactly what to change;
you copied it — duplicating an agent, importing one, restoring an older version: the action goes through and any limit above the platform ceiling arrives lowered to that ceiling. Blocking a duplicate over a number somebody else typed would get in the way of the ordinary "copy the company agent and adapt it" workflow, and lowering it prevents the same escalation.

Either way you cannot end up holding a budget an administrator did not grant. An admin's copy keeps its values, since they could have typed them.

What happens when a budget runs out

The run stops making model calls and writes its final answer from the work it completed, using the reserve. The answer says which limit was reached, so the outcome is a shorter result rather than a failure with nothing to show.

Time is the exception in one respect: the duration limit is folded into the run's existing deadline, and whichever bound is tighter applies.

When a budget does not apply

Cost needs pricing. The cost ceiling is enforced from the platform's model price table, which is seeded for known models. If no price is known for the model a run uses, cost is not enforced for that run — a warning is logged and the token, tool-call and time limits still apply. Missing pricing data must never be able to stop every run on the platform.

Platform limits bound autonomous work. The platform's own settings apply to assignment runs. A budget you configure on an agent applies wherever that agent runs, chat included — configuring a budget is an explicit act, so it takes effect everywhere.

A budget smaller than one call

A limit below the cost or tokens of a single call to the configured model makes every run stop immediately and return a summary of tool output. The platform logs an explicit warning at the start of such a run. Raise the budget, or lower the model's maximum output tokens.

Next step

Read Agentic flow for how agents hand work to each other, or Agents for the other tabs on an agent's detail page.