BUILDING Parts of this page describe what is not built yet, and each says so.
AI budgets and spend
Model spend is the money your organisation spends when an agent calls an AI model. It is counted in dollars (a trial is given $10.00, or $5.00 with a personal email address) and in nothing else, and it is capped: you set a budget, and once models are called each call will be paid for against it before the call is made, and a call that would take a budget past its cap will be refused before anything is sent. Only agents will stop at a cap. Queries, runs and everything else will carry on.
No model has been called on any plane, so every figure reads $0.00. The first call is built, for one task (explaining a validation problem), and is not running anywhere. Before that call, its cost at most is held against every budget that covers it, after it the cost is settled in dollars, and a call that would pass a cap is refused with no request sent to the provider. Everything below is how the budgets are set and read.
What a budget is
A budget is a cap in dollars for a UTC calendar month. There is one at each of these levels, and a call is held to every one that applies to it, so a project's cap never makes up for the organisation's:
| Budget | What it caps |
|---|---|
| The organisation's | Every model call in the organisation, in every project. |
| A project's | The calls made for one project. A project with no cap is held only to the budgets above it. |
| An agent's, in a project | What one of the four agents may spend in one project. |
| An action's | What one agent may spend on one kind of work in one project. |
| The shadow budget | What the organisation may spend on trying a new model release beside the current one. Its default is nothing, which keeps that trial off. |
On a team plan, no model call is made until an owner or an admin has set the organisation's monthly cap. A budget with no cap is shown as no cap and never as $0.00: the first means nothing has been set, the second is a cap of nothing, which stops every call that budget holds. The shadow budget is the one budget that starts as a cap of nothing, so it reads $0.00 until it is raised.
The trial’s grant
A trial has one grant for model spend, given once for the whole trial and not each month: $10.00, or $5.00 when the trial was started with a personal email address. It is a budget like the others and is read the same way, but it has no cap to set. When it is used up, agents will stop and nothing else will.
Setting a budget, lowering it and raising it
Budgets are set in AI settings, on the Budgets tab, by an owner or an admin. A cap can be lowered from anywhere: the console, or the command line with a credential that holds the right, and lowering it to $0.00 stops every call that budget holds. A budget is raised only in the console, with a confirmation from the last five minutes. A team organisation's first cap is a raise, so it is set in the console too.
A command-line credential cannot give that sign-in, so the command that lowers a budget reads the cap you hold first. A figure above it makes no change, sends nothing and says so. The control plane would refuse it anyway, and a refused change is a line in the audit record that tells nobody anything.
Held and settled
Before a model is called, the most the call could cost will be held against every budget that applies to it. When the call finishes, what it really cost will be settled and the rest of the hold let go. What is left is the cap less what is settled and what is held, so a call that is still running will already count. The hold is why a cap will hold even when many calls start at once.
The control plane marks a budget that is near its cap or at it, and the console and the command line show that mark as it was sent. Nothing here works out a percentage of its own.
Reading spend
In the console, the Spend tab of AI shows what a project's budgets allow and what has been spent against them, and the Budgets tab of AI settings lists every budget you may read. On the command line:
rl model spend -project <project> -env <environment>prints the budgets that apply to a project: the cap, what is settled, what is held and what is left, in dollars.rl model budgetsprints every budget your credential may read, or those of one project with-project.rl model budget set -scope project -project <project> <dollars>lowers one budget's cap. The scope says which budget:organisation,project,agent,actionorshadow, with-agentand-actionwhere the budget needs them.
The tables print dollars, worked out exactly from the whole micro-cents (a millionth of a cent) the control plane holds. Each command prints
the response as one JSON document, as sent and so in micro-cents, with -o json. None of the commands has an option that could carry
a key or a secret. They exit 0 when it worked, 1 when the control plane refused it, and
2 when nothing was read or changed, which includes a figure that would be a raise. The rest of the command line is on
The rl command line.
What this page leaves out
It does not say how a call is priced, because no price list is published yet and no call is made. It does not describe a call being refused at a cap, an agent stopping, or a model's bill being checked against what was settled: those are written when they exist.