Alerts and cost

Cost

Where an Environment’s money goes: spend by model, service and day, judge spend separated from model spend, what “unpriced” means, and the online-rule budget card.

After this page you can see where an Environment’s spend goes, drill from any number to the traces behind it, tell model spend from judge spend, and change the online-rule budget from where the spend is visible.

The Cost page

Cost is environment-scoped. A range picker (24 h · 7 d · 30 d · month to date), totals (LLM and Judge spend, and simulation spend when there is any), by model (model · calls · tokens in and out · cached tokens · cost, with an unpriced marker), by service and by day (stacked LLM against judge), and the online-eval budget card. Clicking a model row opens Traces filtered to that model and sorted by tokens; a trace opens with its estimated cost and Optimize ranks its cost issues with a fix prompt. Cost needs no plan feature.

How a number is computed

Spans carry token counts, not dollars. The trace store returns tokens per model (deduplicated per trace); the control plane prices them from the model catalogue’s price sheet — with any price override a platform admin has set applied on top — in dollars per million tokens: uncached input, cache-read and cache-write tokens (each a subset of the input tokens, never double counted) and output tokens, each at its own price. Judge spend is the sum of the judge cost recorded on every trace evaluation in the window — manual and online — and is already separate from model spend, so nothing is added twice. Cost is computed on read, so a price change in the catalogue applies immediately; costs stored on finished runs were frozen when they ran.

Unpriced means unknown, never zero. A model without a price row, or an evaluation whose judge model had no price, makes its cost null; any null in a sum makes the sum null, and the page shows the part that could be priced as a floor (“at least $x, n unpriced”). A window with no calls is a real $0.

The budget card

The card reads the Environment’s online rule: spent today of the daily budget, whether it is paused, what it will resume at (UTC midnight), and how many traces sampling skipped. Change budget saves a new dailyBudgetUsd; when the new budget exceeds what was spent, the rule resumes immediately. Alert me opens an alert rule on the Judge spend metric so the next day’s spend pages you before the pause. Details in Online rules.

Estimate a workload

With the cost-calculator plan feature, Estimate a workload walks from a description of an agent through a use case, a workflow graph with tokens per node, scenarios with probabilities, and an estimate per conversation and per month with a breakdown and a confidence score; reports are saved per Environment and can be verified by a simulation run, whose observed cost per conversation is then recorded beside the prediction — see Launch and read a run.

Limitations

Traces sort by tokens as a proxy for cost; the trace list has no cost sort.
There is no re-pricing of stored history after a catalogue price change; trace and span views re-price on read, judge costs on finished runs do not.
Per-user cost is summed in the browser from the user rollup’s tokens and is approximate.