Models and cost
How the agent uses a strong and a cheap model, what a run costs, and the budget, profiles and vendors an organization chooses.
An investigation uses two models. A lead model (strong) plans it and writes the answer; a worker model (cheaper) calls the tools and reads what they return. Most of a run's tokens are tool output, so that's where the cheaper model saves the most, and the answer you read is still written by the strong one.
Who can: Owners and Admins change the settings (Agent settings). Owners, Admins and Managers see what runs cost (Usage).
Profiles
Each kind of run has a profile: how it uses the two models, how many steps it may take, and how much it may spend.
| Profile | What the models do | Default for |
|---|---|---|
| Economy | The worker investigates; the lead writes the answer. | Alerts, reports |
| Balanced | The lead plans first, the worker investigates (and can ask the lead when it's stuck), the lead writes the answer. | Investigate |
| Thorough | The lead does every step: the most capable, and several times the cost. | — |
Agent settings shows each profile with the models it uses for your organization right now, its steps and its spending limit per run. When a run reaches 80% of its limit, it stops investigating and answers with what it found. On sremcp, it shows the credits a run on each profile uses instead, and Thorough is on the Business plan.
Quick and deep
In Investigate, Quick asks for a short look: the most likely causes first, fewer steps, a smaller budget. Deep is a full investigation, and the default. Alert investigations are quick.
What keeps runs small
- Tool results are shortened for the model. Long logs, time series and Kubernetes objects are tidied and capped; the model gets a note saying what was left out and can read the rest on demand. You still see the full result.
- Older results become notes as an investigation goes on, and a follow-up question doesn't send earlier answers' tool output again.
- Fewer approved tools. Every approved tool's definition goes to the model on every step. Select recommended on a server's tools approves the read tools that answer most questions. See Approve tools.
The monthly budget
On sremcp, runs on sremcp's models use credits, with their own limit in Billing; this budget applies to runs on your own keys.
Set a budget in dollars for each calendar month (UTC), and what happens when it's spent:
- Keep investigating on Economy, as quick looks: answers keep coming, at the lowest cost.
- Refuse new runs until the month turns: questions get a message saying when it frees up, and alerts and reports are recorded as not run.
Where your data goes
Where the data may go limits the vendors whose models read your questions and tool output: Google, Anthropic, OpenAI, or open-weight models (DeepSeek, GLM, Qwen and the like). Open-weight models only ever run on US/EU hosts, never on their makers' own APIs. If no model is left from the vendors you choose, the settings say so instead of saving.
Your own keys
Add your organization's own key for Google AI Studio, Anthropic or the Vercel AI Gateway, and runs on that provider
are billed to your account. On sremcp, own keys are on the Team plan and up, and runs on them use no credits. With a key, you can also choose your own lead and worker models, for example
anthropic:claude-haiku-5-5 as the worker. Keys are encrypted for your organization and never shown again; removing
one puts runs back on the deployment's.
Usage
On sremcp, Usage shows the credits used this billing period against your plan, credits per day for the last 7, 30 or 90 days, where they went (by kind of run and by person) and the runs that used the most. Runs on your own keys are shown in dollars below, as follows.
For runs on your own keys, Usage shows this month's spending against the budget, spending per day for the last 7, 30 or 90 days, and where it went: by kind of run, by model (and its role), by person, and the most expensive runs. Every model call is priced at its provider's rate as it happens, cached input at its lower rate. Each run in Runs shows its cost, and its page lists every step.