Copilot Studio computer use costs 5 credits a step: set a ceiling

Copilot Studio's computer use tool bills each step at 5 Copilot Credits, or 15 on a premium model. The number of steps a run takes is not known when it starts. So the cost ceiling has to come from how you limit the task, not from a field on the tool. A Sume Agent Completion works the other way: generation_spend_cap_usd is required, so the request is refused before any run starts unless you named a ceiling.

Billing figures are from Microsoft's computer use page, read on 2026-10-03; the Sume side is in Agent Completions.

What does one computer-use run cost?

Microsoft's licensing section says each run relies on a model that executes a sequence of steps, and each step consumes 5 Copilot Credits, or 15 with a premium model. Its example is a four-step time-sheet form: 20 credits on a standard model, 60 on a premium one. The page does not turn credits into dollars, so do that with your own Copilot plan.

Why is the ceiling hard to set?

A form that moved a button may take six steps instead of four. Nothing on the page describes a per-run credit limit, so budget by the worst case you can tolerate: the maximum steps you accept times the tier rate. The helper below does that arithmetic.

def credit_ceiling(max_steps, premium=False):
    rate = 15 if premium else 5
    return max_steps * rate

print(credit_ceiling(4))
print(credit_ceiling(4, premium=True))
print(credit_ceiling(12))

How does Sume bound the same kind of work?

Sume bills generation, not clicks. An Agent Completion returns a 202 receipt, and generation_spend_cap_usd has no default: omit it and the request fails with 400 invalid_request. The docs call the cap the substitute for the spend-approval prompt that exists in the chat UI but not for a backend caller. The receipt echoes the ceiling as usage.generation_spend_cap_usd_micros, and usage.billable_amount_usd_micros on a Format or schedule receipt shows the spend attributed to the run.

Before work starts, admission also checks balance: if Sume cannot reserve the estimated cost, the submit fails with 402 insufficient_credits and no provider work begins.

When should an agent use an API instead of the screen?

• When the target has an API, call it. A screen agent pays 5 credits for every step it spends finding a button.
• Use computer use for the UI-only parts, and keep media generation in a Sume call with its own dollar cap.
• Set a step budget in your instructions and treat a stop as a signal to check the task, not to rerun it blindly.