BreathingRoom
Guide

What an AI Agent Costs Per Month to Run

Build cost gets quoted. Run cost decides whether the thing survives its first budget review — models, platform, infrastructure, review time and maintenance, itemised.

The bill after the build

Build cost gets quoted. Run cost usually does not, and it is the one that decides whether the thing survives its first budget review. Here is what is actually on it.

LineWhat drives itPredictable?
Model usageTokens in and out. Rises with volume, and with every retryNo
Orchestration platformPublished per-seat, per-run, per-credit or per-task feesOnly if you know your volume
Hosting / infrastructureMetered compute. AWS bills $0.0895 per vCPU-hour, Google $0.085No
Monitoring and loggingUsually bundled, sometimes notMostly
Human reviewSomebody reads the output. This is real time and it is not freeYes, and usually underestimated
MaintenanceAn API changes, a field is renamed, an edge case appearsNo, but it is not optional

Why “it depends on usage” is a problem, not an answer

A usage-based bill means your cost rises exactly as the automation succeeds. For a large company that is fine — the savings scale too. For a company of forty people with a fixed monthly budget, an unpredictable line item is a genuine risk, and the finance person is right to push back on it.

It is also why the billing unit matters more than the headline rate. We compiled what 38 vendors actually publish and found 24 different units in use — credits, tasks, runs, actions, outcomes, resolutions, conversations, sessions, minutes, vCPU-hours. Two quotes with the same number on them can differ by an order of magnitude in practice.

Where we sit

$2,500/month. Month-to-month. Cancel anytime. Flat means the variance sits with us. That is not generosity — it is the only version of this that a company your size can plan around, and it forces us to care about efficiency in a way a metered vendor does not.

Work your own numbers with the cost calculator — it includes the review and maintenance lines rather than quietly dropping them.

Where these numbers come from. The AWS and Google infrastructure rates are quoted from those vendors’ own pricing pages, read on 22 August 2026. Everything else on this page is a cost structure rather than a price, because the actual figure depends on your volume and your vendor’s billing unit — of which we counted 24 distinct kinds across 38 vendors.

Related