- Format
- Statistics
- Territory
- AI consulting
- Family
- Readiness and measurement
- Basis
- First-hand
Reviewed
Statistics
What AI costs to run: where the monthly bill actually comes from
What a handed-over AI system costs to run each month, from the studio's own record: model tier by task, token volume, a monitoring window measured in weeks, then minimal human time.
Reviewed by Ameya Sahasrabudhe and Swati Thakur,
What this page records
What an AI system costs to run each month depends entirely on the specific business workflows it enables and the nature of the work it does - there is no single maintenance number, and this page does not invent one. In the studio’s operating record, four things set the bill. Which model tier the task demands: strategy and reasoning-grade work needs frontier-tier models, while simpler high-volume work, such as parsing a large volume of documents into task items, runs on smaller, cheaper models. How much work flows through the system: token cost scales roughly linearly with volume up to a point, after which batching and optimisation bring it down. A monitoring window of a few weeks right after deployment, when checking the system’s quality is the main human cost. And after that window, once the system has been tested in real use and shown to hold up, a human-time cost that falls to minimal. Each record below states the claim, how the studio knows it, and what it rests on. This is the running cost of the systems this studio builds, stated as operating experience - not a market survey.
Two model tiers
What a workflow's task nature selects: frontier-tier models (Opus-class) for strategy and reasoning-grade work, smaller fast models (Haiku-class) for high-volume mechanical work such as parsing large document sets into task items
Linear, to a point
How token cost moves with work volume: roughly linear as volume grows, up to a point past which batching and other optimisation passes bring the cost per unit of work back down
A few weeks
The window right after deployment in which monitoring the system for quality is the main running cost; once a system has been tested in real use and shown to hold up, the ongoing human-time cost is minimal
Low to higher
The budget span across workload types: a law firm automating the creation and tracking of court-order compliance tasks runs on a very low budget; a creative agency generating images and video carries heavier token consumption and a higher one
Almost always
How often AI has come in cheaper than a dedicated human resource doing the same job, irrespective of the scale and complexity of the system - the studio's stated position from its own record, not a cited market statistic
| Number | What it measures | Period | Source |
|---|---|---|---|
| Two model tiers | What a workflow's task nature selects: frontier-tier models (Opus-class) for strategy and reasoning-grade work, smaller fast models (Haiku-class) for high-volume mechanical work such as parsing large document sets into task items | As of August 2026 | Foxnut Studios engagement records (not externally retrievable) |
| Linear, to a point | How token cost moves with work volume: roughly linear as volume grows, up to a point past which batching and other optimisation passes bring the cost per unit of work back down | As of August 2026 | Foxnut Studios engagement records (not externally retrievable) |
| A few weeks | The window right after deployment in which monitoring the system for quality is the main running cost; once a system has been tested in real use and shown to hold up, the ongoing human-time cost is minimal | As of August 2026 | Foxnut Studios engagement records (not externally retrievable) |
| Low to higher | The budget span across workload types: a law firm automating the creation and tracking of court-order compliance tasks runs on a very low budget; a creative agency generating images and video carries heavier token consumption and a higher one | As of August 2026 | Foxnut Studios engagement records (not externally retrievable) |
| Almost always | How often AI has come in cheaper than a dedicated human resource doing the same job, irrespective of the scale and complexity of the system - the studio's stated position from its own record, not a cited market statistic | As of August 2026 | Foxnut Studios engagement records (not externally retrievable) |
What most running-cost estimates get wrong
The first mistake is pricing the model instead of the workflow. Per-token prices are public and real - the model vendor publishes a rate card - but the bill is set by which tier the task actually needs and how much work flows through it. A law firm tracking court-order compliance tasks and a creative agency generating video sit on the same rate card and end up with very different budgets, because one moves small amounts of structured text and the other moves media. The second mistake is extrapolating the first invoice forward in a straight line. Cost does scale roughly linearly with volume at first, but past a point batching and optimisation passes bend the curve down, so the month-three bill is not the month-one bill times three. The third mistake is treating the human cost as a permanent salary line. In the studio’s record it concentrates in the first few weeks after deployment, while the system is being monitored for quality; once it has held up in real use, the ongoing human time is minimal.
Cheaper than the hire: the studio’s position
One claim on this page deserves to be labelled with special care. In the studio’s experience, irrespective of the scale and complexity of the system being implemented, AI is almost always cheaper than a dedicated human resource for the same job. That is the founder’s own conclusion from the systems this studio has built and handed over. It is presented here as exactly that - a position taken from an operating record - and not as a market statistic, because no public benchmark supports or refutes it at the scale this studio works at. It is also a claim about the jobs these systems do, not a claim that every AI deployment everywhere beats every hire.
What this page cannot show you
The two engagements behind the low-to-higher contrast are real and anonymised: no client is named on this site beyond the published client list, and their monthly bills are client-confidential, so the contrast is stated as a span rather than as figures. The circulating market ranges for what AI costs to run each month are deliberately absent: the ones this page’s research surfaced trace to vendor marketing rather than to primary publications, and an untraceable number is not a statistic. Per-token model prices are also not restated here - they are the vendor’s published numbers, they change, and what a unit of model work costs is a different question from what a running system costs its owner. And this page does not state what the studio charges to build a system; that number lives in one place on this site, deliberately.
How these records were compiled
Every record on this page comes from one source: the studio’s own engagement records, summarised by the founder on 5 August 2026 and edited for publication without adding numbers. None of it is externally retrievable, and each record says so rather than dressing assertion up as public data. The single external source cited is the model vendor’s published price list, fetched and read the same day - it is what makes the two-tier claim in the first record checkable, since the price gap between frontier-tier and small models is public. Two classes of figure were excluded on principle: monthly cost ranges that could not be traced past vendor marketing to a primary publication, and any repetition of per-model unit prices, which belong with their publisher. Where this page asserts, it says it asserts.
Sources
Foxnut Studios works on briefs like this one from Bengaluru and Paris. If you want the shape of that before you talk to anyone, here is what an AI engagement costs before you ask.