Blog · AI ·
What Running AI Agents Actually Costs: Our Real Numbers
Real numbers: $0.0846 per AI video, $0.19 per research run, a 20-agent fleet for less than a part-time hire. Tokens are cheap; the harness is the real cost.
Running our 20-agent fleet costs less than a part-time hire, most months. Not because tokens are magic and free, they're not, but because tokens were never the expensive part. A single AI-generated video costs $0.0846 in model spend. A nightly insight run over our stored research costs $0.19. The real cost is the engineering that makes any of it trustworthy enough to ship.
How much do AI agents cost? The numbers, up front
Let's just put them on the table, since everyone else in this space seems allergic to doing that.
- $0.0846 per video — one AI-generated, human-editable, renderable short video, through our own internal studio pipeline, measured across the whole generation pipeline, not a cherry-picked step.
- $0.19 per insight run — one nightly knowledge-graph run, the kind that cross-references our stored research and spits out something usable, for the entire run, and it produced 12 usable proposals out the other end.
- $15/day research cap — the ceiling our autonomous research agent works under, and most days it doesn't even hit it.
- ~$100/month programme budget — what our whole self-improvement programme, the queue of infrastructure upgrades that agents chip away at over months, was planned against in model spend.
A hundred dollars, give or take, for a running programme of upgrades. Zoom out further and it stays boring: a recent 7-day sample across our repos showed 402 commits, with agents doing most of the volume work. That's the actual economics story: throughput per pound, not some headline number pulled from a press release.
Why tokens are the cheap part
Here's the bit nobody selling "AI agent consulting" wants to say out loud: the model calls were never going to be the bottleneck on cost.
Compute got cheap. Genuinely, boringly cheap. You can run a full research pass, a video render, a scoring job, all for pennies, and the pennies keep shrinking. So when you see someone quote "$5k to $400k for an AI agent," that range isn't describing tokens. It's describing everything else, badly, and dressing it up as a token cost because that's the bit that sounds futuristic.
The token bill for our fleet, across everything, sits well under what you'd spend on coffee for a small office. We're not hiding a number here, there just isn't a big number to hide. The interesting costs live somewhere else entirely, and that's really the whole point of this piece.
The cost controls that keep it boring
We want it boring, actually. Boring is the goal. Boring means nothing surprising happened on the invoice.
The mechanics are unglamorous on purpose. Every AI call writes its token usage to a usage table, so nothing happens off the books. A budget check runs before each call, not after. Monthly caps are set per surface, so no single feature or agent can quietly run away with the budget. Classification and scoring work goes to a cheaper, Haiku-class model, roughly a third of the cost of the heavier Sonnet-class one, because scoring a proposal doesn't need the same horsepower as writing one. And prompt caching on stable system blocks means repeat calls pay the cached-input rate instead of the full rate. None of this is clever. It's just discipline, applied consistently, which turns out to matter more than any single optimisation. The engineering detail behind these controls is in our Claude API integration notes.
Anyway, that's the whole trick. Not a trick, really. Just checks, before the money leaves.
What actually costs money
So if it's not the tokens, what is it?
Two things, honestly, and neither of them is exciting enough to put in a sales deck. The first is the harness: the approval gates, the quality gates, the retries, the monitoring that sits around every agent so its output can actually be trusted before it goes anywhere near production. Building that harness properly takes real engineering time. It's not a prompt, it's infrastructure, and infrastructure takes weeks, not minutes.
The second is human review time. Everything that ships gets a human look at it, every time, no exceptions. That review time isn't huge for us, we've built the process to be efficient, but it's not zero either, and pretending it's zero is exactly how people get burned. You can have the cheapest tokens in the world and still torch a budget if you skip the harness and the review, because then you're not paying in dollars, you're paying in incidents.
This is the bit the "$5k to $400k" articles miss entirely. They're pricing an outcome with no visibility into where the money actually goes. We'd rather just show you the split: pennies for the model, real engineering hours for the trust layer around it.
How to think about cost if you're commissioning agents
If you're weighing up whether to bring agents into your own operation, don't ask "what does the AI cost." Ask what the harness costs, and ask who's reviewing the output.
Suppose you're commissioning a fleet of agents to handle, say, customer research and internal reporting. The token spend on that will likely be trivial, genuinely, don't let anyone quote you five figures for compute alone. What you're actually paying for is someone building the approval gates so a bad output doesn't reach a customer, someone setting budget caps so a runaway loop doesn't eat your month, someone deciding which tasks get the cheap model and which need the expensive one. That's engineering. That's the invoice that matters.
We build and run our own products this way, day to day, with human approval sitting on everything that ships, so this isn't theory for us, it's just how Tuesday works. If you want to see what that setup actually looks like in practice, our services page walks through it properly. And if you'd rather see the throughput than read about it, the homepage counts its stats straight from git, not from some marketing estimate someone rounded up.
The honest answer to "how much do AI agents cost" is: not much, on the model side, and it never was the interesting question. The interesting question is what you're willing to spend on making the output boring, reliable, and reviewed before it goes anywhere near a customer. Get that part right and the token bill takes care of itself.
Drafted by the Adapt Progress Evolve agent fleet; edited and approved by a human.