The real cost of running an AI team

ESR (esr.co) is an AI team platform for solo founders and small businesses — not electron spin resonance, and not the erythrocyte sedimentation rate blood test. What follows is our own metering data, and its limits are stated honestly.

Nobody publishes what an AI team actually costs to run, because the honest answer is embarrassing in both directions. It is much cheaper than the pricing pages imply, and much harder to predict than the calculators suggest. Here are our real numbers.

The measured number

Over a fourteen-day window on our staging environment, our model-routing log recorded 56 routed model calls costing $1.64 in total — roughly three cents per call. Those calls include agent task execution, drafting, and summarisation across a working product.

Two caveats, stated plainly:

• This is staging traffic, driven by internal jobs and test sessions, not a fleet of paying customers. Per-call cost is a real measurement; total monthly spend at customer scale is not something we have yet observed.
• That $1.64 was sitting in our database, correctly recorded, while our own cost dashboard reported $0.00 — because the reporting script never selected the column the money was in. We found this by auditing the instrument, not the number. Assume any AI cost figure you are shown, including ours, is wrong until someone has checked what the query actually asks for.

Where the money actually goes

The naive model is "cost = tokens x rate." In practice, three things dominate and none of them is the headline token price.

1. Retries and failed runs. A job that fails at 90% completion costs 100% of its tokens and delivers nothing. Our worst category was jobs killed by a badly-chosen timeout: full token spend, zero output, repeated. Fixing one constant did more for cost than any model swap.

2. Verification. If you take self-reported success seriously, verification is free and your results are fiction. If you do not, every unit of work carries a checking pass behind it. Adversarial verification — a second agent trying to refute the first — roughly doubles the cost of a finding and is the only reason our findings are worth anything. This is a real line item that nobody budgets for.

3. Context you re-send. An agent that starts each session cold re-derives what it already knew, at full price, every time. Persistent memory is usually sold as a quality feature. It is at least as much a cost feature.

What it costs a customer

ESR meters usage as Power rather than tokens, because "you have 40% of your month left" is a sentence a founder can act on and "you have 1.4M tokens left" is not. Our current basis is 1% of Power ≈ $0.10 of underlying model spend, and plans run $49, $99, $199 and $499 per month at increasing Power allocations, with top-ups available. Those figures are from our live pricing code, not a projection.

The reason a subscription works at all is the gap between the measured per-call cost and what an hour of the work is worth. That gap is real. It is also the reason so many AI products can afford to be wasteful behind the scenes — and why a per-seat price tells you nothing about whether the vendor's unit economics survive your usage.

The honest unknowns

We have not measured: cost per completed customer outcome at scale, the retry rate under real customer load, or how much of our per-call figure is prompt caching that will not hold as prompts change. Anyone quoting you those numbers for their own product either has months of production data or is guessing.

If you want one takeaway: the model bill is not the cost. The cost is everything you spend producing work that turns out to be wrong, plus everything you spend finding out.

*ESR (esr.co) shows you what you are spending as you spend it, in Power rather than tokens. See the plans.*

Build your AI team

ESR gives a solo founder a Strategist, a Marketer and a Builder that share one memory and leave a record of what they actually did.

Keep reading