FREE TOOL · QUADRATIC MODEL
What your AI agent costs in production
Price per call is the visible part and the least important one. The bill is driven by calls per task, context re-sent every turn, and silent retries.
What your AI agent costs in production
Estimated cost
- Cost per task
- —
- Tokens per task
- —
- Per day
- —
- Per month
- —
- Per year
- —
Estimate excludes infrastructure, vector storage and human supervision. Token prices are yours to enter: they change too often to be frozen into a tool.
Input cost grows quadratically, not linearly
This is the mistake that blows budgets. On every turn, a conversational agent re-sends the whole conversation history. On the sixth call it does not pay six times the initial context: it pays the sum of six accumulated contexts, twenty-one times a single turn.
The formula is n(n+1)/2. Going from three to six calls per task does not double the input bill, it multiplies it by three and a half. That is why an estimate made by simply multiplying price per call is systematically optimistic by a factor of three to five.
The retries nobody counts
An agent in production replays calls: invalid output format, a tool that fails, an answer rejected by validation. These retries appear nowhere in a spreadsheet estimate, and commonly account for fifteen to thirty percent of the bill.
They are also the easiest line to cut. Constraining output format with a strict schema costs a few extra tokens per call and removes much of the replaying. It is almost always a winning trade.
What this calculator does not cover
Infrastructure, vector storage, human supervision and engineering time are not counted. On many projects they exceed model cost — especially human supervision, which stays necessary until the error rate has come down far enough.
Token prices are yours to enter rather than hardcoded: they change too often for a frozen tool to stay accurate. Check them with your provider before building a budget on them.
Common questions
Why not simply multiply price per call?
Because context is re-sent. Total input cost for an n-call task is n(n+1)/2 times one turn's context, not n times. That gap is exactly the difference between a naive estimate and the real bill.
Do the default prices match a specific model?
No, deliberately. Three and fifteen dollars per million tokens is a common order of magnitude for a mid-range model. Replace them with your provider's real rates: they move too fast to be frozen.
How do I estimate my retry rate?
If the agent is not in production yet, twenty percent is a reasonable assumption for an unconstrained output format, five to ten with strict validation. Once live, measure it: it is often the first month's surprise.
Are my figures sent anywhere?
No. All the computation happens in your browser, with no network request.
An AI project rarely dies of a bad model.
It dies of a scope with no owner, no measurement and no acceptable failure threshold. Costing is the easy part — framing decides the rest.