What you are paying for in an agent build
The model calls are the smallest part of the build. What costs money is everything that makes it safe to let a model act:
| Part of the build | Why it costs | What moves it |
|---|---|---|
| Tools and integrations | Every system the agent touches brings its own authentication, failure modes and edge cases | The number of systems — the single biggest driver |
| Evaluation set | Real cases with known correct actions, scored on every change | How many distinct situations the agent must handle |
| Review step and thresholds | Below a confidence threshold the agent proposes and a person approves | How much the agent may do without a person |
| Audit trail | Every action traceable to its input, its source and the step that produced it | Compliance scope |
| Retry-safe writes | Agents retry, and a retried write must not act twice | Which actions cannot be undone |
| The orchestration itself | Prompts, tool calling, routing between models | Smaller than expected — rarely what decides a quote |
The retry row is a database problem before it is a model problem, and it is the one most agent estimates skip. One constraint that makes a retried booking harmless is worth more than any amount of prompt wording asking the agent not to repeat itself.
The running cost is per step, not per agent
Every step an agent takes is a model call, and every call costs money and time. Using the one system whose numbers we have measured — a retrieval-backed call at about $0.011 and 1.9 seconds to the last token — and multiplying, which is arithmetic rather than a measurement of any agent:
| Model calls per task | Cost per task | Time per task, steps in sequence | At 40,000 tasks a month |
|---|---|---|---|
| 1 | about $0.011 | about 1.9 s | about $440 |
| 3 | about $0.033 | about 5.7 s | about $1,320 |
| 5 | about $0.055 | about 9.5 s | about $2,200 |
| 10 | about $0.11 | about 19 s | about $4,400 |
Read the table as a shape, not a quote. Real steps differ in length, some run in parallel, and a step that needs no retrieval is cheaper than one that does. What survives every variation is the shape: cost and waiting time grow with the number of steps, so the cheapest agent is the one with the fewest steps that still does the job. For your own model’s prices and step count, the AI token cost calculator does this arithmetic per task and per month.
Three ways to make an agent cheaper
- Do not build an agent. If the steps are always the same, it is a pipeline with one model call in it — cheaper, faster and testable line by line. A surprising share of agent requests turn out to be this.
- Shorten the context. On the system above, generation is about $0.009 of the $0.011, and most of that is the retrieved passages arriving as input tokens. Fewer, better-ranked passages cut cost on every step.
- Route the easy steps. Not every step needs the largest model. Classification and routing steps can usually go to a smaller one, and repeated lookups can be cached.
Why agent quotes vary so much
Because two things that are both called an agent can differ by an order of magnitude in what has to be built. A single workflow that reads, proposes and waits for a person to approve is an integration. A system of agents that act on several business systems without supervision needs an evaluation set for every action it may take, an audit trail and retry-safe writes everywhere it touches — which is why the top of the AI product band exists.
We quote managed delivery fixed-price after a discovery call, with a written scope inside 48 hours. If the agent will keep changing with your product, AI engineers on your team at $4,000–$7,000 a month is usually the better shape. What we build and how is on the AI agent development page.