Agent run cost estimator
An agent resends its history on every step, so the context it carries grows with the square of the step count. This prices one run, one month, and the share that is context you are paying for again.
Estimated cost, each month
$15,070
That is $0.3014 per run, across 50,000 runs.
- Model calls per run, with retries
- 9.2
- Input tokens read per run
- 82,064
- Of that, context resent from earlier steps
- 45,264
- Resent context, as a share of the bill
- 45.1%
An estimate. It models a fixed-length loop with a constant amount added per step, which is the common shape but not every shape. Rates last checked 2026-08-20.
Why do agent costs grow faster than agent usage?
Because every step reads everything that came before it. A ten step run does not send the prompt ten times, it sends the prompt ten times plus nine steps of accumulated replies and tool results, and it pays the input rate on all of it. Output grows in a straight line with the number of steps. The context carried forward grows with the square of them. That is why an agent that was cheap in a demo of three steps becomes the largest line on the bill once it runs twelve steps against real tickets, and why the growth arrives without anybody shipping a change.
Double the steps in a loop and the context it carries forward roughly quadruples.
Nothing about the workload changed. The loop just got longer.
Four levers, in the order they usually pay.
None of these reduces what the agent does. Each is a cheaper route to the same completed run.
- Give the loop a terminal state
- A ceiling on steps and a definition of done. Loops without one do not fail loudly, they run long and bill quietly. This is the largest single lever on almost every agent.
- Cache the stable head
- The system prompt and tool schemas are identical on every step of every run, and prompt caching bills them at roughly a tenth of the input rate. This is the direct answer to resent context. Price it on the caching calculator.
- Trim what each tool returns
- Most tools return a full payload where the model reads three fields. Every extra token is carried forward into every later step of the run, so trimming a tool result once saves it many times over.
- Split the loop across models
- Routing, tool selection and extraction rarely need a frontier model, while the final synthesis often does. Running the cheap steps on a small model is the difference between the top and the bottom of the price list.
About agent costs
Why does an agent cost so much more than the number of steps suggests?
Because the history travels with it. On step one the model reads the system prompt and the task. On step two it reads all of that again plus its own first reply and the first tool result. By step ten it is reading nine steps of accumulated context before it does anything.
Output grows in a straight line with the step count. The context carried forward grows with the square of it. Double the steps and that carried context roughly quadruples, which is why the bill climbs faster than the step count does.
What counts as a step?
One call to the model. A single agent run that reads a ticket, searches twice, calls an API, drafts a reply and checks its own work is six steps, and every one of them is a separately billed request carrying everything that came before it.
Why does the retry rate matter so much?
Because a retry that eventually succeeds never looks like a failure. It does not appear in an error dashboard, nobody files a ticket, and the only trace it leaves is on the invoice. A loop with a 30% retry rate is a third more expensive than the one anybody designed.
Retries also land late in the run, where the accumulated context is longest, so each one costs more than an average step.
What actually brings this number down?
Four things, in the order they usually pay. Cap the loop so it has a terminal state. Turn on prompt caching for the stable head of the conversation, which is the direct answer to resent context. Trim what each tool result adds to the history, since most tools return far more than the model reads. Then move the steps that do not need a frontier model onto a smaller one, which is often most of them.
How exact is this?
It models one shape of loop: a fixed number of steps, a constant amount added to the history per step, and retries as extra steps. Real agents branch, summarize, and drop context, so treat the figure as an order of magnitude and the trend line as the real finding. Rates were last checked 2026-08-20.
Keep going
Prompt caching savings calculator
Price the direct fix for context you resend on every step.
Read moreAI token calculator
Compare what one call costs across Anthropic, OpenAI and Google.
Read moreAgent and workflow cost control
How this looks once it is running in production, across every agent you have.
Read more