WHITEBOXX / CALCULATOR

What will your AI agent
cost to run?

Answer four questions to estimate what an AI agent costs to run each month, per run and over three years. In our worked example, model usage is about 3% of the monthly cost. Checking the results costs far more.

By Gautam Mehra · Published · Updated
2,000
2How much work is each run?
3Which size of model?
5 min
Estimated $12,217 a month See breakdown ↓
Adjust the assumptions

Exact values behind the answers above, plus people and maintenance costs.

ESTIMATED RUNNING COST

$12,217a month

Human review is 82% of the monthly cost. Model usage is 3%.

  • Human review$10,000
  • Maintenance$1,333
  • Hosting and data$500
  • Model usage$384
Per run
$6.11
Over three years, including the build
$519,824

An estimate for comparing options, not a quote.

How much does it cost to run an AI agent?

In our worked example, an agent that runs 2,000 times a month costs about $12,217 a month to operate, or $6.11 per run. Model usage is $384 of that. Reviewer time is $10,000, because a person spends five minutes checking each result.

The example shows a common pattern. Once an agent is in production, the cost of checking its work usually exceeds the cost of the model. Change the answers above to match your task.

PRICES, NOVEMBER 2026

The model price tiers in the calculator.

PRICES, NOVEMBER 2026
TierInput, per million tokensOutput, per million tokensTypical use
Small$0.10$0.50Classification, extraction and routing
Mid-size$2$10Most agent steps, including reading and drafting
Large$5$25Hard reasoning and long, complex tasks

Where do the prices come from?

The tiers reflect published list prices from the main model providers in November 2026. Each provider offers models near each tier. Prices change often, and batch processing or cached input can cut them by half or more, so check the provider’s pricing page before you set a budget.

INPUTS

How to fill in the inputs.

  1. 01

    Runs per month

    Count how often the task happens today. Use a busy month.

  2. 02

    Model calls per run

    An agent usually makes several calls per task: planning, each tool step and the final answer. Five to fifteen is common.

  3. 03

    Tokens per call

    A token is about three-quarters of an English word, so a page of text is roughly 500 to 700 tokens.

  4. 04

    Retries and overhead

    Allow for failed steps, retries and longer runs. 20% is a cautious starting point.

  5. 05

    Review time

    The time a person spends checking each result. Use the review pattern you plan to use, such as checking every result or only exceptions.

  6. 06

    Build and maintenance

    Use your own estimate or a quote. Many teams budget 15% to 25% of the build cost each year for maintenance.

Work on it in Excel.

The same cost model as an Excel workbook, with every formula visible and the assumptions labelled.

Download the cost model (.xlsx) ↗

What does the calculator leave out?

It does not include buying outside data, security reviews, staff training or the time your team spends defining the task. It also assumes a steady volume. If usage will grow, run the numbers again at the higher volume.

Treat the result as a first estimate for comparing options, not a quote.

How do you lower the cost?

Review time is usually the biggest lever. Reviewing only flagged exceptions instead of every result can cut review cost sharply, but only if each run keeps the sources and steps a reviewer needs. After that, use a smaller model for simple steps, and cache input that repeats, such as instructions and reference documents.

Read the full cost guide