---
title: "Agent cost modeling · estimate the monthly cost to run an AI agent · Agent Fieldbook"
url: https://agentfieldbook.org/economics/cost_modeling/
description: "An interactive estimator for what it costs to run an AI agent per month: model inference from vendor-published token rates, infrastructure, storage, retrieval, gateway markup and human review, summed to a total, per-user and per-call figure. A source-graded prototype that shows why, at real review rates, human oversight usually costs more than the model."
section: "Market · Cost Modeling"
source: Agent Fieldbook — generated from the published page
---

# What will it cost to run this agent

Illustrative model, not a quote. Pick a model, an infrastructure, data + retrieval + gateway posture. Set your user + call volume + human review rate. Get a monthly estimate, with a full breakdown across model, infra, storage, retrieval, gateway and HITL, grounded in vendor-published rates.

## $59 per user $0.134 per call 44k calls / month

## Common questions

### How is the monthly cost of an AI agent calculated?

Seven lines summed: model inference, infrastructure, storage, retrieval, an optional gateway markup, scheduled jobs and human review. Inference is calls per month x tokens per call x the vendor's published per-token rate, counted separately for input and output. The others are either a fixed monthly figure or a per-event rate x volume.

### How are token costs worked out?

Tokens divided by one million, multiplied by the published price per million, done separately for input and output because output usually costs several times more. Rates come from the model catalogue, which carries each vendor's own published figures. A model with no published rate falls back to a conservative default rather than a guess.

### How is human review priced?

Calls per month x the review rate you set x a per-review cost of $2.20, a mid-market fully-loaded rate. It is deliberately volume-sensitive rather than a fixed overhead, which is why moving the review percentage moves the total more than changing model does. Substitute your own loaded cost per review for a real estimate.

### Where do the prices come from?

Vendor-published rates only. Model prices come from the foundation-model table, infrastructure from the cloud table, each row carrying its own source link and verified date. The only exception is a model with no published rate, which falls back to a conservative default rather than a guess.

### What does the estimate leave out?

Batch and streaming discounts, cached-input pricing, multi-model routing, cross-region egress, and compliance overheads such as EU-only, HIPAA or FedRAMP infrastructure. Each of those moves a real bill materially, so treat the total as an order of magnitude for comparison, not a quote.

### Why does the total change so much when I adjust review?

Because inference is cheap and oversight is not. At realistic review rates a supervised agent spends more on the person checking its work than on the tokens it generates, so the review slider dominates the total. That relationship is the main thing this tool exists to show.

## Next in the learning path

- [Pricing The per-seat to per-outcome shift](https://agentfieldbook.org/economics/pricing/)

- [Cost Trends Why inference is the cheapest part](https://agentfieldbook.org/economics/cost_trends/)

- [Billing Shift The flat to usage pivot](https://agentfieldbook.org/economics/billing/)

- [Models The reasoning cores you cost here](https://agentfieldbook.org/software/models/)

- [Cost Calculator Headcount versus agent cost](https://agentfieldbook.org/workforce/calculator/)

## Inputs

| Input | Min | Max | Default |
| --- | --- | --- | --- |
| Active users | 1 | 10000 | 100 |
| Calls per user per day | 1 | 200 | 20 |
| Context tokens / call | 500 | 200000 | 6000 |
| Response tokens / call | 100 | 20000 | 1200 |
| MCP / tool calls per session | 0 | 30 | 4 |
| Human review rate | 0 | 100 | 5 |