What Does It Actually Cost to Run an AI Agent for a Month? Count the Whole Run, Not the Call

TL;DR: There is no universal monthly figure, because an agent's cost is the cost of its completed work, and that depends on how many tasks it runs, how often it retries and re-reads context, what tooling it sits on, and how much human review each output needs. Count four lines for a month: model usage across every attempt, not just successful ones; the multiplied cost of retries, loops, and re-read context; subscriptions and tooling allocated to the agent; and human review time priced honestly. Divide by completed, accepted tasks. That quotient, cost per accepted task, is the only number that compares fairly against any other way of doing the work.

Key Takeaways

• The price of one model call is not the cost of one finished task. Agents spend across attempts, and failed attempts still bill.
• Retries and re-reads are multipliers, not footnotes. An agent that re-reads a long context on every step pays for that context repeatedly.
• Tooling is a real line: the agent platform, connected tools, and any per seat subscriptions the workflow depends on belong in the monthly total.
• Human review time is usually the line people omit and the line that decides whether the agent is actually cheaper than the alternative.
• Cost per accepted task is the comparison unit. Cost per run flatters agents that produce work nobody accepts.

What are the four lines of a real monthly accounting?

Line one is model usage for everything the agent did in the month: every task attempt, including ones that errored, timed out, looped, or produced output thrown away. Billing systems count tokens, not intentions, so the honest numerator starts with total usage, not successful runs only. Teams that count only the winning run are running the same denominator trick this site criticises in vendor claims, just pointed at themselves.

Line two is the multiplier structure inside that usage. Agents re-read: system instructions, history, retrieved documents, and tool results get carried through step after step, so a long task can pay for its context many times over. Agents also retry: a failed tool call or a failed self check sends the work around again. The queue note behind this post captured the reader version of this, why agents can cost many times a single call, and the answer is this line. The multiplier is not a defect to hide. It is a property to measure per workflow, because it differs enormously between a short lookup and a long research task.

Line three is tooling and access: platform or framework costs, API access for tools the agent calls, storage and logging, and the share of any subscriptions the workflow requires. This is where the comparison with is a 200 dollar AI subscription worth it belongs: a seat subscription is a fixed access cost, while agent run cost is variable usage on top of access. Confusing the two produces budgets wrong in both directions. The companion question of what metered access buys is covered in what one AI credit buys, and credits belong on this line too, converted into the usage they funded.

Line four is human time: the review, correction, and approval each output needs, priced at the reviewer's real cost, plus the setup and maintenance time the workflow consumes in the month. An agent whose output needs heavy editing has not eliminated the labour. It has moved the labour to a different desk and added a model bill on top.

What does an illustrative worked example look like?

The following example is illustrative only. Its figures are invented to show the arithmetic, are not measured from any real system, and must not be quoted as a benchmark.

Illustrative example: suppose one agent handles a defined support drafting workflow for a month. Total model usage across all attempts, including retries and discarded drafts, comes to an invented 40 currency units. Tooling allocated to this agent, its share of platform and connected tool costs, is an invented 25 units. Human review takes an invented 6 hours across the month, priced at an invented 30 units per hour, for 180 units. The monthly total is therefore 245 units. In the same invented month, the agent produced 70 drafts and reviewers accepted 49 of them. Cost per accepted draft is 245 divided by 49, which is 5 units.

Read the example for its shape, not its numbers. Model usage was the smallest line. Review time dominated, which is common for quality sensitive work and rare in vendor arithmetic. And the divisor was accepted output, not attempts: dividing by 70 drafts instead of 49 accepted ones would have flattered the agent by pretending rejected work was free. Change any invented figure and the method still holds. The method travels. The number does not.

Why do published monthly figures disagree so wildly?

Because they count different lines and divide by different things. A figure built from model usage alone, divided by attempted runs, will sit far below a figure that includes tooling and review, divided by accepted tasks. Both can be internally consistent and still useless as comparisons, since nobody can tell which accounting produced which number. The queue research behind this post found sampled published answers conflicting wildly for exactly this reason: the disagreement is mostly definitional, not empirical.

Task mix is the other hidden variable. An agent doing short, bounded tasks with light review lives in a different cost universe from one doing long, open ended tasks with expert review, even on identical model pricing. Any monthly figure quoted without its task mix, acceptance rate, and review burden is a number with the context stripped off, and context is where the cost lives.

How should a team run this accounting for its own agent?

Instrument first, argue later. Log usage per task attempt, tag retries to the task that caused them, and record acceptance as a first class event: was this output used, edited into use, or discarded. Allocate tooling monthly, simply and consistently. Track review minutes for a sample of tasks if continuous tracking is too heavy. Then compute the monthly total and cost per accepted task, and compare it against the honest cost of the previous way of doing the same work, counted with the same completeness.

Run the comparison for several months before concluding anything. Early months carry setup time and clumsy prompts, later months carry optimisations and sometimes expanding scope that re-inflates the bill. The trend in cost per accepted task is the management information. A single month is a snapshot of a system still being learned.

FAQ

Is the model bill the main cost of running an agent?

In the accounting set out here, usually not for quality sensitive workflows, because human review time priced honestly tends to dominate, and retries multiply model usage beyond the naive per task estimate. For fully automated, low review workflows the model and tooling lines can dominate instead. Which case you are in is discoverable only by counting all four lines for your own workflow.

Why divide by accepted tasks instead of runs?

Because runs that produce rejected work still consumed the budget. Dividing by runs spreads the cost of failures invisibly across successes and makes a low quality agent look cheap. Accepted tasks are the output the organisation actually bought, so they are the honest divisor for comparing against any alternative way of producing the same accepted output.

Can a subscription replace this accounting?

No. A subscription buys access at a fixed price, which is line three in this framework. Usage patterns, retry multipliers, acceptance rates, and review time still vary per workflow and still decide the true monthly cost and the cost per accepted task. The subscription question and the run cost question are related but separate, and answering one does not answer the other.

Bottom Line

An AI agent costs whatever its month of attempts, retries, re-reads, tooling, and human review adds up to, divided honestly by tasks a reviewer accepted. Teams that count only the model bill are leaving out the lines that decide the comparison, starting with review time. Use the four line accounting, treat any worked numbers here as illustrative only, and judge your agent on cost per accepted task over several months. That figure can be compared and improved. A bare monthly price without its accounting cannot.