What does an AI agent cost to run, per month and per task?

01The formula

Cost per task is the input tokens the agent reads, multiplied by the model's input price, plus the output tokens it writes, multiplied by the output price, summed over every model call the task makes. Monthly cost is cost per task multiplied by tasks per month, plus the tools the agent calls and the retries it makes along the way.

Two things make the real number larger than the first estimate. An agent rarely makes one call per task: it plans, calls a tool, reads the result and calls the model again, and each step re-reads the conversation so far. And output tokens are usually priced several times higher than input tokens, so an agent that writes at length costs more than one that reads at length.

02What the models cost

Prices are quoted per million tokens, separately for input and output, with a lower price for cached input. The public price lists of three major model providers and one open-weights host, read on 2 October 2026, fall into four bands.

• Frontier models: $2 to $10 per million input tokens and $12 to $50 per million output tokens.
• Mid-tier models: about $0.75 to $2 input and $3.75 to $10 output.
• Small models: about $0.10 to $1 input and $0.50 to $5 output.
• Open-weights models through a hosted provider: from about $0.14 input and $0.28 output for small models to $3 input and $15 output for the largest.

Three details change the bill. Cached input is priced at a tenth of fresh input or less. Batch processing, for work that can wait, is usually half price. And output usually costs four to five times as much as input, so an agent that writes at length costs more than one that reads at length.

03Two worked examples

The arithmetic below uses the top of each price band. It is illustrative: your token counts will differ, and prices move.

1. A triage task that reads a message and writes a short label and summary: 8,000 input tokens and 800 output tokens, one model call. On a frontier model at $10 input and $50 output, that is 8 cents plus 4 cents, so 12 cents a task. On a mid-tier model at $2 and $10, 2.4 cents. On a small model, between a tenth of a cent and 1.2 cents. Run 30,000 times a month, that is $3,600 on the frontier model, $720 on the mid-tier, and $36 to $360 on a small one.
2. A multi-step agent that researches and drafts: about 20 model calls, 400,000 input tokens of which 300,000 are cached, and 20,000 output tokens. On a frontier model, about $2.08 a task with caching and $5 without. On a mid-tier model, about 46 cents. On a small model, about 23 cents. Run 1,000 times a month, that is about $2,075 on the frontier model with caching, $460 on the mid-tier and $230 on a small one.

Two lessons sit in the arithmetic. Model choice moves the cost of the same task by ten to a hundred times. And on long agent loops, caching roughly halves the frontier bill.

Agents also use more tokens than people expect. The FinOps Foundation reported in May 2026 that reasoning and agentic workloads consume five to thirty times more tokens per task than the equivalent chat interaction.

04The number to govern: cost per outcome

Cost per task tells you what the agent spends. Cost per outcome tells you what the business gets for it: the total cost of the workflow divided by the claims processed, contracts reviewed or tickets resolved. It is the number to put next to the old cost of the same work, and the one that shows whether a cheaper model, a shorter prompt or a cap is worth the change.

It also moves. Model prices change during the year, a workflow's volume grows, and an agent's prompts get longer as people add to them. A cost per outcome measured once at launch and never again is a forecast that nobody checked.

05Where the cost hides

• Retries and loops: a step that fails and retries is paid for every time.
• Long context: each step carries the whole task so far, so the last step costs most.
• The default model: the largest model chosen during the demo, kept in production for work a smaller model would do.
• Idle schedules: an agent that runs every hour checks for work whether or not there is any.
• Evaluation and testing: the runs nobody counts as production, billed at production prices.

06Four ways to bring it down

1. Route each call to the model that should take it. Most steps in most workflows do not need the largest model.
2. Cap the task. A token budget per run stops a task that has gone wrong after a known cost.
3. Cache what repeats. Providers price cached input tokens below fresh ones; a long instruction block that every run shares is the obvious candidate.
4. Measure per outcome, monthly. The change that matters shows up there first.