What an AI Agent Costs to Run: Tokens, Retrieval, Hosting and Review Hours

Most AI agent budgets are written for the build and then surprised by the bill. The AI agent running cost is not one line; it is four: the model calls the agent makes on every task, the retrieval and tool infrastructure it depends on, the hosting and monitoring around it, and the human hours spent checking what it did. This guide works through each one with the model providers’ own list prices and a worked example you can swap your own numbers into.
The short answer
For an agent that completes a bounded business task in about eight model calls, model fees at November 2026 list prices run from under a cent to about 64 cents per task depending on the model tier. At 5,000 tasks a month that is roughly USD 30 on a small model, USD 640 on a mid-tier model and USD 3,200 on a top-tier model, before caching discounts. Infrastructure adds a few hundred to around USD 1,700 a month. Human review of a sample of the agent’s work is usually the largest line of all, and the one most budgets leave out.
Why an agent costs more to run than a chatbot
A chatbot answers a question with one model call. An agent plans, calls tools, reads the results, checks its own work and sometimes retries. Each of those steps is a model call, and each call carries the instructions, the task state and the tool outputs gathered so far. A task that a person would describe as one job can easily be six to twelve model calls, and the input to each later call is larger than the one before it because the context keeps growing. That is why the same model that costs cents per conversation in a chatbot can cost ten times as much per task in an agent. Our guide to AI chatbot development cost covers the single-call case; this one covers the multi-call case.
Line one: model fees per task
The table assumes one task of eight model calls, each sending about 6,000 input tokens (instructions, task state, retrieved passages, tool results) and receiving about 400 output tokens: 48,000 input and 3,200 output tokens per task. Prices are the standard-tier list prices per million tokens as published by Anthropic and OpenAI on November 12, 2026.
| Model tier | Examples and list price (input / output per million tokens) | Cost per task | 5,000 tasks a month | 25,000 tasks a month |
|---|---|---|---|---|
| Small, fast models | Claude Haiku 5.5 or gpt-6-luna: USD 0.10 / 0.50 | About USD 0.006 | About USD 32 | About USD 160 |
| Mid-tier models | Claude Sonnet 5.5 or gpt-6.1-sol: USD 2 / 10 | About USD 0.13 | About USD 640 | About USD 3,200 |
| Top-tier models | gpt-6-astra: USD 10 / 50 (Claude Fable 5.2 is USD 8 / 40) | About USD 0.64 | About USD 3,200 | About USD 16,000 |
Two things follow. The model tier moves the bill by a factor of a hundred, so the design question is which steps need a strong model and which do not. In most agents, planning and final verification benefit from a stronger model while routine tool calls and formatting do not, and a mixed design cuts the fee sharply. Second, the number of calls per task matters as much as the price per token. An agent that retries three times on a flaky tool costs three times as much on that step, so reliability engineering is cost engineering.
Caching changes the arithmetic
Most of what an agent sends on each call is the same as the previous call: the system instructions, the tool definitions and the task so far. Both providers offer prompt caching that charges a reduced rate for input the model has already seen; Anthropic lists cache reads at a tenth of the base input price, and both offer batch processing discounts for work that can wait. If 70 percent of the input in the example above is served from cache at a tenth of the price, the mid-tier task falls from about 13 cents to about 7 cents, and 5,000 tasks a month from about USD 640 to about USD 340. Caching is not automatic; the prompts have to be structured so the stable part comes first and stays identical between calls. It is one of the first things we check when a client’s agent bill looks high.
Line two: retrieval and tools
An agent that answers from your documents needs a vector store to search them; an agent that acts needs access to the systems it acts on. Published 2026 guides from development firms put a managed vector database at roughly USD 50 to 500 a month at small and mid scale, and the embedding calls that keep it current are usually small next to the agent’s own calls. The costs that surprise people here are the third-party ones: paid APIs the agent calls on every task (search, enrichment, document parsing) and per-seat licences for the systems it logs into. Price those per task, not per month, because they scale with volume exactly as the model fees do.
Line three: hosting, monitoring and maintenance
The orchestration layer, the queue that holds tasks, the logs that record every call and the dashboard that shows what the agent did all run somewhere. Published estimates put hosting at USD 200 to 1,000 a month and monitoring at USD 50 to 200 a month for a typical deployment (Monocubed, 2026). Maintenance is the bigger number: development firms consistently put it at 15 to 25 percent of the build cost each year, covering prompt and tool updates as models change, re-testing after provider updates, and the work of keeping the knowledge the agent relies on current. For a build in the ranges covered in our guide to what an AI agent costs to build, that is often more than the model fees.
Line four: the human hours
Review is a running cost, not a launch cost
Every agent that does something consequential needs a person checking a sample of its output, forever. Take the 5,000-task example: if one task in ten is reviewed and each review takes three minutes at a fully loaded USD 35 an hour, that is about USD 875 a month, which is more than the mid-tier model fee for the same volume. At 25,000 tasks it is about USD 4,375. The sample rate falls as confidence grows, but it never reaches zero for anything that touches money, customers or compliance, and in regulated settings the review records are themselves a deliverable.
This is where good design pays back. For Pfizer, Appson built an AI platform, to HIPAA and GxP standards, that puts the expertise of senior researchers within reach of every junior researcher; every answer carries a match percentage and a rating of good, average or below average, so reviewers spend their minutes on the answers that need them rather than reading everything. An agent that reports its own confidence honestly is cheaper to run than one that does not, because the review budget goes where the risk is.
A worked monthly budget
Putting the four lines together for the 5,000-task agent on a mid-tier model with caching in place, for illustration only:
| Line | Monthly amount (USD) | Basis |
|---|---|---|
| Model fees, mid-tier, 70 percent cached input | About 340 | List prices, November 12, 2026 |
| Vector store, embeddings and third-party API calls | 100 to 600 | Published 2026 ranges; depends on the tools used |
| Hosting and monitoring | 250 to 1,200 | Published 2026 ranges |
| Human review, 10 percent sample, 3 minutes each | About 875 | Illustrative assumption at USD 35 an hour |
| Maintenance reserve | 15 to 25 percent of build cost a year, spread monthly | Development-firm estimates |
The pattern holds across most deployments we see: the model is rarely the biggest line, people and maintenance are, and the agent’s design decides how large those two become.
Five levers that cut the bill
- Route by step, not by agent. A strong model for planning and final checks, a small model for everything routine.
- Structure prompts for caching. Stable instructions and tool definitions first, identical on every call.
- Cap the loop. A maximum number of calls and retries per task, with a clean hand-off to a person when it is reached.
- Trim the context. Retrieve the passages the step needs, not the whole document; summarise tool outputs before passing them on.
- Measure cost per task from day one. Log tokens, calls and review minutes per task so the budget is managed on numbers, not estimates. Our guide to what an AI pilot should prove puts cost per task on the list of things to measure before funding production.
Frequently asked questions
How much does an AI agent cost to run per month?
For a bounded-task agent handling about 5,000 tasks a month, model fees at November 2026 list prices run from roughly USD 30 on a small model to USD 3,200 on a top-tier model, infrastructure adds a few hundred to around USD 1,700, and human review of a sample typically adds several hundred to a few thousand dollars. Volume, model choice and the review sample rate decide where you land.
Why does an agent use so many more tokens than a chatbot?
Because one task is several model calls, and each call carries the instructions, the task state and everything gathered so far. Eight calls of 6,000 input tokens is 48,000 tokens for a task a chatbot would handle in one call of a few thousand.
Does prompt caching really reduce the cost?
Yes, when the prompts are built for it. Cached input is charged at a fraction of the base price, and in an agent most of each call’s input repeats from the previous call. In the worked example it roughly halves the mid-tier model fee.
Is the model the biggest running cost?
Usually not. Human review and maintenance are typically larger than model fees for a mid-tier agent at moderate volume. The model becomes the largest line only at high volume on a top-tier model without caching.
Should every process get an agent?
No. Many processes are cheaper and more reliable as ordinary workflow automation with a model called only where judgment is needed. Our comparison of AI agents and workflow automation gives a three-question test for deciding.
Budgeting for the agent you will actually run
Price the agent per task across all four lines before it is built, design it so the strong model and the human reviewer are used only where they change the outcome, and measure the real cost per task from the first week. Done that way, the running cost stops being a surprise and becomes a number you can manage down each month.
If you are scoping an agent now, our agentic AI services team can build the per-task cost model with you before any code is written. Talk to our team to start.
Have a project in mind?
Appson Technologies builds AI powered applications, custom software and cloud solutions for businesses in the US and Dubai. Tell us what you are trying to build and we will tell you what is realistic and what it costs.
Get My Free Consultation
