How Much Does It Cost to Run an AI Agent? Tokens, Tools, Hosting and People

The cost of running an AI agent is the model tokens it uses, plus tool and API fees, hosting, monitoring and, usually the biggest line, the human time spent reviewing its work. The useful way to measure it is cost per completed task, not cost per month, because the monthly bill only makes sense once you know how many tasks the agent does and how much each one costs.

This guide breaks the cost into its parts, works through a token example with the arithmetic shown, and lists the levers that bring the number down. Prices in this post come from Anthropic's published API pricing as of November 2026. Prices change, so check the vendor page before you plan a budget. Building the agent is a separate, one-off cost: it depends on scope and is quoted after a call, so it is not covered here.

What are the parts of an agent's running cost?

Most first-time buyers put all their attention on the first row. In practice the people line is often larger than the token line, and the maintenance line is the one that surprises.

How do model tokens get priced?

Models are billed by the million tokens ("MTok"), with different rates for input (everything you send, including instructions, tool definitions, retrieved documents and conversation history) and output (what the model writes). A token is roughly a short word or word-fragment. Output tokens cost several times more than input tokens.

An agent is more expensive than a single chat reply because it works in a loop: read the task, choose a tool, read the tool result, choose the next step, and so on. Each turn of the loop sends the growing context again. Anthropic's engineers report that agents typically use about four times the tokens of a chat interaction, and multi-agent setups about fifteen times, which is why they recommend reserving the heavier designs for tasks valuable enough to justify the spend.

Current Anthropic list prices (November 2026)

Last verified November 2026 on Anthropic's pricing page. Haiku 5.5 charges more for requests whose prompt exceeds 100,000 tokens, and other models and tiers have their own rates, so read the page rather than relying on this table. Anthropic also notes that its newer tokenizer (Claude 4.7 and later) produces roughly 30% more tokens for the same text than older models, so a token count measured on one model does not carry over exactly to another. These are API list prices; other vendors, and cloud resellers, price differently.

What does one task actually cost? A worked example

This is an illustrative calculation to show the method. The numbers are assumptions, not measurements of any real agent, and your own usage will differ.

The assumed task: an agent triages one inbound support email. It makes 4 model calls (read the email, look up the order, check policy, write a draft). Each call sends about 6,000 input tokens: roughly 3,500 of fixed material (instructions and tool definitions) and 2,500 of task-specific content. Each call writes about 400 output tokens.

Totals per task: 24,000 input tokens and 1,600 output tokens.

The caching row assumes a warm cache, meaning the same prefix was used within the cache lifetime. Writing to the cache costs more than a normal read (1.25 times base input price for the default 5-minute cache), so a prefix that is rarely reused saves less than shown.

The spread is the lesson. The same task costs from a fraction of a cent to about thirteen cents depending on the model tier and prompt design, a gap of roughly forty times between the cheapest and dearest rows. Whether the cheaper model is good enough is an evaluation question: run both on your real cases and compare, as covered in how to test and evaluate AI agents.

From per-task to per-month

Take 5,000 tasks a month, still illustrative:

• All on Sonnet 5.5, uncached: 5,000 × $0.064 = $320 in tokens.
• Routed so 70% of easy tasks go to Haiku 5.5 and 30% to Sonnet 5.5, uncached: 5,000 × (0.7 × $0.0032 + 0.3 × $0.064) = 5,000 × $0.02144 = $107.

Now add the human line. If a person reviews 10% of tasks at 2 minutes each, that is 500 reviews, about 16.7 hours a month. At an assumed loaded cost of $20 an hour, that is about $333, which exceeds the whole token bill of the cheaper setup. Change the review rate from 10% to 5% and you save more than any model discount will. This is why cost per task, including review time, is the figure to track.

Which levers lower the cost?

Prompt caching

If every call starts with the same instructions, tool definitions or reference documents, the API can read that prefix from cache at a fraction of the input price. Anthropic's prompt caching docs and rates make a cache read 10% of base input on most models and 5% on Sonnet 5.5 and Opus 5.5, and the cache lasts five minutes by default, refreshed each time it is used, with a one-hour option at a higher write price. There are minimum prompt lengths for caching to apply (512 tokens for Sonnet 5.5 and Haiku 5.5, more for some other models). Changing tool definitions invalidates the cache, so keep the stable parts stable.

Model-tier routing

Not every step needs the most capable model. Anthropic's guide to building agents describes routing: classify the input, send easy requests to a cheaper model and hard ones to a stronger one. A small model can handle classification, extraction and simple lookups, with the larger one reserved for judgement-heavy steps.

Batch processing

For work that does not need an immediate answer, such as overnight document processing or running an evaluation set, Anthropic's Batch API gives a 50% discount on input and output tokens. Anthropic says most batches finish within an hour, but results are not immediate, so it suits back-office jobs and not live conversations.

Shorter loops and tighter context

• Cap the number of steps per task, and stop the run when it is exceeded. A looping agent is the most common cause of a surprise bill.
• Send only the fields needed from a record, not the whole thing.
• Summarise long histories instead of resending them.
• Use ordinary code for anything that does not need a model: arithmetic, formatting, validation, date logic.

Do not default to many agents

Splitting work across several cooperating agents multiplies token use. It can be worth it, but only when the task needs it; see single-agent vs multi-agent systems.

What other fees show up in the bill?

• Paid model tools. Anthropic's API charges for its web search tool per search, listed at $10 per 1,000 searches as of November 2026, on top of the tokens the results add. Code execution can carry runtime charges when used on its own. Check the current pricing page for each tool you enable.
• Data residency. Specifying US-only inference on Anthropic's API applies a 1.1 times multiplier to token prices, per its pricing page. Location requirements can therefore affect cost.
• Workflow platforms. If the agent runs inside an automation tool, you may pay per execution or per task there as well. We compare the models in n8n vs Make pricing; read the vendor's current plan page before budgeting.
• Third-party data and messaging. Enrichment lookups, SMS, WhatsApp or voice minutes are billed by their own providers.
• Hosting. A small agent can run on modest infrastructure; costs grow with queues, databases, logs and uptime requirements. The amount depends on your setup and provider, so we do not quote a figure here.

Why is cost per task better than cost per month?

A monthly figure hides the two things you can control. If volume doubles, the monthly cost doubles but the per-task cost stays the same, so you cannot tell whether the agent got worse or just got busier. And a per-month figure cannot be compared with the alternative, a person doing the job, which is naturally measured per item.

Track four numbers for each task type: model and tool cost, review minutes, rework (how often a human had to fix it) and the agent's success rate from your evals. A task that costs 6 cents in tokens but fails 30% of the time and needs a 3-minute fix is not a 6-cent task. Put the whole thing against the cost of handling the same item manually, then decide.

It also tells you what not to automate. If a task is rare, high-stakes and needs heavy review, the numbers may favour a person. Our post on AI agents, chatbots and automations helps with the earlier question of whether you need an agent at all.

How do you build a cost estimate for your own agent?

1. Pick one task type and collect 20 real examples.
2. Run a prototype on them and log input tokens, output tokens, tool calls and steps for each.
3. Price it with the vendor's current page, for two or three model tiers.
4. Add the people line: review rate times minutes times hourly cost.
5. Add fixed costs: hosting, monitoring, a maintenance allowance.
6. Apply a margin for variance. Agents have long tails: a few tasks use ten times the average. Set a per-task ceiling.
7. Re-measure after a month of live use. Real traffic is rarely like the sample.

That gives you a per-task cost with a range, which is a sounder basis for a go or no-go than a headline subscription number. The one-off cost of building the agent depends on scope, so we scope it with you on a call and quote it, rather than publishing a price list. If you want that done around a real process, see our agentic AI development and n8n automation work.

Frequently asked questions

How much does it cost to run an AI agent per month?

It depends on tasks per month times cost per task, plus fixed costs and review time. In the illustrative example above, 5,000 triage tasks came to between roughly $107 and $320 in model tokens depending on model choice, before review time, hosting and maintenance. Your real figure needs a measured sample.

Is the model usually the biggest cost?

Not always. For tasks that need human review, the people time often exceeds the token bill. For high-volume, fully automated tasks that use a large model and long contexts, tokens can dominate.

Does prompt caching really help agents?

Yes when calls repeat a long, stable prefix such as instructions and tool definitions, which agents usually do. Savings are smaller if the prefix changes often or is used too rarely to stay in the cache.

Can a cheaper model do the job?

For many narrow steps, yes: classification, extraction and routine drafting. Test it on your own examples with an eval set rather than assuming, and route only the harder steps to a larger model.

Are the prices in this post fixed?

No. They are Anthropic's published list prices as of November 2026. Model names, tiers and rates change, so check the vendor's pricing page before budgeting.