AI agent cost per month comes down to four line items, and three of them have a published price per unit: input tokens, output tokens, and the time the session is running. The fourth appears on no price list and is often the largest one: the time your own people spend reviewing what the agent did.

This is for a founder or director who wants a number that belongs to their own operation instead of a range from a vendor blog. Every price below comes from Anthropic’s price list, retrieved on 18 October 2026, and stays in dollars as the source publishes it.

What does an AI agent cost per month?

The bill has these line items. You fill in two numbers yourself: runs per month, and your own hourly cost.

Line itemPrice per unit (price list 18-10-2026)Your volume per monthYour amount
Input tokens (the instruction, the context, the tool results)$4 per million on Claude Opus 5.5, $2 on Sonnet 5.5, $0.10 on Haiku 5.5 for prompts up to 100,000 tokens[1]
Output tokens (what the agent writes and decides)$20 per million on Opus 5.5, $10 on Sonnet 5.5, $0.50 on Haiku 5.5[1]
Session runtime (time the session is running)$0.08 per session-hour[1]
Review time of your own peopleyour own fully loaded hourly cost
Searches, if the agent uses the web$10 per 1,000 searches, on top of tokens[1]

Note the third row. Runtime accrues only while the session has status running; time spent idle, which means waiting for your next message or for a tool confirmation, does not count[1]. An agent that sits switched on all day without doing anything costs nothing in runtime.

Why does an agent consume more than a chatbot?

Because every tool call is another trip through the model. A chatbot takes a question and returns an answer. An agent takes a question, decides it needs to call a system, receives a result, and that result goes back into the model as new input. The documentation calls that a tool round trip, and names the case where it is not needed: questions the model can answer from training alone, such as summarising or translating[2].

The fact that the agent makes that decision is the point. With the default setting the model determines on each turn whether to call a tool or respond directly, and it calls one when the request maps to that tool’s described capability and the answer is not already in context[3]. Tool use is a contract: your application specifies which operations exist and what shape their inputs and outputs take, and the model decides when and how to call them[2]. It is designed as the bridge between a natural-language request and the system that fulfils it, such as a database or an internal API[2].

Three consequences for the bill. Every tool result is input you pay for. The more steps a task has, the more round trips. And an agent should only exist where the work is too open-ended for a fixed workflow[4]: if the order of steps is fixed, a workflow is cheaper, because it has nothing to decide. The difference between the two is covered in AI agent vs chatbot, and what an agent actually is in what is an AI agent.

How do you get from cost per run to cost per month?

The documentation works one session out line by line, and that is the shape to copy. The example is a one-hour session on Claude Opus 5.5 consuming 50,000 input tokens and 15,000 output tokens, with 40,000 of the input tokens served from cache[1]:

Line itemCalculationCost
Uncached input tokens10,000 x $4 / 1,000,000$0.04
Cache read tokens40,000 x $4 x 0.05 / 1,000,000$0.008
Output tokens15,000 x $20 / 1,000,000$0.30
Session runtime1.0 hour x $0.08$0.08
Total$0.428

Without caching the same session lands at $0.58: input tokens 50,000 x $4 / 1,000,000 = $0.20, plus $0.30 output, plus $0.08 runtime[1]. The 0.05 factor in the second row is not a rounding: on Claude Opus 5.5 and Sonnet 5.5 a cache hit costs 5 percent of the standard input price, where other models use 10 percent[1].

From here to a month is one multiplication. If the agent runs 200 times a month, that is 200 x $0.428, so roughly $85.60 in model and runtime. At 2,000 runs it is ten times that. The number of runs is the only assumption in that figure, and it is one you can look up rather than guess: check how often the process ran last month.

What you will not find in this calculation is a price per user or a monthly subscription. That is deliberate. An agent on the API is billed on consumption, and anyone budgeting seats for a team is answering a different question.

What does your own review time cost?

This is the fourth line item and no vendor publishes it. Work it out like this: minutes of review per run, times the share of runs you review, times your own fully loaded hourly cost.

An example of the shape, with your numbers to fill in: 200 runs a month, reviewing half of them in the early months, at 4 minutes each, is a little over 6 hours of work. Put your hourly cost next to that and compare it with the $85.60 in tokens and runtime. In most first months, review time is the biggest number on the bill.

WeAdapt quotes no average hourly rate and no average review time here, because no source makes those numbers hard. What is established is why the line item exists: the agent decides for itself when to call a tool[3], and that discretion is exactly why somebody watches until the pattern is known. Strike that line and you have priced an agent nobody checks.

Setting this against the return, using the hours the process costs today, is covered in calculate the ROI of AI automation. The four line items of a whole project, including building and maintenance, are in how much AI automation costs. This article is only about the monthly bill of one agent.

Why a cheaper model is not always cheaper

The price list puts Haiku 5.5 at $0.10 per million input tokens and Opus 5.5 at $4[1]. That looks like an easy decision, and there is a catch that rarely gets mentioned: how many tokens the same text produces depends on the model’s tokenizer, the language and the type of content[1]. Claude 4.7 and later models use a newer tokenizer that produces approximately 30 percent more tokens for the same text than earlier models[1].

Two practical consequences. Comparing two models on price per million tokens is not a comparison of your bill, because the token count differs per model. And if you extrapolate a pilot from one model to another, your token count no longer holds. Count them again, with token counting on the model you intend to use[1].

The documentation keeps its own advice simple: Haiku for simple tasks, Sonnet for most production workloads, Opus for the most complex reasoning[1]. Which model belongs at which step of your agent is the kind of choice WeAdapt makes per step when building AI agents, rather than once for the whole system.

What lowers the bill, briefly

Two things are published as multipliers. The Batch API processes large volumes of requests asynchronously at a 50 percent discount on both input and output tokens[1], and a cache hit costs 10 percent of the standard input price, or 5 percent on Opus 5.5 and Sonnet 5.5[1]. Those multipliers stack with other pricing modifiers, and batch and caching discounts can be combined[1].

One caveat specific to agents: the batch discount does not apply to managed agent sessions, because sessions are stateful and interactive[1]. Caching does.

How to lower the bill systematically is its own question with its own trade-offs, and it does not belong in an article about what something costs. Want to know what your process will cost per month before anything gets built? Book a call, or read in the FAQ what a project includes. WeAdapt builds custom AI tools and works the monthly bill out up front with your volume, not with a range.

Sources

  1. Anthropic, Pricing (platform.claude.com, accessed 18 October 2026)
  2. Anthropic, How tool use works (platform.claude.com, accessed 16 September 2026)
  3. Anthropic, Tool use (platform.claude.com, accessed 20 September 2026)
  4. n8n, Build and manage agents (github.com, accessed 20 September 2026)