How much does an AI agent really cost per month?

An AI agent’s monthly cost includes any subscription, model usage, connected tools, hosting, and human oversight. There is no universal flat rate: in the detailed simulation below, model calls cost $14 for 1,000 simple tasks, before other expenses. To plan your budget, start with the cost per accepted task and your expected monthly workload. Price depends first on the model: see for example the detailed rates of Mistral AI.

What the advertised price actually includes

Providers do not all charge for the same unit. Some sell access per user; others charge for automation runs, credits, or an outcome defined in the contract. A monthly price alone is therefore not enough to compare two solutions.

Before choosing from the AI tools organized by profession, list what is included and what you will need to pay for separately. A plan may cover the interface and hosting while leaving some external usage on another bill.

Billing model	Unit to track	Additional expense to check	Limit to examine
Individual subscription	User and billing period	Overages, add-ons, external access	Quotas and available features
Automation platform	Execution or credit	Connected model and third-party services	How retries and subprocesses are counted
Agent built on an API	Tokens and tools used	Hosting, storage, maintenance	Rate limits and technical caps
Business agent billed by outcome	Billable event	Human agent seats, channels, and integration	Contractual definition of an outcome

A token is a unit of text processed by the model, rather than necessarily a whole word. Your budget depends on the text sent and generated, including the documents, instructions, and tool results the agent handles.

The difference between an AI agent and a chatbot explains part of the cost gap: a request to an agent may trigger several consecutive calls. Count those calls across the entire task instead of measuring only the final response.

Calculating model usage costs

For an API billed by text volume, the basic calculation is straightforward:

Model cost = millions of billed input tokens × input rate + millions of billed output tokens × output rate.

Then add any separately priced categories, such as cache reads and writes, paid tools, or specialized processing. Use the usage counters returned by the provider, because the length of the visible text does not always capture everything that is billed.

Take Claude Sonnet 5.5. As of October 24, 2026, the official documentation lists $2 per million input tokens and $10 per million output tokens, excluding caching and optional features. These prices serve only as the basis for the following simulations; they say nothing about the quality you will achieve for your use case. Source: official pricing documentation.

Calculation assumption, not a field measurement: a task uses a total of 2,000 input tokens and 1,000 output tokens across all calls. That costs $0.004 for input and $0.01 for output, or $0.014 in total. For 1,000 identical tasks per month, the model cost is $14.

If each task instead uses 20,000 input tokens and 5,000 output tokens, the same calculation gives $90 for 1,000 tasks. These volumes are illustrative assumptions; replace them with your observed usage.

The LLM pricing calculator lets you adjust volumes and rates. Keep your estimate in US dollars when that is the provider’s billing currency, then record any amount actually charged in euros separately.

Three budgets for different uses

Personal use, internal automation, and a public-facing customer service operation have different constraints. The first depends mainly on your own activity; the others must also accommodate incoming requests, incidents, and service continuity.

For individual use, start with the plan that provides the features you need. Check whether the agent draws on included usage or an external account. Add the time spent preparing requests and reviewing results: a small software bill can still come with substantial rework.

For internal automation, separate the platform cost from AI processing. As of October 24, 2026, n8n lists its Starter plan at €20 per month, billed annually, with 2,500 monthly executions. That is the published plan price, rather than an estimate of your agent’s total cost. Source: official pricing page.

The page also lists AI features and included credits. Examine what they cover before adding an API budget: identify which work uses the included capabilities and which work runs through your own paid connections. A workflow execution, by itself, is not a measure of token usage.

For customer service, Fin is listed from $0.99 per billable outcome as of October 24, 2026. Assuming 500 outcomes that actually qualify for that rate gives $495, before other applicable charges. The provider may count certain handoffs after a procedure as outcomes, so check the definition that applies to your contract. Source: official pricing page.

These examples illustrate different ways of calculating costs. They are neither a market price range nor a purchase recommendation.

Adding setup, hosting, and human work

A complete budget should distinguish setup from ongoing operations. Connecting software, preparing data, and defining approval steps are setup activities. Reviewing results, resolving incidents, and adapting connections are part of day-to-day operation.

To spread the initial investment across your budget, choose an allocation period. Entirely hypothetical example: €600 in configuration costs spread over 12 months amounts to €50 per month. This accounting calculation does not change the upfront payment; it helps you compare different ways of completing the same task.

Human time deserves its own line item. With an internal assumption of 2 hours of monthly oversight valued at €40 per hour, supervision costs €80. That hourly rate is a simulation parameter, rather than an observed professional service rate.

Hosting costs depend on what you run. Hosting the program that orchestrates calls to an API is different from running the model on your own hardware. In the latter case, include the purchase or rental of the machine, its usage costs, electricity, and administration.

Remember backups, logs, and test environments as well. Even when they use existing infrastructure, assign them a reasonable share of its cost if you want to calculate the full economic cost.

The guide to AI agents in business helps define responsibilities for a pilot. The person paying for the subscription is not necessarily the person maintaining connections or approving actions.

What makes the bill grow

A lengthy task can cost more than a series of small, clearly bounded requests. With each call, the agent sends context that may include instructions, conversation history, and previous results. Without context management, the same information can be billed repeatedly.

Tools are another expense. The pricing documentation distinguishes tokens associated with tool use from any separate fees for tools executed on the provider’s side. For your budget, request both pieces of information: the number of external calls and their effect on the context sent to the model.

Retries matter too. An unreadable document, an expired connection, or an unusable response can trigger another attempt. If you count only successful tasks, you hide the usage consumed by attempts that were eventually abandoned.

Prompt caching can reduce the processing cost of repeated content under certain conditions. The official documentation distinguishes cache writes from cache reuse; the savings depend on how stable the content is and whether it is actually reused. Do not deduct theoretical savings before you see them in the usage counters. Source: how prompt caching works.

Finally, adding more agents can increase the number of calls, exchanges, and checks. Require a measurable improvement in results before retaining that complexity. The guide to reducing AI agent costs covers the options to examine after taking those measurements.

Measuring cost per accepted task

The useful metric is total operating cost divided by the number of tasks accepted under your criteria. A task the software marks as finished may still be incomplete for your business: it could have the wrong recipient, a missing document, or an answer that needs rewriting.

Define acceptance before starting the pilot. For invoice extraction, that could mean the required fields are correctly extracted and verified. For document research, it could mean an answer supported by authorized documents, with the necessary references.

Keep a record of:

• The number of requests received and their categories.
• Model and external service usage.
• Failures, retries, and abandoned attempts.
• Time spent reviewing and correcting results.
• The number of results ultimately accepted.

Simulation separate from provider pricing: if operations cost €120 and produce 300 accepted results, the unit cost is €0.40. With the same spending and only 150 accepted results, it rises to €0.80. Quality therefore changes the cost of useful output, even when the bill stays the same.

Then compare this with the existing process, accounting for the time actually saved and the work required to handle exceptions. For customer requests, the article on customer service AI agents adds the relevant operational perspective.

Setting a budget and stopping rules

Build a baseline estimate and a high-activity scenario using your own data. Vary request volume, request length, retries, and review time. Present those assumptions alongside the total: a number on its own can create a misleading impression of precision.

Also distinguish alerts from hard limits. A usage notification does not guarantee that calls will stop. Check which controls are available, then set appropriate limits in the application: maximum duration, number of attempts, permitted actions, and a way to interrupt tasks that have stopped making progress.

Decide what happens when the budget is exhausted. Requests can wait, be handed to a person, or be declined with a clear explanation. That operational decision belongs in the budget because it determines how much work remains to be handled elsewhere.

Your monthly budget should bring together subscriptions, usage, infrastructure, maintenance, and oversight. Keep setup on a separate line, including any allocation across later months. You can then judge an AI agent’s price against a service that has actually been delivered and checked, with identifiable costs.

Frequently asked questions

Can you use an AI agent for free?

A free plan can help you evaluate a use case, but it does not guarantee enough capacity or continued service. For software you run yourself, assess model usage, hardware, and maintenance costs separately. A free download does not mean a zero-cost operation.

Does an AI subscription include API calls?

That depends on the contract and how the agent connects to the model. Check which account supplies the model and where its usage is billed. Adding an API key to an application can generate charges separate from the application's subscription.

How do you request a quote for a business AI agent?

Provide your expected workload, the software you need to connect, the actions the agent may take, and your acceptance criteria. Ask for setup, operations, usage, and maintenance to be itemized separately. Have the provider specify overage charges, contractual commitments, and exit terms.

Should you choose monthly or annual billing?

An annual commitment may lower the advertised monthly price, but it locks in spending before your usage has settled. Compare the total payable over the term and the cancellation conditions. For a pilot, contractual flexibility matters as much as price.