
An AI agent that answers well in a demo can become expensive the moment it runs every day. The initial quote covers the build. It says almost nothing about what you will pay in month three, once conversations have doubled and the model has moved to a new version.
So the question is not "how much does an AI agent cost", but "which lines come back every month, and which one drifts first". Here is the breakdown, with public supplier prices to anchor the orders of magnitude.
Key Takeaways
- An agent in production is paid across four lines: tokens, infrastructure, human supervision, evaluations and security. Only the first is truly variable.
- Tokens are usually the most visible line and often the lightest: at Anthropic, Sonnet 5.5 bills $2 per million input tokens and $10 per million output tokens, Haiku 5.5 costs ten to twenty times less.
- Human supervision is the line that tips a budget: a single engineer day per month, at the market's average day rate, weighs more than millions of tokens.
- An agent's cost is not read per month but per conversation handled: the only ratio that stays comparable when volume changes.
- An agent without automated evaluations looks cheaper and costs more, because regression is discovered at the customer's desk.
The four lines of a production invoice
An agent in production consumes four distinct things. Confusing them is the first cause of a wrong budget.
| Line | Nature | What makes it vary |
|---|---|---|
| Tokens (model API) | Variable, proportional to usage | Conversation length, context size sent, model chosen |
| Infrastructure | Fixed, monthly | Server, database, queue, test environments |
| Human supervision | Semi-fixed | Engineer days per month, criticality level |
| Evaluations and security | Semi-fixed | Test frequency, logging, access review |
An internal agent used by five people and a customer-facing agent exposed to a thousand conversations a day do not share the same cost profile, even when they run on the same model.
Line 1: tokens, the only truly variable cost
This is the line everyone watches, and the easiest to price, because suppliers publish rates per million tokens (MTok).
At Anthropic, according to the pricing page consulted on 18 November 2026, Sonnet 5.5 is billed $2 per million input tokens and $10 per million output tokens. Haiku 5.5, presented as the fastest and most cost-efficient model, drops to $0.10 input and $0.50 output for requests of 100K tokens or less. Opus 5.5 sits at $4 input and $20 output. Batch processing is billed at half price, and reading the prompt cache costs $0.10 per million tokens on Sonnet 5.5.
The arithmetic takes three lines. Take an agent handling 20,000 conversations a month, with 2,000 input tokens and 300 output tokens per conversation:
- input: 20,000 × 2,000 = 40 million tokens, so 40 × $2 = $80;
- output: 20,000 × 300 = 6 million tokens, so 6 × $10 = $60;
- total: $140 per month on Sonnet 5.5, excluding cache and batches.
The same volume on Haiku 5.5 costs $4 in and $3 out, so $7. On Opus 5.5, $160 in and $120 out, so $280. The gap between the cheapest and the most expensive model is therefore a factor of forty at identical volume. That is where the first architecture decision happens, not in the number of conversations.
Two effects push this line up without any change in volume. The first is context growth: an agent that reloads the full history and documents on every turn pays several times for the same information. The second is verbosity: asking for a structured answer in three paragraphs rather than ten lines halves the output bill, which is the expensive side.
Line 2: infrastructure, a fixed cost that is often underestimated
An agent is not just an API call. It needs somewhere to run, a database for conversations and task state, a queue when processing is slow, logs, and at least two environments so testing does not happen in production.
This line is fixed: it barely depends on the number of users as long as the agent stays within a reasonable range. A single server is often enough to start, with a managed database and a backup. The bill counts in tens of euros per month, to be checked on your host's pricing page, because prices move and depend on the region.
Three traps make this line more expensive than announced. The first is the forgotten staging environment: it runs all the time for occasional use. The second is an oversized database kept for comfort. The third is the absence of rate limiting: an agent without a ceiling can be called in a loop by a script and blow up the token line, not the server line.
An agent that must stay available during business hours does not need the same redundancy as a banking service. Sizing infrastructure to the real criticality level often saves more than switching models.
Line 3: human supervision, the line nobody budgets
An agent in production is watched. Someone has to read the cases where it got things wrong, adjust instructions, fix the tools it calls, arbitrate edge cases. This work does not disappear with model quality: it moves to the rare situations, which are precisely the ones that cost the most when they happen.
What does it represent? One engineer day per month is a common order of magnitude for a stable business agent. According to the Free-Work IT salary barometer, consulted on 18 November 2026, the average day rate for an engineering role is €417 direct and €435 through an intermediary. One day per month at €417 is €5,004 per year, thirty-six times the token cost in the previous example.
That ratio surprises leadership teams who compared offers on build price. An agent built for €15,000 and supervised one day per month costs, in the first year, almost as much in supervision as in design. And if the agent touches a critical process, supervision moves to two or three days per month, with on-call cover.
The maintenance of an AI agent therefore deserves to be contracted from the start, with a committed number of days and a response time. Without it, supervision becomes a diffuse cost nobody tracks.
Line 4: evaluations, observability and security
This line covers everything that tells you whether the agent still works. It has three parts: a test set replayed on every model or instruction change, logging of calls and decisions, and a review of access to the data being used.
It costs little in infrastructure and a lot in time at the start. Automated evaluations require building a set of real cases, often about a hundred, with the expected answer. That is an investment of a few days, then a negligible running cost: replaying a hundred cases on a small model costs a few cents.
The alternative, evaluating nothing, does not remove the cost. It moves it to late detection: a regression found by a customer costs reputation, an emergency fix, sometimes a refund. That is why evaluating a demo before signing is not enough; you also need to check that the control mechanism exists after go-live.
Security adds a modest but non-zero line: API key management, data segregation per customer, access logging. On an agent handling personal data, this line belongs in the budget and not only in the documentation.
Subscription or API: two cost curves
Many teams start with a per-seat subscription, then move to the API when the agent becomes a service. The two curves do not cross at the same point depending on usage.
According to OpenAI's pricing page consulted on 18 November 2026, a standard ChatGPT Business seat costs $20 per month on annual billing ($25 monthly) and a premium seat $100 per month on annual billing. For a team of ten, that is $200 per month, a fixed and predictable amount.
The API follows consumption. For a low-volume agent, it costs less than a subscription. For an agent heavily used by a large team, the subscription can remain cheaper, provided you accept usage limits. The right question is not which of the two is cheaper, but which curve stays under control when usage triples.
One practical point: a subscription does not give you an API. An agent that must plug into your management software needs API access, and therefore a consumption budget, however modest.
How to price your case in an hour
Three numbers are enough to reach a defensible order of magnitude in front of an executive committee.
- Volume: number of conversations or tasks per month, at peak and on average. Peak drives infrastructure, average drives tokens.
- Exchange size: input and output tokens per exchange, measured on twenty real cases. Eyeballed estimates are almost always too low on input, because the context sent on every turn gets forgotten.
- Target model: an economical model for triage and simple answers, a stronger model for complex cases. An agent that routes between two models often costs less than a well-tuned single-model agent.
Multiply, add infrastructure, then add supervision in days per month. The result is not an accounting forecast; it is an order of magnitude that lets you compare two scenarios and spot the dominant line. If you want a framework to compare two commercial proposals, ask each supplier for the full monthly cost, supervision included, and not only the build price.
FAQ
How much does an AI agent cost per month, roughly?
It depends almost entirely on volume and supervision. A low-volume internal agent on an economical model can stay under $20 of tokens per month. A customer-facing agent at 20,000 monthly conversations on a mid-range model runs around $140 of tokens, at Anthropic's November 2026 rates, plus infrastructure and supervision.
Can token costs drift without volume increasing?
Yes, through two effects. Sending a long context on every turn multiplies input tokens, and longer answers hit the output side, billed five times more than input at Anthropic. A token ceiling per request and an output format instruction are the two most profitable guardrails.
Should you prefer a small model to cut the bill?
Not systematically. A model that is too weak produces answers that need fixing, and fixing costs engineer days, a far more expensive line than tokens. The right practice is routing: an economical model for simple cases, a stronger model for cases identified as hard.
Is human supervision billed as a flat fee or by time?
Both exist, and the flat fee often hides a low number of days. Ask for the number of days included, the incident procedure and the rate for an extra day. A contract that mentions no volume of days does not frame supervision, it defers it.
What happens if the supplier changes its prices?
Pricing pages state explicitly that prices may change. An agent whose cost depends on a single supplier and a single model therefore carries a budget risk. Building an abstraction layer that allows a model change reduces that risk, at the price of a little integration work.
Conclusion
The cost of an AI agent in production is not just an API bill. It reads across four lines, two of which are fixed and one of which is human. Tokens are the most visible part and often the lightest: in our examples, a single day of supervision per month weighs more than millions of tokens.
The method that avoids nasty surprises comes down to three numbers: real volume, measured exchange size, model suited to the case. The rest is a matter of contract and control, not rates.
If you are preparing a project and want to check what your quote really covers, or if you are looking to have an agent designed that holds up in production, let's talk: the contact page is open, and we can also start with an audit of what already exists before deciding what to build.


