What goes into the monthly bill of an AI agent?
The first line is the AI model itself. It is billed per token, a small chunk of text (often part of a word), in two columns: what the model reads (your instructions, the email, the quote, the conversation so far) and what it writes. At Anthropic, output costs five times more than input: $40 versus $8 per million tokens for Fable 5.2.
Then come the tools the agent uses. A web search costs $10 per 1,000 searches on Anthropic's API, plus the text of the results, which is billed as input. Third-party services plugged into the agent are billed per call, as French business site Journal du Net points out.
Last, there is hosting (the server the agent runs on, the database that keeps its logs) and oversight time: someone reviews what goes out, handles edge cases and adjusts the instructions. Both depend on your setup. Ask for them as separate lines in the quote.
How much does it cost to run an AI agent with each model?
For the same workload, the model you pick can change the price per token by a factor of 80 within a single provider. Here are Anthropic's public prices as of 10 November 2026, per million tokens (input / output):
- Claude Fable 5.2: $8 / $40
- Claude Opus 5.5: $4 / $20
- Claude Sonnet 5.5: $2 / $10
- Claude Haiku 5.5: $0.10 / $0.50 (for requests under 100,000 tokens; above that, $0.50 / $2.50)
Fable 5.1, the previous version of the top model, cost $10 per million tokens read and $50 per million written. Anthropic's own advice is Haiku for simple tasks, Sonnet for most production workloads and Opus for the most complex reasoning.
Price per token is not enough to compare models. On the same set of tests, Artificial Analysis, quoted by Journal du Net, measured a cost of nearly $27 for Fable 5.2 against around $10 for Opus 5.5. That gap, almost threefold, is wider than the twofold gap in token prices. Journal du Net recommends keeping Fable 5.2 for complex tasks that other models fail.
So the useful question for a small business is: which is the cheapest model that does this specific task well? You find out by testing two or three models on your real emails and quotes before you commit.
Why does the bill change from one month to the next?
Because it follows your activity. A month with twice as many requests means twice as much text read and written. Overruns are common: in an IDC survey quoted by Journal du Net, 67% of companies said they had exceeded their compute and model budget for agents by more than 10% over twelve months.
Some causes stay hidden until the invoice arrives. Journal du Net describes a prospecting agent sold on a flat fee that started looping overnight after its instructions were changed. Its cost per lead went up fourfold, and the vendor only found out from the monthly bill.
Conversation length matters too. On Anthropic's API, web search results are counted again as input on every later turn of the conversation. The longer an exchange runs, the more each new answer costs.
How can you cap what an AI agent spends?
Set the cap before the agent goes live. Journal du Net suggests a budget per job: €100 to enrich 200 leads, or 50 cents a lead. If the cost climbs to €2 a lead, the budget stops at 50 leads instead of letting spending run on.
- One budget per agent and per job, with a ceiling and a named owner. When it runs out, the agent stops and alerts someone.
- A limit on tool calls. Anthropic's API, for example, lets you cap the number of web searches per turn.
- Caching for repeated text. At Anthropic, reading text that is already cached generally costs 0.1 times the normal input price. Fable 5.2 has a reduced rate of 0.025 times: $0.20 per million tokens instead of $8. Writing to a 5-minute cache the first time costs 1.25 times the normal price ($10 for Fable 5.2). This helps when the agent rereads the same instructions or product catalogue on every request.
- Batch processing for anything that can wait. On Anthropic's API, batched requests cost 50% less, and most batches finish within an hour. A monthly report or sorting yesterday's documents are good candidates.
- The right model for each step: a cheaper model to sort and classify, a more advanced one only for the hard cases.
- A monthly usage report per agent, set against the work done: emails handled, quotes drafted. Cost per case handled tells you whether the agent pays for itself.
What should you ask a developer before launching an agent?
Ask for two separate figures: the build price, and an estimate of the monthly running cost with the volumes it assumes. Also ask whose account the model usage is billed to (yours or the developer's) and how you will track it.
At NUOPS, every AI integration project starts with a limited scope, human approval on anything that leaves your company and full logging of what the agent does. The model choice (Anthropic, OpenAI or Mistral, hosted in Europe when required) is explained in the quote. You get a ballpark figure on the first 30-minute call, then a firm quote within 5 days.