How much does the OpenAI, Mistral or Claude API cost for an SME?

Direct answer: for typical SME usage, a RAG chatbot on your documents or an automation agent, the monthly API bill comes to tens of euros, not thousands. Often even single-digit euros. The real cost item of an AI project is building the system, not calling the model. And the choice between OpenAI, Mistral and Claude weighs less on the final bill than the architecture around them.

I spend my week running this kind of system in production, on BeForBuild.com and at my clients'. Here are the public prices verified in mid-2026, a concrete calculation for an SME RAG, and the criteria that really matter when choosing.

In short

• 1 million tokens represents about 750,000 words. Input (your question + the context) and output (the answer) are billed separately, with output 3 to 6 times more expensive.
• Frontier models: on the order of $5 for input and $25 to $30 for output per million tokens (GPT-5.5, Claude Opus). Intermediate: $2 to $3 / $6 to $15 (Mistral Large, Claude Sonnet). Small models: under $1 (Mistral Small, Claude Haiku for input, Ministral).
• An SME RAG at 1,000 questions/month: under 1 euro with a small model, around 10 euros with an intermediate, around 18 euros with a frontier.
• An automation agent consumes more tokens per task, but stays within the tens of euros per month for SME volume.
• The cost that matters is the build: a RAG starts at 2 640 € excl. tax, an agent between 440 € and 1 760 € excl. tax. The API, behind it, is a marginal cost.

How an AI API is billed

All language model APIs bill per token, the unit of text the model reads and writes. A token is about three quarters of a word: 1 million tokens ≈ 750,000 words, the equivalent of about ten novels. You pay two distinct counters: input tokens (your question, plus all the context the system sends to the model: document excerpts, instructions, history) and output tokens (the generated answer), the latter 3 to 6 times more expensive. In a RAG, input dominates by far: you send a lot of context to get a short answer.

The orders of magnitude by tier, public prices in mid-2026 (prices move, check the official pages cited in the sources):

• Frontier models (the high end): GPT-5.5 is listed at $5 per million tokens for input and $30 for output, Claude Opus at $5 / $25. It is the most expensive tier, and the least often necessary.
• Intermediate models: Claude Sonnet at $3 / $15, Mistral Large at $2 / $6, GPT-5.4 at $2.50 / $15. This is the workhorse tier for production: more than capable enough to answer from provided documents.
• Small models: Mistral Small at $0.10 / $0.30, Claude Haiku at $1 / $5, Ministral under $0.10. For bounded tasks (classification, extraction, simple answers on a given context), they are often enough.

Two mechanisms cut the bill further: prompt caching (up to 90% reduction on context repeated from one call to the next, typically the system instructions) and batch processing (50% reduction when the answer can wait a few hours). For an agent that processes documents overnight, it is half price.

The concrete calculation: an SME RAG

Let's take an internal assistant that answers your teams' questions about your documentation: 1,000 questions per month, each question sending about 2,000 tokens of context (the document excerpts found by the search, plus the instructions) and receiving a 300-token answer. Monthly total: 2 million input tokens, 300,000 output.

• Small model (Mistral Small, $0.10 / $0.30): 2 × 0.10 + 0.3 × 0.30 = $0.29 per month. Yes, under a euro.
• Intermediate (Claude Sonnet, $3 / $15): 2 × 3 + 0.3 × 15 = $10.50 per month.
• Frontier (GPT-5.5, $5 / $30): 2 × 5 + 0.3 × 30 = $19 per month.

Same calculation for an automation agent, which consumes more per task because it chains several steps: say 500 tasks per month at 10,000 input tokens and 1,500 output each, that is 5 million input and 750,000 output. With an intermediate model: about $26 per month. With a frontier: about $48. With a small model for the simple steps: under $1.

The conclusion fits in one sentence: at SME volume, API cost is not the issue. It only becomes one at high volume (tens of thousands of requests per day) or with a wasteful architecture that sends ten times too much context on every call. I detailed these traps in adding AI to a SaaS without blowing the budget.

Sovereignty and data: the non-price criterion

At these price levels, the provider choice often plays out somewhere other than on the rate. Mistral is French, offers hosting of its API in the European Union, and publishes open source models: for an SME subject to GDPR or working with sensitive data, that is an argument that weighs more than the cents of difference. OpenAI and Anthropic, for their part, offer professional options with zero data retention (your requests are neither stored nor used for training), to be enabled explicitly and verified contractually.

A third path: self-hosted open source models (Mistral, Llama and others). There, no more per-token bill, but an infrastructure cost to set against it: a GPU server rents for several hundred euros per month. It is only profitable at high volume or under strong confidentiality constraints. I compared the options in detail in sovereign and self-hosted RAG.

Why the model choice matters less than people think

This is the point I repeat most often in audits. In a well-built RAG, the quality of the search weighs more than the model: if the right passages of your documents surface, an intermediate model formulates a correct answer; if they do not, the best frontier in the world will answer beside the point, with confidence. A well-fed Mistral Large beats a poorly fed GPT-5.5, for a tenth of the output price.

And above all: a properly architected system isolates the provider. The model is a configuration parameter, not a foundation. If Mistral lowers its prices, if OpenAI changes its own, if a new model comes out, you change one line and rerun the tests. That is exactly how I build the RAG chatbot I deliver: the model choice happens at the end, on measured criteria (answer quality on YOUR questions, cost, sovereignty), and it stays reversible.

How much it costs in total

The budget of an SME AI project breaks down into two very unequal items:

• The build (one-off): this is the real investment. With me, a custom RAG chatbot on your documents starts at 2 640 € excl. tax, an automation agent between 440 € and 1 760 € excl. tax depending on complexity, from building blocks already proven in production.
• Operation (monthly): the API bill calculated above, a few euros to a few tens of euros per month, plus hosting, also controllable (a well-chosen edge stack stays in the tens of euros).

In other words: if a provider justifies a high quote with "API costs", ask questions. And if you hesitate to launch a project for fear of an uncontrollable OpenAI bill, that is, at SME volume, not the right place to be cautious.

Where to start

1. Quantify your real volume. How many questions or tasks per month? Multiply by 2,500 tokens and apply the rates above: you will have your order of magnitude in five minutes.
2. Choose your constraints before the model. Sensitive data? EU hosting? Zero retention? These criteria eliminate or impose providers far more reliably than price.
3. Test on your case, not on benchmarks. A sample of your real questions, two or three models, and we measure. That is what I do at the start of a project.

If you want an opinion on your specific case, I offer a free 30-minute audit: we quantify your volume, identify the constraints, and you leave with an honest order of magnitude, whether you work with me or not.

Sources

• OpenAI · official API pricing grid
• Mistral AI · Platform pricing
• Anthropic · official Claude API pricing