European AI: cheap for chat, costly for overnight agents

For chat, the difference is small change. For agents running all night, the calculation is different.

Prices as of September 2026.

I have just built a small app that helps you choose European AI, task by task. Once it was finished, I sat down to calculate the question I hear most often right now: what does going European cost us?

The short answer is that it depends on what you use AI for. For chat and documents, the difference is small change. For agents that need to run all night in your systems, it is a different story. Here is the calculation, layer by layer.

1. The model is the cheap part

An AI bill has three layers: the model, the harness around it, and the compute it all consumes when running. Most people only look at the first.

The three layers of an AI bill

Most people only look at the first.

Do you know which of the three layers drives your bill?

Model 1. An AI bill has three layers: model, harness and compute.

You pay for the model per token. The American frontier models, Claude Fable 5.1 and GPT-6 Astra, both cost around 10 dollars per million input tokens and 50 for output. A European open model such as Mistral Medium 3.5 costs 1.50 and 7.50, roughly one seventh of the token price. Fable and Astra also score higher on Artificial Analysis' Intelligence Index, but for everyday chat, that is not where the difference matters most.

Table 1. Token prices as of September 2026, USD excluding VAT. Sources: Mistral AI, OpenAI, Artificial Analysis.

For everyday chat, it is not the price per million tokens that determines the bill. It is how many millions of tokens an employee actually uses in a month, and that is less than most people think.

2. For chat and documents, European AI is cheaper, including on the invoice

For an employee using AI to write, summarise and look things up during the day, consumption is modest. For that kind of work, a European open model such as Mistral Medium 3.5 is an easy decision. You pay a fraction per token, your data stays in Europe, and you do not give up anything that matters in practice.

3. For overnight agents, token prices are not what drives the bill

The jump comes from somewhere else. An agent that needs to work while you sleep does not send one prompt. It runs for hours. It reads, plans, verifies its own work, tries again and tracks a task across many context windows.

It is everything around the model that determines whether it can do that: memory across windows, subagents, self-verification and guardrails that last all night. And that is exactly where the American frontier models are ahead right now. They can handle that kind of long-running work out of the box.

If you want the same on top of a European model today, you have to build more of the scaffolding yourself, and you pay for it in development time and in how reliable the run is. The token price is lower. The total price of an overnight run, measured end to end, is not necessarily lower.

> Principle: The token price is lower. The total price of an overnight run, measured end to end, is not necessarily lower.

4. A worked example

Suppose an overnight agent run uses 5 million input tokens and 2 million output tokens. On Claude Fable 5.1 or GPT-6 Astra, it costs around 150 dollars. On Mistral Medium 3.5, around 22 dollars. In isolation, the European model is clearly cheapest, even if the run has to be repeated a few extra times.

One overnight agent run

Token prices alone, before development time and failed runs.

But an overnight run costs more than tokens. It costs the developer who builds and maintains the scaffolding, and it costs when the run has not completed cleanly.

Model 2. Worked example: 5 million input tokens and 2 million output tokens. Prices as of September 2026.

But an overnight run that must be complete and correct when you wake up costs more than tokens. It costs the developer who builds and maintains the scaffolding, and it costs when the run has not completed cleanly. That is where the more mature frontier harness earns back its higher token price. Not on the model provider's invoice, but in your own time and in whether the work is done when you arrive.

5. What this means for your decision

The decision is not European or American across the entire organisation. It is task by task.

Chat, documents and bounded tasks: go European, take the savings and sovereignty with you. Long-running, autonomous agents in your core systems: today, the frontier harness is more mature, and if your compliance requirements demand a European model, budget for the scaffolding you need to add yourself.

Three questions are worth asking at your next management meeting.

1. Which of your AI workflows are chat, and which are overnight agents running unattended?

2. What does one overnight agent run cost you today, measured end to end, not just in token prices?

3. And do you know which of the three layers, model, harness or compute, drives your bill?

Most people do not know the answer to the last question. That is the best reason to look into it before the number grows on its own.

Sources

Prices verified in September 2026 against the providers' and Artificial Analysis' own pages.

• Mistral AI, API pricing: Mistral Medium 3.5, 1.50 / 7.50 dollars per million input/output tokens.
• OpenAI, GPT-6 Astra: 10 / 50 dollars per million input/output tokens.
• Artificial Analysis, Intelligence Index and model pages: Claude Fable 5.1, 10 / 50 dollars for input/output.
• My EU AI picker app and model overview, September 2026.

This is general practical guidance, not legal or compliance advice, and I am not a lawyer or compliance specialist. Do not use it as the sole basis for a decision. Always involve your own compliance or legal team and your DPO, and confirm dates and status against primary sources (EUR-Lex, the European Commission, the Danish Data Protection Agency). All prices are in USD excluding VAT and may have changed; always check the providers' own pricing pages. Status as of September 2026.