Working out AI cost per case: from tokens to euros

Prices come per million, your bill runs per case

One division sits between them. Vendors quote amounts per million tokens, a company thinks in cases: one customer email, one invoice, one set of minutes. Tokens are the pieces a model chops text into, and Anthropic's documentation works with roughly four characters or 0.75 words per token in English; German, with its long compound nouns, runs higher.

Nobody has to estimate precisely. Every API reports the tokens actually consumed, input and output, after each call, so twenty real cases give you the average that carries the rest of this calculation.

The formula fits on one line

Two counts times two prices, divided by the exchange rate. Everything below converts at the ECB reference rate of 18 September 2026, 1 euro to 1.1460 US dollars.

• Cost per case in euros = ((input tokens / 1,000,000) × input price + (output tokens / 1,000,000) × output price) / 1.1460
• Example case: sort one customer email and draft a reply, measured at 1,800 input tokens and 400 output tokens.
• With Claude Haiku 4.5 at 1.00 and 5.00 dollars: (0.0018 × 1.00 + 0.0004 × 5.00) / 1.1460 = 0.0033 euros, or 0.33 cents per case.
• At 1,000 cases a month: 3.32 euros.

Five models, one workload, a factor of 45

Same thousand cases, five price tags. The figures below were computed with python3 from the list prices named above and the ECB rate, each at 1,800 input and 400 output tokens per case:

• GPT-5.6 Luna, 0.20 and 1.20 dollars: 0.73 euros a month.
• Gemini 3.5 Flash-Lite, 0.30 and 2.50 dollars: 1.34 euros.
• Claude Haiku 4.5, 1.00 and 5.00 dollars: 3.32 euros.
• Claude Sonnet 5, 2.00 and 10.00 dollars: 6.63 euros.
• GPT-6 Astra, 10.00 and 50.00 dollars: 33.16 euros.

An output token costs five times an input token

Answering is dearer than reading. On Claude Haiku 4.5, Claude Sonnet 5 and GPT-6 Astra an output token costs five times an input token, on GPT-5.6 Luna six times, on Gemini 3.5 Flash-Lite 8.3 times; that is why the 400 output tokens in our workload carry 52.6 percent of the bill on Claude Haiku 4.5 and 64.9 percent on Flash-Lite, while making up only 18 percent of the tokens.

That gives you a lever no vendor switch can match. Force the model to return three fields instead of a finished draft, say category, urgency and template number, and roughly 150 output tokens will do; on Claude Sonnet 5 the monthly bill then drops from 6.63 to 4.45 euros, a third less, with the model untouched.

Two switches cut what is left

Both sit in the documentation rather than in a negotiation. Batch processing comes with a 50 percent discount at all three vendors on input and output, as long as the result does not have to land within seconds; for an overnight run across yesterday's post, that is the right mode.

The second switch is prompt caching: a cache hit costs 0.1 times the input price according to Anthropic's price list. In our workload, 1,500 of the 1,800 input tokens never change, being instructions, the category list and examples. Caching brings the 3.32 euros down to 2.14, batch to 1.66, both together to 1.07. Haggling over the price per million tokens at that level saves you the price of an office chair per year.

Whole documents are where it gets expensive

Context is the cost driver. A 100 kB documentation page comes to roughly 25,000 tokens according to Anthropic's documentation, a 500 kB PDF to roughly 125,000. Put 200 scanned invoices a month fully into the context, 25,000 input tokens each, and expect 600 tokens of answer, and you pay 4.89 euros on Haiku 4.5, 9.77 on Sonnet 5 and 48.87 on GPT-6 Astra.

Two habits push that number further. An agent working through ten steps resends the growing transcript at every step, and screenshots are billed as image input. Our guide to preparing incoming invoices with AI shows how such a flow works without carrying every page along.

Review costs a hundred times the model

This is where your bill is decided. A thousand cases at 30 seconds of visual checking each come to 8.3 working hours; at 40 euros fully loaded per hour that is 333 euros a month, against 3.32 euros for the model. The factor is 100.

So the right question is not which vendor delivers the cheapest million tokens, but how many cases may pass without a human looking. A workflow that decides 70 percent of cases safely and presents the other 30 percent saves more than any model switch. And a model that writes into your systems needs the boundaries from today's piece on what an AI agent should be allowed to do anyway.

Who this pays off for, and when it does not

From around 300 similar cases a month we would run the numbers and then build. Below that, the model bill is pocket change either way, and setup, testing and briefing cost more working time than the workflow saves in its first year. Above it, keep the order: shorten the output first, then cache, then batch, and only then change model.

One situation argues clearly against your own API integration: if the cases differ and everyone in the building needs AI for something else, a paid per-seat plan is the quieter route. Free tiers are no way to save on customer data, as the piece on masking customer data before the prompt shows from the vendors' own terms.

Three questions, then count for a week

Are list prices realistic? For small volumes yes, because discounts start at high volume. Prices do move, though: the ones quoted here were read off the vendor pages on 21 September 2026, and Google already shows a doubling for Gemini 3.8 Flash on 1 January 2027.

How many tokens does our case use? Run twenty real cases and read the input and output token fields out of the API response. That is fifteen minutes of work and replaces every estimate.

Does an expensive model ever pay off? Yes, when it lowers the review rate. If the stronger model costs 30 euros more a month and saves two hours of checking, it has paid for itself.

Next step: this week, put three numbers into a spreadsheet for your most used AI workflow, cases per month, input tokens and output tokens per case, and run them through the formula above. If the result does not match what you thought you were paying, we go through it in a free first call; the wider frame sits under services for internal AI workflows.