Own fine-tuned model or API: where is the break-even?

First: may the data go to an API provider?

Many of our conversations do not start with the bill. They start with data that may not leave the company:

• health data, banking secrecy, trade secrets, case files;
• a data protection officer who will not approve another foreign sub-processor;
• work on a network with no outside connection.

Then the comparison is not "API or own model" but "own model or no model". Set this at the top of the calculator, under "Your data": the API path switches off, the cascade goes to zero and the hard cases go to a person. If the data may not even come to us, only your own hardware remains. GDPR does not require an on-premise model; for ordinary personal data an EU provider under a processing agreement is lawful. More in running a private LLM on your own server.

The second reason is behaviour. A smaller model trained on your examples can follow a strict output format or your terminology more reliably than a prompt does. Whether it does on your task is measured, not assumed.

How the calculator counts

One API request costs input tokens times the input price plus output tokens times the output price. Multiply by requests per month and you have the API path.

The own-model path has three parts:

• Fixed monthly cost: our EU hosting, or your hardware spread over the depreciation period plus electricity, plus the retainer if you choose one.
• The cascade: the share of requests that still goes to the large API. Those you keep paying for.
• Corrections: the share of answers a person must fix, times minutes per fix, times the hourly rate. On an API that share stays flat, because the model never sees your corrections. With your own model the pilot cuts it once, and each retraining cycle under a retainer cuts what is left, never below 30 % of today. Switch it off if you do not want to count corrections.
• The start: Pilot S, Pilot M or a program, paid once.

The break-even month is when the cumulative cost of both paths is equal. The calculator also shows two volumes: where the monthly costs meet, and how many requests you need to repay the start within a year. The second number is usually the honest one.

A worked example with public prices

Defaults: 80,000 requests a month, 1,500 input and 400 output tokens, Claude Opus 5.5 at USD 4 and 20 per million tokens. A month costs USD 1,120, which at the ECB rate of 1.1298 is about €991.

If the model runs on a workstation you already have, it costs about €30 of power a month. With the 10 % that still goes to Claude Opus 5.5 that is about €129 a month, a saving of €862. Pilot S (€3,200) pays back in month 4, and over 24 months you pay about €6,300 instead of €23,800.

On our EU hosting at €600 a month the own model costs about €699 a month and Pilot S pays back in month 11. With a Lite retainer at €690 a month on your own machine, Pilot S pays back on token savings alone in month 19, and with corrections (2 % of answers) in month 4: a retainer is justified by fewer corrections, new tasks and a better base, not by the API bill.

A cheaper API raises the line. On Claude Sonnet 5.5 (USD 2 and 10) the same traffic costs about €496 a month: on your own machine Pilot S pays back in month 8, on our hosting only after two years. The default calculation also counts corrections (2 % of answers, 3 minutes, €25 an hour); the amounts above leave them out.

At a bill of a few tens of euros, keep the API

That is the plain conclusion. 20,000 requests a month on Claude Sonnet 5.5 cost about €124. Even on a machine you already have the saving is about €80 a month and Pilot S pays back only after two years. An own model then makes sense mainly for the data or the answer format, not for the price.

A laptop is a real option for a 4B model. Our Sokol 4B runs on one, reads images and calls tools. A workstation with one graphics card, even an older one, now runs a quantised 27B model, such as our Sraka 27B. Whether it handles your traffic depends less on model size than on how many requests it must serve at once: for a handful of people such a machine is enough, for steady customer traffic it usually is not.

What the calculator leaves out

• API discounts. Prompt caching and batch processing lower the API bill; Anthropic lists 50 % off for batch. If you use them, enter lower prices.
• Token counts differ by model. The same text gives a different number of tokens with different models. Take the counts from your own API usage report.
• Your team's time. Someone has to look after a machine you own. With our hosting, that is included.
• Quality. The calculation assumes the small model is good enough. Sometimes it is not, and then the task stays with the large model.

What our side costs

• Evaluation, €1,900, one day. Up to five models on a frozen set built from your examples, a results table and a written recommendation. Credited against a pilot. Sample report.
• Pilot S, €3,200, 2–3 weeks. One fine-tuned model on data you already have, base-vs-tuned scoreboard, GGUF build.
• Pilot M, €9,000, 4–5 weeks. When we build the data.
• Programs from €15,000. Retainer with retraining from €690 a month; Watch without training €190 a month, and a retraining cycle on demand €1,490.

All prices in EUR, excluding VAT, on the price list. If you are not sure your data is ready, the 12-question data check tells you first.