EU LLM API prices compared

This page puts the posted prices of the open-model APIs side by side, for the models that more than one of them serves. It names the currency each vendor quotes in, and it does not convert: an exchange rate moves, and a converted price reads as a promise.

Checked on 2026-11-07. Every vendor changes prices, so treat this as a starting point and confirm at the source before you plan a budget. Every source is linked at the foot of this page.

DeepSeek-V4.1-Flash-1017

Per million tokens, input / cached input / output.

Four platforms, one model, and the three that publish a figure in dollars publish the same figure. The difference is not the rate. It is whether the vendor will tell you which region ran the request.

GLM-5.3-Flash

The same pattern: an identical rate, and a different answer on location.

Qwen3.8-27B

Blade's figure is a floor, not a price. Its rate card is explicit that an eligible model "follows its GPU class's hourly schedule" and can rise above the list rate; its published ceiling for this model is €0.21 / €2.76. Against that ceiling our fixed price is lower on both input and output.

DeepSeek-V4-Flash

gpt-oss-120b, which we do not serve

A comparison page that only lists races you win is an advertisement. OVHcloud is the cheapest EU route to gpt-oss-120b, and our public catalog does not carry that model at all. If it is the model you want, buy it there.

Reading a rate card without being misled

A list price is not a bill. Four things move the number you actually pay, and none of them is on the headline row.

• Cached input. A long system prompt, repeated on every request, is most of the input bill. The cached rate on DeepSeek-V4.1-Flash-1017 is $0.0060 per million tokens here, two orders of magnitude under the uncached rate. A vendor without a cached price is quoting you the expensive path.
• The output ratio. Output costs three to ten times input. A summarizer and a reasoning agent on the same model have different economics, so price your own mix rather than the midpoint.
• A floor is not a price. A dynamic rate card quotes the cheapest hour. Ask for the ceiling, and for the history of what the rate actually was.
• The currency. EUR before tax and USD after are different numbers. We quote USD; Scaleway, OVHcloud and Blade quote EUR.

A region costs nothing extra on an open model. That is the single most useful finding on this page. On the two models where four platforms compete, the EU price and the unnamed-region price are identical. If a vendor charges a premium or a multiplier for a named region, that is a commercial choice, not a cost.

Beyond the token

Three things in a bill are not tokens at all, and they are where budgets go wrong.

• OCR. We bill PDFs and images per page, at $0.0044 for the automatic model and $0.0022 for the fast path. Most rivals run OCR through a vision model and bill per token, which makes a document budget a guess. Mistral bills per 1,000 pages.
• A flat fee. Helmcode sells an API key, not tokens, from €399 a month. Above roughly a billion tokens a month on two models that is cheaper than any meter; below it, much more expensive. See the Helmcode comparison.
• GPUs by the hour. Together AI lists an HGX H100 at $5.49 an hour and an HGX B200 at $8.99. That is a different product from an API, and the right one if you train.

Where to go next

• Our own prices, per model and per tier
• The models we serve, with the price of each
• The comparisons in detail: Scaleway, OVHcloud, Mistral AI, Blade, Together AI, Fireworks AI, Helmcode, Nebius

Sources

Checked on 2026-11-07: OVHcloud AI Endpoints catalog, Blade rate card, Blade hourly pricing, Together AI pricing, Fireworks AI pricing, Helmcode pricing, Mistral AI pricing, Scaleway Generative APIs FAQ.

Scaleway's price page refuses an automated read, so its figures were recorded on 2026-10-22 from its pricing page.

Munito AI prices apply from 2026-11-04 and come from the model catalog.