European LLM APIs
I have been researching European clouds that offer LLM inference APIs. Buying European reduces risk of LLM-mediated IP theft, minimizes risk of US and Chinese spying and meddling, and provides a practical intermediate solution before transitioning to local LLMs. This article is just exploration of offerings. I did not actually test any of these yet.
Mistral
The obvious first stop is Mistral, a French company that trains its own models. They are weaker than American frontier models, but they are also cheaper. Annoyingly, the prices are the same when denominated in euros, so you will pay 15% more in euros with current exchange rates. Mistral offers some interesting models listed below. Prices are per 1M input/output tokens.
- Mistral Medium 3 ($0.40/$2): Best non-reasoning multi-modal model from Mistral, offered at an attractive price. Its performance in benchmarks is quite weak compared to frontier models, but people voting on LMArena seem to like it anyway.
- Magistral Medium ($2/$5): Mistral's best multi-modal reasoning model. It performs better in benchmarks, even in coding, but it's glitching a lot, which is probably why people hate it on LMArena. I see it more as an experiment than a product at this point. It's a fine-tune of Mistral Medium, which makes the price surprising, but then reasoning models really cost more to run.
- Codestral ($0.30/$0.90): Affordable hosted non-reasoning FIM model for auto-completions in IDEs.
- Mistral Large 2 ($2/$6): Large but outdated model. Skip.
- Devstral Medium ($0.40/$2): Agentic but non-reasoning. Too weak for coding. Skip.
Considering that even the best model from Mistral is not competitive with the best opensource models, I will have to look for European LLM inference APIs that can run opensource models. Mistral itself does not offer this service.
Scaleway
Scaleway is a French cloud company offering LLM APIs for a wide range of opensource models at attractive prices. The best model they offer at the moment is qwen3-235b-a22b for €0.75/€2.25. Notice the low 1:3 input-to-output cost ratio. Many models have 1:1 ratio. This points to some inefficiency in their inference engine. Ditto for the surprisingly low output token limits.
OVHcloud
OVHcloud is another French cloud company. It has AI Endpoints service, i.e. LLM APIs. There's a price list, but I find the offering quite limited. The biggest model is only Llama 3.3 70B and corresponding DeepSeek finetune. Input and output token cost is the same, €0.67 for Llama 3.3 70B. They will hopefully expand the service in the future.
Nscale and Nebius
These two aren't quite European. Nscale has offices in London even though their datacenters are in continental Europe. Nebius is based in Amsterdam, but their leadership includes several Russians, including the CEO. Both companies have some interesting offerings that Scaleway and OVHcloud do not have though. Nscale offers GPT OSS 120B for $0.10/$0.40. Nebius offers DeepSeek R1 0528 for $0.80/$2.40.
What's next?
I think it's high time to move lighter tasks to European clouds. I am going to do so as well. Mistral Medium and Scaleway's Qwen 3 235B A22B look most promising. Capabilities are increasing quickly and applicability of European LLM APIs will grow with them. European clouds however cannot handle hard tasks like programming until they can offer frontier models.