A region name on a price list does not commit a provider to process your data there. The written processing commitment decides your shortlist, and token price only ranks what is left.
Which EU-hosted LLM API should we shortlist when our DPA review rules out US endpoints? That question usually arrives when a customer's security team reads the data flow and stops a rollout. A DPA (data processing agreement) is the contract that says how a provider handles your personal data. Below, we put European LLM API providers side by side on region, price, compatibility and retention. A shortlist then takes one read rather than nine browser tabs.
The providers serving open models from Europe
We are Lyceum, a sovereign European AI inference provider for open-source models. Our serverless models run in Paris and Finland. Serverless means you call a model per token and run no servers yourself.
European clouds, model vendors and inference-only providers all sell per-token access to open models. Some US providers compete on price from US data centres.
Open-weight models are published once on a hub, and many providers can then serve them. A matching model name still does not prove identical weights or runtime settings. Quantisation (storing weights at lower precision to save memory) and chat templates (the format a model expects its conversation in) can differ too. Each of these can change quality, speed and token counts, so test the endpoint you will actually call.
Why look now? Since January 2026, the best open-weight models have trailed the best closed models by about four months. Epoch AI measures that on its capability index, which combines many benchmarks into one score. Epoch AI published that result on 29 May 2026. For many workloads, a European team no longer needs a US API to get near-frontier quality.
Models, regions and per-token prices compared
We read each provider page below on 20 October 2026. Prices are per million tokens, as input / output, in the currency the provider publishes. For each provider we list the point we checked, with the value from its own page. Only identical model names support a direct price comparison.
| Provider | What its own page says (read 20 Oct 2026) | Lyceum (us) on the same point (read 20 Oct 2026) |
|---|---|---|
| Mistral AI | Default endpoint api.mistral.ai is Global, with no commitment to an inference location; the EU endpoint api.eu.mistral.ai bills 1.1x standard list prices | EU processing at list price |
| OVHcloud AI Endpoints | Catalogue lists 22 models, none from Kimi, GLM, DeepSeek or MiniMax; Qwen3-Embedding-8B $0.12 | Qwen3 Embedding 8B $0.02 |
| Scaleway Generative APIs | glm-5.3 EUR 1.80 / EUR 5.50; its Claude Code guide runs the third-party y-router proxy, a translator between Anthropic and OpenAI request formats | GLM-5.3 $1.40 / $4.40; Anthropic-compatible endpoint for Claude Code |
| Nebius Token Factory | Public endpoints show Region Global, and the processing location can change without prior notice | Serverless models in Paris and Finland |
| IONOS AI Model Hub | Qwen3.8-27B $0.45 / $2.70; no Kimi, GLM, DeepSeek or MiniMax model listed | Qwen3.8 27B $0.40 / $2.40 |
| STACKIT AI Model Serving | Shared text models overlap with ours only on Qwen3.8 27B; no Kimi, GLM, DeepSeek or MiniMax model listed | Kimi, GLM, DeepSeek, MiniMax and Qwen models |
| TensorX | z-ai/glm-5.3 $1.75 / $4.50 | GLM-5.3 $1.40 / $4.40 |
| DeepInfra | Runs in US-based data centres | Serverless models in Paris and Finland |
Our own models get the columns a procurement review asks for. Each model page states the model's processing region and price. All three models below are called through our OpenAI-compatible API or our Anthropic-compatible endpoint.
| Reference model (API string) | Processing region | Input / cached input / output per 1M (USD) |
|---|---|---|
| GLM-5.3 (z-ai/glm-5.3) | EU-hosted | $1.40 / $0.26 / $4.40 |
| DeepSeek V4 Flash (deepseek/deepseek-v4-flash-0731) | EU-hosted | $0.25 / $0.06 / $0.30 |
| Qwen3.8 27B (qwen/qwen3.8-27b) | EU-hosted | $0.40 / $0.10 / $2.40 |
A worked example shows what these list prices mean for a monthly bill. The volume is hypothetical: 100 million input tokens and 20 million output tokens a month on GLM-5.3, with no caching.
# Hypothetical volume: 100M input and 20M output tokens a month on GLM-5.3.
# List prices per 1M tokens, read 20 Oct 2026. EUR converted at the ECB rate of that day.
EUR_USD = 1.1215
prices = {"Lyceum": (1.40, 4.40), "TensorX": (1.75, 4.50), "Scaleway": (1.80 * EUR_USD, 5.50 * EUR_USD)}
for name, (inp, out) in prices.items():
print(name, round(100 * inp + 20 * out, 2))
# Lyceum 228.0 / TensorX 265.0 / Scaleway 325.23On this volume, we bill $228 a month and TensorX $265. Scaleway bills EUR 290 (100 x EUR 1.80 + 20 x EUR 5.50), about $325 at the ECB reference rate of 20 October 2026. Cached input (the repeated start of a prompt, billed at a lower rate) changes the result. If 60% of the input came from cache, our bill would be $159.60. That is 40M tokens at $1.40, 60M at $0.26 and $88 of output. The example uses list prices only, so negotiated discounts are not in it, and the cache share is an assumption. It does not show quality, latency or rate limits.
OpenAI compatibility and what breaks on a swap
An OpenAI-compatible API accepts the same request format as OpenAI's chat completions API. You change the base URL, the key and the model name. Our serverless API works this way, and Claude Code connects through our Anthropic-compatible endpoint, which accepts the request format of Anthropic's Claude API. A compatible request format does not promise identical behaviour. Three things usually break:
- Tool calling. Tool calling is the model asking your code to run a named function. Serving stacks (the software that runs a model behind an API) parse it per model. Our docs record one case: tool_choice "required" is currently not enforced on deepseek/deepseek-v4-pro-0813, deepseek/deepseek-v4-flash-0731 and z-ai/glm-5.3. On those three models, a reply can be plain text while finish_reason still says "tool_calls". It is a documented limit, so check for tool_calls in the message itself.
- Streaming shape. Streaming sends the answer in chunks, and the fields in those chunks differ between serving stacks. Reasoning models add another field. A reasoning trace is the text a model writes while it works towards its answer. On our API, most models return it in message.reasoning_content, while some backends use message.reasoning.
- Token accounting. Tokenizers (the rules that split text into tokens) use different algorithms from model to model, and usage fields differ between providers. On our API, prompt_tokens_details is absent on a cache miss (a request that found no reusable prompt start) rather than reported as zero.
Read the usage object defensively, so a missing field does not break your cost tracking:
details = resp.usage.prompt_tokens_details
cached = details.cached_tokens if details else 0Record the model, settings and client version of every test, so you can compare a later change. Our guide OpenAI-compatible APIs: what breaks when switching models covers each failure in more depth.
Retention and training policies compared
Article 28 of the GDPR sets the contract your processor has to meet. The processor acts only on documented instructions from the controller. It needs authorisation before it engages another processor, and deletes or returns the personal data at the end. Procurement checks each provider's retention terms against it.
What GDPR compliance means for an LLM API
GDPR compliance describes how you use a service, not a label an API carries. You are usually the controller, the party that decides why data is processed, and the API provider is your processor. Chapter V of the GDPR holds its rules on transfers abroad. Sending personal data to a country outside the EU needs a legal basis there. The EDPB, the board of EU data protection authorities, explains when sending data abroad counts as a transfer. GDPR does not ban every US service. An adequacy decision is the Commission's ruling that a country protects data well enough. One such decision covers only US companies certified under the EU-US Data Privacy Framework. An EU-hosted endpoint simplifies that review but does not replace it. For the legal side of residency, read AI data residency, explained properly.
Review retention at the product and endpoint level. Ask separately about prompts, outputs, application logs, abuse monitoring, billing records and backups. Check whether zero retention is automatic or needs approval, and whether batch jobs, files or agents are exceptions. A no-training statement does not by itself mean no storage.
Our position is short. We don't keep your prompts or outputs, and we never use them for training. Nothing is written to disk. Our DPA is Annex 1 of our terms, so it applies when you accept them. The TOMs (technical and organisational measures) are available on request. A sub-processor is another company that processes data for us. We tell you before we add or replace one, and you can object within 10 business days. This is our stated policy, not a certification.
Where the published data is thin or missing
Four topics are thin in what providers publish, so ask each vendor about them:
- Processing scope: get a commitment that covers the exact endpoint, model, failover and support access.
- Control plane: a regional endpoint can still handle API keys, billing and usage analytics outside the region. Ask where that data is kept.
- Retention: separate request content from operational metadata, and confirm the exceptions.
- Capacity: request account-specific rate limits and test realistic concurrency.
Sometimes a provider page does not state a value. A dated "not published" is then more useful to procurement than a guess, because it names the question to ask in writing.
We hold ourselves to the same standard. A service credit is a refund when a promised service level is missed. We publish no service-level agreement, uptime target or service credit for serverless inference, and business terms are agreed per contract. Our audit for ISO 27001:2022, the international standard for information security management, is in progress. We send our sub-processor list by email on request. Live status, per-model latency and past incidents are on our status page.
Which provider fits which workload
The provider that fits passes your written requirements for the model you will call. Price decides only among those that pass.
A three-question rule for choosing an EU-hosted LLM API
- Does the endpoint you will call commit in writing to process in the EU, for the model you need? If it does not, it is not EU-hosted for your review, whatever the price.
- Are prompts and outputs stored by default, and does the DPA say so? If storage is the default, ask what switches it off and whether that needs approval.
- Does the model you need run on that endpoint at a dated price? Only then compare cost on your own volume, as in the worked example above.
The first question removes candidates. The second decides what your legal team signs. The third ranks what is left.
Then map your workload to the check that matters most:
- Coding agents: test tool calls, reasoning fields and retries. Claude Code reaches open models such as GLM-5.3 through our Anthropic-compatible endpoint.
- Retrieval and search: embeddings are billed on input only, and Qwen3 Embedding 8B costs $0.02 per 1M tokens on our API.
- Steady, high volume: a dedicated endpoint reserves GPUs for your model alone. Compare it with your token bill, as in pay per token vs dedicated GPU inference.
This comparison is a dated snapshot of provider pages read on 20 October 2026, so recheck live rates before you buy. Compare the model you plan to call, at the price and region published for it today, in our EU-hosted LLM API model roster.
Sources
- huggingface.co
- huggingface.co
- epoch.ai
- ecb.europa.eu
- https://docs.lyceum.technology/docs/inference/serverless
- docs.vllm.ai
- huggingface.co
- gdpr-info.eu
- gdpr-info.eu
- gdpr-info.eu
- https://www.edpb.europa.eu/system/files/2023-02/edpb_guidelines_05-2021_interplay_between_the_application_of_art3-chapter_v_of_the_gdpr_v2_en_0.pdf
- https://commission.europa.eu/law/law-topic/data-protection/international-dimension-data-protection/eu-us-data-transfers_en
- https://docs.mistral.ai/api