EU LLM gateway: how to choose one in 2026

What an EU LLM gateway is

An EU LLM gateway — also called a European AI gateway — is a single API that sits in front of several large language model providers based in Europe. Your application sends every request to the gateway with one key; the gateway picks a provider that matches the jurisdiction you asked for, forwards the request, returns the answer and bills you once.

The word that matters is "EU". Plenty of gateways can reach a European region. An EU LLM gateway is one where staying in Europe is the default behaviour, not a setting you have to remember on every request: the company is European, the providers it routes to are European, and when one of them fails, the fallback is European too.

It helps to separate it from three neighbours it is often confused with:

• An inference provider runs the models on its own GPUs — OVHcloud, Scaleway, Mistral AI, Infomaniak, AKI.IO and others. It sells one catalogue, its own. A gateway runs no GPUs: it puts several such catalogues behind one interface. Our list of European inference providers covers the operators themselves.
• A general router with an EU option reaches providers worldwide and lets you restrict a request or a project to Europe. That can be enough, but the default route may leave Europe, so sovereignty depends on configuration being right everywhere, forever.
• A self-hosted gateway — LiteLLM, Portkey's open-source gateway and similar — is software you run yourself. It gives you full control of routing and logs, but you sign every provider's contract and operate the service.

Why use a gateway at all

If one provider serves every model you need, buy from it directly: it is cheaper and one contract is simpler than two. A gateway earns its place when:

• No single provider has your models. The open-weight models teams ask for — Qwen, GLM, DeepSeek, Llama, Mistral, Gemma — are spread across European providers. A gateway lets you call all of them with one account and one balance.
• You need failover without leaving Europe. Providers have outages and rate limits. A gateway retries on another eligible provider; an EU gateway guarantees that "eligible" means European.
• You need proof. Auditors and customers ask where data was processed. A good gateway returns that information with every response, so it lands in your logs instead of a slide.
• You want to keep your code portable. With an OpenAI-compatible gateway, moving between models or providers is a model ID, not a rewrite.

How to evaluate an EU LLM gateway: ten questions

Ask these of any gateway, ours included. Each one is answerable from the vendor's own legal notice, documentation and sub-processor list.

1. Who is the company, and who controls it? Incorporation, headquarters and parent company. US law follows control: the CLOUD Act lets US authorities require a company under US jurisdiction to disclose data it controls, wherever the servers are. A European company sits outside its direct reach — not outside every foreign law. Our GDPR vs. CLOUD Act guide goes through the detail.
2. Can the default route leave Europe? Read the routing documentation, not the landing page. "EU region available" and "EU only" are different answers.
3. What happens when the EU provider fails? Does the request move to another European provider, fail with an error, or quietly fall back to a US one? Only the first two keep your guarantee.
4. Is there per-call evidence? Response headers or fields naming the provider that served the request, and whether a fallback happened.
5. What does the gateway itself keep? Prompts and completions in logs or a database, for debugging, caching or analytics — and for how long.
6. What do the upstream providers keep? A gateway with zero retention in front of a provider that keeps prompts for 30 days protects nothing. Check each provider's retention sentence word for word.
7. Who are the sub-processors? The hosting of the gateway itself counts: many European gateways run on US clouds. That can be acceptable for transit, but it should be disclosed.
8. Is there a data processing agreement? A DPA under GDPR article 28, available before you sign up, naming the sub-processors.
9. Which APIs does it speak? OpenAI-compatible chat completions is the baseline; embeddings and the Anthropic Messages format matter if you run RAG or tools like Claude Code.
10. How is it priced, and who can buy? Markup on each token, a fee on each top-up, or a subscription; VAT included or not; open to individuals or businesses only.

European gateways and routers compared

Five services come up when you search for an EU or European LLM gateway. The facts below come from each company's own pages, on the dates given; the links are in each line.

cortecs

cortecs routes to about sixteen providers, European ones such as AKI.IO, Berget, IONOS, OVHcloud, Scaleway and TensorX, and the EU regions of the three US hyperscalers, at the upstream rate plus a 5 % fee on top-ups (checked 2026-10-13). Its sovereign-cloud page says that with eu_native: true "only EU providers are used", and that with it off, data "might be processed by entities of non-EU companies (Microsoft, Google, AWS, ...)" (sovereign cloud, read 2026-10-19). A good fit if you want the widest choice and are happy to set that flag on every project.

Melious

Melious routes across the twelve EU inference providers on its sub-processor page and says every request runs inside the EU, with no US or APAC fallback (sub-processors). It is cheaper than us on comparable models — on Llama 3.3 70B its default route costs less than ours, checked 2026-10-12 — sells to businesses only, and its router picks the provider for you. The full, sourced comparison is in Frontière AI vs Melious.

LowRouter

LowRouter passes the upstream provider's price through with no per-token markup and earns a fee on each top-up instead; on Qwen3 235B served by Scaleway in Paris, the same model on the same provider as ours, it is cheaper even with the fee (checked 2026-10-12). It routes to EU and US providers alike, closed US models included, and its own privacy policy says its US-established providers remain subject to US law, including the CLOUD Act. A region pinned in the model ID fails closed; without one, its router chooses. See Frontière AI vs LowRouter.

Requesty

Requesty is a UK company whose EU endpoint runs on AWS in Frankfurt; its privacy policy says that where inference happens "follows the model you select" (privacy, checked 2026-10-13). It is a global router with a European entry point rather than an EU-only gateway.

Global routers such as OpenRouter are compared in the best OpenRouter alternatives, which also looks at which keep traffic in Europe by default.

The self-hosted option

You can also run the gateway yourself. LiteLLM is the most common choice: a proxy that speaks the OpenAI format and routes to dozens of backends. Portkey's AI Gateway is an MIT-licensed alternative you can run with Docker or Node.js. Self-hosting gives you full control over routing rules and logs, and nothing leaves your infrastructure until it reaches a provider. In exchange, you sign and manage a contract with every provider, keep their keys, run the service and patch it.

The two approaches combine: point a self-hosted LiteLLM at Frontière AI as one upstream and keep your own policy layer on top. Our LiteLLM guide shows the configuration, tested end to end, and our comparison of open-source AI gateways covers the alternatives.

How Frontière AI works as an EU LLM gateway

Frontière AI is published by FLEECE AI SASU, a French company. It runs no GPUs; it routes to European inference providers and labels every model by jurisdiction before you call it. Today, 17 EU-sovereign models are live, served by Scaleway (France), Mistral AI (France), Infomaniak (Switzerland) and AKI.IO (Germany). 7 more are offered under a separate fast-access label: they run at US companies and are never presented as sovereign.

Routing and failover

• Sovereign by default, per model. A model labeled sovereign stays on sovereign providers, failover included. If its EU providers are all unavailable, the call fails with a 503 rather than leaving Europe.
• Per-key policy. On the Keys page, a key can be set to "Sovereign only" — fast-access models are then refused — or to "All providers", which explicitly allows a sovereign model to fall back to a US provider if its EU provider fails. The default is neither: sovereign models stay sovereign, fast-access models keep their labeled routes.
• Per-request header. X-Data-Sensitivity: sensitive tightens any request to sovereign providers only. A header can tighten a key's setting, never loosen "Sovereign only".

Evidence on every response

Each response carries X-Routed-Provider (the provider that served it), X-Sovereign (true or false) and, when the first choice failed, X-Fallback-Provider. Store them with your logs and you have a per-call record of where processing happened. In Python:

from openai import OpenAI

client = OpenAI(
    base_url="https://getfrontiereai.eu/api/v1",
    api_key="sk-front-...",  # your Frontière AI key
)

raw = client.chat.completions.with_raw_response.create(
    model="qwen3.5-397b",
    messages=[{"role": "user", "content": "Bonjour"}],
)
print(raw.headers.get("x-routed-provider"), raw.headers.get("x-sovereign"))
print(raw.parse().choices[0].message.content)

Here qwen3.5-397b is served by Scaleway. Every model ID, with its provider and label, is in the catalogue.

APIs

• /v1/chat/completions: the OpenAI contract, with streaming and tool calling — change base_url and the key, nothing else (documentation).
• /v1/embeddings: for RAG pipelines, see embeddings.
• /v1/messages: the Anthropic Messages format, so Claude Code and the Anthropic SDK work with ANTHROPIC_BASE_URL pointed at Frontière AI. It goes through the same routing, billing and headers. Not supported yet: /v1/responses and legacy /v1/completions.

Data

Prompts and completions are not written to our database. The gateway itself runs on Vercel's Paris region: Vercel is a US company, the traffic is in transit, and it keeps request logs but not prompts or responses — our privacy policy lists every sub-processor. Upstream, retention follows each provider: OVHcloud keeps billing data only; Scaleway keeps no prompts except requests that error or are flagged as abusive, for two weeks at most. Our sovereignty methodology shows how each provider is scored.

Price and access

Prepaid from €10, VAT included, no subscription, no top-up fee; open to individuals as well as companies. Our margin is built into the per-token price, so on the same model and provider, Melious and LowRouter are cheaper — their own comparison pages with us say so. If the lowest price on EU-hosted models is your only criterion, use them. If you want one key across several European providers, a sovereign default per model and the provider named on every call, that is what Frontière AI is for.

Related reading

For the bigger picture, read what frontier AI is and how open weights make a European route possible, what EU-hosted AI models really means, and what sovereign AI means in practice.

FAQ