One API for Europe's AI models

One gateway, three jobs

Everything that normally sits between your client and a model provider — the routing, the record, and the limits — in one place instead of three.

• The gateway

  One API. Every model.

  One base URL reaches every lab on the leaderboard. The European ones lead the list and carry the country they answer from, so choosing one is a decision rather than a guess.

  See every model →
• The record

  Every call, accounted for.

  Which model answered, how many tokens it took, what it cost and how it came back — written down per call, and queryable per organization, per key, per day.

• The limits

  Keys that actually stop.

  A ceiling and an allowlist on every key, enforced at the gateway rather than reported afterwards. A key that runs out stops answering instead of running up a bill.

Change the base URL. That's the integration.

There is no SDK to adopt and no adapter to write per vendor. Pick a model by its routing ID and the gateway handles the rest — including the vendors whose APIs are not OpenAI-shaped.

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://ai.rock8router.com/v1",
  apiKey: process.env.ROCK8_API_KEY,
});

const response = await client.chat.completions.create({
  model: "mistral/mistral-large",
  messages: [{ role: "user", content: "Bonjour" }],
});

curl https://ai.rock8router.com/v1/chat/completions \
  -H "Authorization: Bearer $ROCK8_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "mistral/mistral-large",
    "messages": [{ "role": "user", "content": "Bonjour" }]
  }'

Leading European models

Against the whole field

What you get

The parts everyone rebuilds in front of a model provider, built once.

• One OpenAI-compatible endpoint

  Change the base URL and the API key. Chat completions, streaming, tool calls and embeddings all keep the shapes your client already speaks.

• European models, first-class

  Mistral, Aleph Alpha, Silo, LightOn and the rest sit beside the American and Chinese labs, tagged with where they are answered from.

• Budgets that actually stop

  A spend ceiling per key, enforced at the gateway rather than reported after the fact. A key that runs out stops answering instead of running up a bill.

• Model allowlists per key

  Give a key three models and it can reach three models. No prompt-level policy to write, and nothing to keep in sync with the vendor's catalogue.

• A usage record you can audit

  Every call, its model, its token counts, its cost and its response code — queryable per organization, per key, per day.

• Published benchmarks, not our own

  The figures on this site are Artificial Analysis's measurements, shown unchanged. We reorder the list to lead with Europe and print each model's real rank beside it.

Questions we get

Do I have to rewrite my code?

No. The gateway is OpenAI-compatible, so you change the base URL and the API key and keep the client you already use. Streaming, tool calls and embeddings all keep the shapes they have today.

Which models can I reach?

Whichever ones your deployment offers. Each is tagged with the country it is answered from, and the ones inside the EU are marked as such, so you can restrict a key to European models and know the restriction is real rather than a preference.

What happens when a key runs out of budget?

It stops answering. The ceiling is enforced at the gateway as calls are made, not reported afterwards, so a key with no budget left returns an error instead of running up a bill you find out about later.

Can I use my own providers?

Yes. A provider is configuration: its base URL, its credentials and the models it offers. Anything that speaks an OpenAI-compatible API — including a model you host yourself — can be added and routed to alongside the rest.

Where do the benchmark figures come from?

Artificial Analysis, an independent source that publishes them. We show their measurements unchanged and reorder the list to lead with European labs — every row still carries the rank they gave it, so you can see exactly where a model places overall.

Why does one model appear several times?

The source scores a reasoning model once per effort setting and publishes a row for each, so a model measured at three settings is three rows. The setting is named in the row, in brackets after the model's name.