Model providersScaleway (Iliad Group, France)

Scaleway

Decision matrix

Scaleway provider decision matrix
Overview

Scaleway Generative APIs is an OpenAI-compatible, pay-per-token API for open-weight and European models such as Mistral, Qwen, gpt-oss, GLM, DeepSeek and Gemma, hosted in a Paris data centre. The first million tokens are free, and Scaleway says it doesn't read or reuse your prompts. Pick it for EU-hosted chat, code and transcription. There's no hard spending cap.

Other good fits
  • Trying a model at no cost: the first 1,000,000 tokens and 60 minutes of transcription are free.
  • Overnight bulk jobs: requests sent through the Batches API cost 50% less.
Pros
Read pros
  • Pick it when data should be processed in France by a European company. Serverless models run in a Paris data centre operated by OPCORE. Scaleway says it doesn't collect, read or reuse prompts or outputs, except when traffic causes errors or looks malicious, and doesn't use them for training.
  • Pick it when you want more model choice than most EU clouds. The list includes Mistral Medium 3.5, GLM-5.2, DeepSeek V4 Flash, Qwen3.8, Gemma 4, gpt-oss-120b, Llama 3.3 70B, Whisper and embedding models.
  • Pick it when your code already speaks the OpenAI API. Chat completions, responses, embeddings, audio and batch endpoints are supported, with some listed exceptions.
  • Pick it when you later need fixed costs. Dedicated Deployment (formerly Managed Inference) runs a model on GPUs reserved for you, billed by the minute.
Cons
  • Pick something else when you need a closed frontier model. Use OpenRouter or the Claude API.
  • Pick something else when the price of gpt-oss matters most. OVHcloud lists gpt-oss-120b at €0.08 in and €0.40 out.
  • Pick something else when data must stay on your own hardware. Serve a model yourself with vLLM or Ollama.

Avoid if

  • You need a hard monthly spending cap. You can set billing alerts, but nothing blocks usage at a threshold.
  • Your contract requires a data centre outside Paris. All serverless models run in Paris today.
Connection

Create a sufficiently permitted Scaleway API key, keep its secret private, and call https://api.scaleway.ai/v1/chat/completions with a supported model ID. The documented service region is Paris (fr-par).

Billing
Read billing

Prices checked 23 September 2026 (Paris region, prices before tax):

ModelInput per million tokensOutput per million tokens
mistral-medium-3.5-128b€1.50€7.50
qwen3.8-27b€0.60€3.30
gpt-oss-120b€0.15€0.60
whisper-large-v3€0.003 per audio minute

The first 1,000,000 tokens and 60 minutes of audio are free. Batch requests get 50% off. The minimum billing unit is 1,000 tokens.

Running effort: no servers on the serverless side. Someone manages API keys, sets billing alerts and watches the model lifecycle page, because older models are eventually retired. The bill grows with tokens. Scaleway keeps aggregated, anonymised usage data for up to six months.

Data handling

This review did not establish retention or training controls for a particular Scaleway configuration. Confirm the current product documentation and contract before sending non-sample data.

First request

Documentation-based suggestion — not run here.

Read first request

Translate a product catalogue. A 15-person lab-supply shop needs 40 English product descriptions in German and French.

  1. Create a Scaleway account and project, then create an API key in the console.
  2. Open the Generative APIs playground, choose mistral-medium-3.5-128b and paste one description with: "Translate into German and French. Keep product names, numbers and units exactly as written."
  3. When the result reads well, run all 40 through the API using the code from the quick start (opens in a new tab).
  4. Have a native speaker check five translations, especially units and safety wording. The whole job fits inside the free million tokens.
Links & evidence
Documentation-based · not hands-on · Sep 23, 2026

Based on Scaleway's model-as-a-service price page and its Generative APIs quick start, FAQ and data-privacy docs, checked 23 September 2026. We did not run it for this review.

  • Model-as-a-service pricing · Sep 23, 2026 — Per-model prices in Paris region before tax; free tier 1M tokens and 60 audio minutes; Batches API -50%.
  • Generative APIs FAQ · Sep 23, 2026 — Serverless models hosted in Paris (OPCORE); Managed Inference renamed Dedicated Deployment, billed per minute; no hard billing threshold, billing alerts only; free tier details.
  • Data privacy · Sep 23, 2026 — Prompts not collected, read or reused except for abuse or errors; not used for training; anonymised usage data kept up to 6 months; data in Paris, France.
  • Quick start · Sep 23, 2026 — API key creation and OpenAI-compatible first request.
Further review notes & alternatives
Catalogue facts

Call open models through an OpenAI-compatible API in French data centers. Rent GPU Instances when you want to run vLLM, Ollama, or another engine. Choose hosted inference or self-managed hardware based on required control.

Scaleway (Iliad Group, France) · paid · cloud

  • pay-per-token api
  • gpu rental
  • eu sovereign
  • openai compatible
  • low carbon
  • gdpr compliance
  • model inference
  • automate
  • open-source
  • build-app

Browse all tools