Scaleway Generative APIs
Overall Rank #45 ⭐ Consider
⚠️ Proxy required for mainland China. The service runs in a single region (Paris, France) inside the EU, so latency and stability from the Chinese mainland need your own measurement | 🌍 International
💰 Token Pricing
| Type | Price | Note |
|---|---|---|
| Input | 按每百万 token 计费，EUR 目录价（未含税）：qwen3-embedding-8b 与 bge-multilingual-gemma2 0.10、mistral-small-3.2 0.15、gpt-oss-120b 0.15、gemma-4-26b-a4b-it 与 qwen3.6-35b-a3b 0.25、deepseek-v4-flash-0731 0.40、deepseek-v4.1-flash / qwen3.5-397b-a17b / qwen3.8-27b 0.60、qwen3-235b-a22b-instruct-2507 0.75、llama-3.3-70b-instruct 0.90、mistral-medium-3.5-128b 1.50、glm-5.3 与 glm-5.2 1.80、lightonocr-3-0.8b 2.20 | per million tokens |
| Output | 按每百万 token 计费，EUR 目录价（未含税）：gemma-4-26b-a4b-it 0.50、gpt-oss-120b 0.60、deepseek-v4-flash-0731 0.80、llama-3.3-70b-instruct 0.90、qwen3.6-35b-a3b 1.50、qwen3-235b-a22b-instruct-2507 2.25、qwen3.8-27b 3.30、qwen3.5-397b-a17b 3.60、glm-5.3 与 glm-5.2 5.50、mistral-medium-3.5-128b 7.50；嵌入与 OCR 模型不计输出费 | per million tokens |
💡 Free Credits: The free tier covers the first 1,000,000 tokens and includes 60 minutes of audio transcription; billing starts at token number 1,000,001
🤖 Supported Models (22)
qwen3-embedding-8b (Embeddings, 32k ctx, EUR 0.10/M input, output free)bge-multilingual-gemma2 (Embeddings, 8k ctx, EUR 0.10/M input, output free)mistral-small-3.2-24b-instruct-2506 (Chat and Vision, 128k ctx, EUR 0.15/0.35)gpt-oss-120b (Chat, 128k ctx, EUR 0.15/0.60)gemma-4-26b-a4b-it (Chat and Vision, 256k ctx, EUR 0.25/0.50)qwen3.6-35b-a3b (Chat, code, and Vision, 256k ctx, EUR 0.25/1.50)deepseek-v4-flash-0731 (Chat and code, 256k ctx, EUR 0.40/0.80, cached input EUR 0.08)deepseek-v4.1-flash (Chat, code, and Vision, 256k ctx, EUR 0.60/1.80, cached input EUR 0.12)qwen3.5-397b-a17b (Chat, code, and Vision, 250k ctx, EUR 0.60/3.60)qwen3.8-27b (Chat, code, and Vision, 256k ctx, EUR 0.60/3.30, cached input EUR 0.12)qwen3-235b-a22b-instruct-2507 (Chat, 250k ctx, EUR 0.75/2.25)llama-3.3-70b-instruct (Chat, 100k serverless / 128k dedicated, EUR 0.90/0.90)mistral-medium-3.5-128b (Chat and Vision, 180k ctx, EUR 1.50/7.50)glm-5.3 (Chat and code, 256k ctx, EUR 1.80/5.50)glm-5.2 (Chat and code, 256k ctx, EUR 1.80/5.50; deprecation 2026-10-16, EOL 2026-11-16)lightonocr-3-0.8b (OCR, PDF only, EUR 2.20/M input, output free)whisper-large-v3 (Audio transcription, EUR 0.003 per audio minute, output free)devstral-2-123b-instruct-2512 (Chat and code, 200k serverless / 260k dedicated)qwen3-coder-30b-a3b-instruct (Code, 128k ctx)voxtral-small-24b-2507 (Chat and audio understanding, 32k ctx)pixtral-12b-2409 (Chat and Vision, 128k ctx, 12 images per request, 4k max output)holo2-30b-a3b (Chat and Vision, GUI analysis, 22k ctx, CC-BY-NC-4.0)
✨ Pros
- ✓One of the few European sovereign clouds publishing a per-1M-token EUR price list: 17 serverless models across 18 price points, before tax
- ✓OpenAI-compatible endpoint at api.scaleway.ai - migration means changing base_url and the model name; the official OpenAI Python SDK and LangChain work unchanged
- ✓The Batches API is a flat 50% discount with no rate limit, which puts bulk re-embedding at EUR 0.05/M input and is the cheapest route for offline evaluation runs
- ✓Residency is documented, not implied: every serverless model is hosted in a single secure data center in Paris, France operated by OPCORE, with the vendor stating serverless models stay hosted in Europe
- ✓Generative APIs publish a 99.9% Monthly Availability Rate, with credits at 99.0-99.9% / 95.0-99.0% / below 95.0% of the monthly invoice
- ✓DeepSeek V4 Flash 0731 and V4.1 Flash price cached input at exactly one fifth of the standard input rate, the cheapest repeated-prompt path in the catalog
⚠️ Cons
- ×No mainland China endpoint and a single Paris region, so traffic from the Chinese mainland needs a proxy and your own latency measurement
- ×Cached-input pricing is published for only three models; for everything else you cannot derive the cache-hit cost from the official price list
- ×A strict single-region residency requirement forces you onto Dedicated Deployment, where the bill jumps from per-million-tokens to per-GPU-hour (H100-SXM-8 is roughly EUR 21,944/month)
- ×Serverless output caps are modest across the board (typically 16k-32k, and pixtral-12b only 4k), so long-output workloads need careful model selection
- ×The model roster has a published retirement schedule: glm-5.2 was deprecated on 2026-10-16 with EOL on 2026-11-16 and traffic rerouted to glm-5.3, so check the deprecation list before you integrate
🎯 Best For
Teams that need EU data residency; anyone keeping OpenAI as a second supplier to avoid a single point of failure; bulk offline work (re-embedding, classification sweeps, evaluation runs) that can use the 50% Batches rate; and compliance cases where a fixed location is worth a per-GPU-hour bill