Scaleway Generative APIs Review 2026: EU Per-Token Pricing and the Paris-Only Residency Rule

Scaleway is a French cloud provider that started as a bare-metal and GPU-hosting company and grew into a full European hyperscaler with GPU Instances, Object Storage, Containers, and a managed inference product called Generative APIs. For an API catalog, Scaleway is interesting for one reason above all others: it is the rare provider that publishes a clean, self-hosted, per-million-token price card in euros for models you can actually call, hosted inside the European Union, with no marketplace markup and no reseller chain in between.

That combination puts Scaleway into a narrow band. It is not the cheapest provider in the market — Groq, Cerebras, DeepInfra, and Fireworks all undercut several of these rates in USD. It is not the widest catalog either — OpenRouter carries 457 models against Scaleway's 17 serverless entries. What Scaleway is, and what almost nothing else is, is a first-party European cloud selling European-hosted inference at published EUR rates, with an SLA attached, a free tier, and a batch discount. If your procurement checklist contains a data-residency line item, that is the reason to read the rest of this review.

We verified every price below against the live Scaleway Model-as-a-Service pricing page and the Generative APIs documentation on October 29, 2026. The table has changed since our last crawl 42 days ago: the lineup now includes GLM-5.3 and GLM-5.2, DeepSeek V4.1 Flash, Qwen 3.8 27B, and Gemma 4 26B A4B, while Qwen 3 235B Thinking 2507 has dropped off the serverless list. Prices are quoted before tax, in euros, and are not normalized to any other currency.

Two deployment modes, two billing models

Scaleway splits its inference product into two distinct offerings, and confusing them is the most common mistake when reading the pricing page.

Serverless is the pay-per-token product. You call a pre-configured endpoint, you pay for tokens consumed, and you never touch a GPU. Billing starts at token 1,000,001 because the first million are free. This is what most developers should start with, and it is where the per-model prices below apply.

Dedicated Deployment is a managed GPU instance running a model you select. You pick a GPU instance type from the Scaleway hardware range, Scaleway provisions it, you deploy an optimized model onto it, and you are billed per GPU hour for the time the deployment exists. Critically, billing only starts once the model is actually deployed — not while you are waiting for it. The rates run from EUR 0.93/hour for an L4 up to EUR 30.06/hour for an 8-way H100 SXM cluster.

The practical read: Serverless is for spiky, unpredictable, or pre-launch traffic. Dedicated Deployment is for steady high-volume workloads where you want predictable throughput, no rate limits, and a hardware choice. A production system often ends up using both — serverless for the long tail of small models and dedicated for the one model that carries your main volume.

Serverless pricing: the full 17-model table

This is the complete serverless catalog as published, with input and output rates per million tokens. Cached-input rates are listed where Scaleway publishes them separately.

Three observations from the table that matter more than the individual prices.

Cached input is published for three models, and the DeepSeek entries are the cheapest way into it. DeepSeek V4 Flash 0731 and DeepSeek V4.1 Flash both carry an explicit cached-input rate at exactly one fifth of the standard input rate, and Qwen 3.8 27B publishes the same one-fifth rate at EUR 0.12/M. Every other model shows no cached rate on this page, which does not mean caching is unavailable — it means the price is not published as a separate line here. If your workload is dominated by a repeated system prompt, those three entries are the only ones where you can price the cache hit from this page alone.

Embeddings are priced per input token with no output charge. Qwen 3 Embedding 8B and BGE Multilingual Gemma2 both sit at EUR 0.10/M. Embedding models are the easiest thing to move because they are small, stateless, and easy to re-index, which makes them the low-risk place to test a new provider.

The two GLM models cost the same, and the GLM-5.3 entry is the one to use. GLM-5.3 and GLM-5.2 are both EUR 1.80/M in and EUR 5.50/M out. Same price for what is presumably a newer model. That is a genuine reason to pick 5.3 over 5.2 — but also a signal worth watching, because flat pricing across generations is unusual in this market and can change without notice.

Dedicated Deployment: per-GPU-hour pricing

When you deploy a model onto dedicated hardware, you are billed for the GPU time, not the tokens. Scaleway publishes the Paris-region Managed Inference rate card, and it is worth reading carefully because the monthly equivalents are large numbers.

The break-even math is straightforward and worth doing before you commit. An H100-1-80G at EUR 3.40/hour costs roughly EUR 2,481 per month if left running continuously. Against that fixed cost, serverless pricing only makes sense for a specific throughput. Run the dedicated instance against the same traffic that would cost more than EUR 3.40/hour in tokens and dedicated wins; below that, it does not. And note the billing rule: you pay for the deployment, not for tokens. An idle dedicated instance bills at full rate while serving nothing.

The residency rule: Paris today, Europe tomorrow

This is the part that most reseller summaries get wrong, so we want to be precise.

Scaleway's own documentation states that all serverless models are currently hosted in a secure data center in Paris, France, operated by OPCORE. The guarantee Scaleway gives is that serverless models reachable through api.scaleway.ai remain hosted in Europe, and that customers will be notified when the hosting location changes from France-only to a wider European footprint.

Two implications follow, and they point in opposite directions.

First, the guarantee you get is European, not French. Your data does not leave the EU. But if you are under a French-specific data-residency commitment — some French public-sector and health-adjacent procurement rules, or an internal policy naming a specific country — Scaleway's current footprint does not satisfy it, because there is only one location and it is France. The Amsterdam region that appeared in earlier third-party write-ups is a Scaleway cloud region for other products, not a second serverless inference site.

Second, Scaleway's own recommendation for strict single-region processing is Dedicated Deployment, because you can pin that deployment to a specific region and instance. So the provider that markets itself on sovereignty routes its strictest requirement to its more expensive product. That is a coherent design — dedicated hardware is how you get a deterministic location — but it means the "EU-only" headline and the "I need to control the exact region" requirement land on different tiers.

The practical version: if your requirement is "must not leave the EU," serverless at EUR 0.10/M is a good deal. If your requirement is "must stay in exactly this place," budget for dedicated deployment.

The Batches API: -50% and no rate limit

The Batches API is the most under-discussed value in this catalog, because two benefits stack.

First, every request submitted through Batches is priced with a -50% discount against standard model prices. Second, Batches requests do not have a rate limit, because the endpoint is explicitly designed for non-real-time workloads.

That combination makes batches the cheapest legal path through Scaleway for anything that does not need a response inside a request timeout. Effective rates on the table above become EUR 0.05/M input for the embedding models, EUR 0.075/M for Mistral Small 3.2, and EUR 0.90/M in / EUR 2.75/M out for GLM-5.3. For a bulk re-embedding, a classification sweep, or an offline evaluation harness, this is the configuration to use.

The tradeoff is latency, obviously, and the documentation is candid that batch turnaround depends on queue depth and on how you split your work. If you submit 10,000 small requests you wait for a queue cycle; if you submit fewer, larger requests you wait longer per item but often finish sooner overall. Scaleway's own guidance for interactive workloads is to split queries into the smallest meaningful parts.

Integration: OpenAI-compatible, with one honest caveat

Scaleway Generative APIs is designed to be compatible with OpenAI libraries and SDKs, including the OpenAI Python client library and LangChain. The endpoint is api.scaleway.ai, with standard /v1/chat/completions and /v1/embeddings routes and standard Bearer authentication. For most integrations this is a base URL swap.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.scaleway.ai/v1",
    api_key="SCWXXXXXXXXXXXXXXXXX",
)

resp = client.chat.completions.create(
    model="mistral-small-3.2-24b-instruct-2506",
    messages=[{"role": "user", "content": "Summarize this in one sentence."}],
    stream=True,
)

for chunk in resp:
    if chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end="")

And the same call over plain HTTP, for the batch path:

curl -s https://api.scaleway.ai/v1/chat/completions \
  -H "Authorization: Bearer $SCW_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "glm-5.3",
    "messages": [{"role": "user", "content": "Write a haiku about EU data residency."}]
  }'

The caveat is about model naming, not protocol. Scaleway exposes its own model IDs — glm-5.3, qwen3.5-397b-a17b, mistral-small-3.2-24b-instruct-2506 — which are not identical to the IDs used by OpenRouter, Hugging Face, or the model vendors. Any provider-switching layer you build needs a per-provider ID map, and if you intend to move between Scaleway and a multi-model aggregator, that mapping is the part that will actually cost you engineering time. The API shape is portable; the model names are not.

One more integration note worth having: Scaleway also appears on the Hugging Face Inference Providers router, where you can select it per request against the same OpenAI-compatible endpoint. If you are already routing through HF, Scaleway becomes a selectable EU-residency backend rather than a separate integration.

Rate limits, quotas, and budgets

Scaleway uses a three-layer control model on serverless, and understanding which layer bites you saves a support ticket.

• Organization quotas are system-level limits Scaleway manages. They apply across your entire organization, measured in tokens per minute, and exist to manage burst.
• Project rate limits are limits you set yourself at the project level. They cannot exceed the top-level organization quota, and they also measure tokens per minute.
• User budgets are per-end-user usage limits enforced on a daily or monthly timeframe, measured in token volume. These are cost controls, not throughput controls.

If you need more throughput, Scaleway lists three routes in order of preference: move non-real-time work to the Batches API (which also has no rate limit and a 50% discount), move to Dedicated Deployment for dedicated capacity without enforced rate limits, or contact your account manager for a volume commitment that raises the quota proportionally.

Note the token-per-minute unit. This is not a request-per-second limit in the style of some US providers — it means your limit scales with the size of the prompts you send, so a workload with large inputs consumes budget faster than one with small inputs even at the same request rate.

Performance: what is guaranteed and what is not

Serverless targets a 99.9% monthly availability rate under the Generative APIs service level agreement. Dedicated Deployment currently has no published SLA — Scaleway says it is still defining one — which is worth noting if you are building a dedicated deployment into a contract with uptime commitments.

On latency, Scaleway publishes order-of-magnitude guidance rather than guarantees, and the guidance is specific: with streaming enabled, time to first token should be under 1 second for most standard chat queries with fewer than 1,000 input tokens, and output generation speed should exceed 100 tokens per second for recent small-to-medium models such as GPT-OSS 120B and Mistral Small 3.2 24B.

Two factors dominate in practice. Model size and architecture matter — smaller, more recent models perform better. And chat time-to-first-token grows proportionally with input context size past a threshold, which means the sub-1-second figure is specifically about short prompts. Scaleway explicitly does not strictly guarantee performance, and it recommends Dedicated Deployment for applications with critical performance requirements.

For a rough throughput sanity check, we compared the published rates against the same models on OpenRouter's rate card. On a like-for-like USD basis, GLM-5.3 on OpenRouter sits near USD 3.60/M output against Scaleway's EUR 5.50/M output, and Mistral Small 3.2 sits near USD 0.20/M against Scaleway's EUR 0.35/M. Once the EUR-USD rate and the fact that OpenRouter carries its own routing layer are taken into account, the two are much closer than the raw numbers suggest. Scaleway's advantage is not necessarily the cheapest token — it is the residency, the EUR invoicing, and the first-party relationship.

Licenses, compliance, and the honest gaps

Model licenses apply on both tiers. You must comply with the license of whichever model you call, and for catalog models you accept the applicable license — including any EULA — before creating a dedicated deployment. For custom models you import yourself, compliance is your responsibility. The licenses are listed per model in the documentation and in the console Playground.

Now the gaps, stated plainly.

The catalog is small and moves. Seventeen serverless models is narrow by 2026 standards, and the lineup changed materially in the 42 days since our last crawl — GLM-5.3 and DeepSeek V4.1 Flash are new entries, Qwen 3 235B Thinking 2507 is gone from serverless. If your architecture depends on one specific model, Scaleway is a single point of failure that changes under you. Mitigate with a fallback provider.

No mainland-China path. There is no Chinese region, no ICP filing, and no local endpoint. Production traffic serving China needs an overseas egress or an aggregator. A multi-upstream OpenAI-compatible router such as FreeModel lets you route China-facing traffic to a domestic upstream while keeping the Scaleway EU endpoint as the fallback for EU-bound workloads, without maintaining two separate client implementations.

No dedicated SLA yet. If you are buying dedicated deployment for predictable performance, you are buying a 99.9% serverless SLA that does not apply to your dedicated instance. Price that risk accordingly, or keep a serverless fallback live.

EUR pricing is not convertible on the fly. Every rate here is in euros before tax. If your finance team budgets in USD, the FX exposure is yours unless you hedge it. That is a cost of the sovereignty guarantee, and for most EU-based buyers it is the point rather than a drawback.

Verdict

Scaleway Generative APIs is the answer to a question most API catalogs do not ask: where do I get inference that is genuinely European-hosted, first-party, priced in euros, and covered by a 99.9% SLA without routing through an aggregator who is not telling me where my data actually sits? The free million tokens remove all risk from a first test, the OpenAI-compatible surface makes that test about an afternoon, and the -50% batch rate with no rate limit is a genuinely good deal for offline work.

It is not the cheapest way to buy tokens, and it does not have the widest catalog. If your only requirement is the lowest cost per million tokens for a specific model, look at Groq, Cerebras, DeepInfra, or Fireworks. If your requirement is "the provider of record has an EU footprint, a published SLA, and no hidden reseller margin," Scaleway is one of very few options that satisfies all three at once.

Start with the free tier on one embedding model, move a real chat workload to Mistral Small 3.2 or GPT-OSS 120B, then measure the actual EUR-per-task cost against your current provider. If the residency line item has real money attached to it in your procurement, the comparison may not even matter — but make it anyway, because the first million tokens are free and the experiment costs you an afternoon.

Sources: Scaleway Model-as-a-Service pricing page, Scaleway Generative APIs documentation (FAQ, supported models, OpenAI compatibility), and the OpenRouter model rate card for cross-provider comparison. All verified October 29, 2026.