European Alternatives to OpenAI API
Looking for a European alternative to the OpenAI API? OpenAI Group PBC is a Delaware company. A European region exists, but it needs OpenAI approval and contract steps, and by default prompts and responses can sit in abuse monitoring logs for up to 30 days.
Fourteen European providers serve open-weight models behind an OpenAI-compatible endpoint, from cloud companies, gateways and small inference specialists in Sweden, Austria, Denmark, Switzerland, France, Germany, the Netherlands and Slovakia.
How we rank these tools — 4-step process
-
1
European ownership, verified
The company is headquartered and incorporated in the EU, EEA or Switzerland, and processes customer data in Europe. A US parent company disqualifies a tool from this page regardless of where its servers are.
-
2
Category fit and hands-on review
What the tool actually does, who it suits, and where it falls short — checked against the vendor’s own documentation, changelog and pricing page rather than its marketing copy.
-
3
Compliance and pricing check
GDPR posture, hosting location and the prices quoted on this page are verified against the vendor’s public pricing before publication, and re-checked when we revisit the category.
-
4
Position on this page
Placement on this page can be paid, and that can affect which tools appear here and the order they appear in. It never buys a good review: a tool that fails the checks above is not here at any price, and payment does not change the shortcomings we write about. A vendor can ask us to correct a factual error — not to remove a criticism.
Vendors can pay for visibility on this page. It never changes what an entry says about a product, including the criticism, and we earn nothing when you click through to a vendor. Paid placement can affect which tools appear here and the order they appear in. Editorial policy
14 European Alternatives to OpenAI API
Berget AI
Swedish inference provider with an OpenAI-compatible API, zero data retention stated on its home page and a published monthly plan from 25 euro
Cortecs
Vienna gateway that routes one API across European and other providers, with an EU-only routing option and payloads held only in volatile memory
2BA.AI
Danish OpenAI-compatible coding API with EU-hosted inference, no prompt logging and a flat 20 euro plan with 4,500 requests per rolling 5 hours
Infomaniak
Geneva cloud company whose AI Services give OpenAI-compatible access to open models from data centres in Switzerland, with no queries stored
Scaleway
Paris cloud provider whose Generative APIs serve open and Mistral models from one data centre in Paris, with prompts not collected or read
Mistral AI
Paris model maker with its own API platform, data hosted in the EU by default with an optional US endpoint and zero data retention on pay-as-you-go
IONOS
Montabaur cloud provider whose AI Model Hub runs all inference in German data centres, never logs prompts and engages no sub-processors
OVHcloud
Roubaix cloud provider whose AI Endpoints is a serverless OpenAI-compatible API on open models, with zero data retention beyond billing
STACKIT
Schwarz Group cloud from Bad Friedrichshall whose AI Model Serving offers an OpenAI-compatible API on open models and states that no customer data is stored
T-Systems AI Foundation Services
An enterprise LLM platform that runs open models on T-Cloud in Germany, with a high monthly minimum
Nebius AI Studio
Amsterdam AI cloud whose Token Factory serves open models over an OpenAI-compatible API, with a fixed EU region only on dedicated endpoints
Heabsy
Bratislava company with an OpenAI- and Anthropic-compatible inference API that separates models on its own EEA hardware from routed models that may run outside the EEA
GreenPT
Utrecht provider of an OpenAI-compatible API for open-weight models, hosted by Scaleway in France with AI sub-processors in Finland
Melious AI
Saarbrücken gateway to more than 60 open-weight models through an OpenAI- and Anthropic-compatible API, routed across EU providers in eight countries
Key takeaways
- The OpenAI API has a European data residency option (eu.api.openai.com), but it needs OpenAI approval, a Modified Retention amendment and carries a 10 percent uplift on models released from 5 March 2026. By default abuse monitoring logs are kept up to 30 days.
- Fourteen European providers are listed. Almost all expose an OpenAI-compatible endpoint, so the switch is a base URL, a key and a model name. Mistral AI is the exception: the pages read describe its own API and SDKs and make no OpenAI-compatibility claim.
- Berget AI (Sweden) and Cortecs (Austria) lead because they are small, independent and state their data handling in their terms. Infomaniak (Switzerland), Scaleway (France) and IONOS (Germany) run their own data centres and name the country.
- Two entries carry a caveat on location: Nebius AI Studio public endpoints show region Global, and Cortecs routes to non-EU providers outside its Sovereign Cloud mode.
- 2BA.AI is a Danish flat-rate coding API at 20 euro a month. Its terms name no company, so its country comes from the vendor and cannot be checked on its pages.
- Expect different models, not GPT. Test your own prompts before you move traffic, and keep a fallback while you do.
Why people leave OpenAI API
The OpenAI API is the default way to put a language model behind a product, and its chat completions format has become the shape most other providers copy. It is a US service, sold by OpenAI Group PBC, a Delaware public benefit corporation. For European buyers the question is rarely whether it works. It is where the prompts are processed, who holds them, and under whose law a dispute is settled.
OpenAI does offer a European option, and a fair page has to say so. A project can be set to a European region and use eu.api.openai.com, which keeps storage and inference in the EEA and Switzerland.
It is not self-serve: OpenAI says you must contact sales to check eligibility, be approved for Modified Abuse Monitoring or Zero Data Retention, and sign a Modified Retention amendment, and models released after 5 March 2026 carry a 10 percent uplift on data residency endpoints. Without those steps, abuse monitoring logs that can contain prompts and responses are kept for up to 30 days by default.
The fourteen providers on this page take a different route. They are European companies that serve mostly open-weight models behind an endpoint that follows the OpenAI format, so a switch is usually a new base URL, a new key and a new model name. What you give up is GPT itself: you choose between models such as Qwen, GLM, Gemma, Llama, gpt-oss and Mistral's own, and you test them on your own prompts.
- EU processing without an approval process On the OpenAI API, European processing is a project setting that OpenAI must approve and that comes with extra contract steps and a price uplift on newer models. On IONOS the data handling page says all inference is performed in its data centres in Germany, on Scaleway the region is Paris, and Infomaniak runs its AI services in Switzerland. There, the location is the default and not a negotiated exception.
- Prompt retention that is stated, not requested OpenAI keeps abuse monitoring logs with prompts and responses for up to 30 days unless you are approved for Zero Data Retention. IONOS says prompts and outputs are never logged, Scaleway applies a Zero Data Retention policy by default, Berget AI says it never stores prompt or output content, and Infomaniak says no queries are stored. These are the providers' own statements, so read the terms before you rely on them, but the starting point is retention off.
- A European contract and court Berget AI contracts under Swedish law, Cortecs under Austrian law with the court in Vienna, GreenPT under Dutch law with the court in Amsterdam, and Mistral AI is a French company registered in Paris. That settles which legal system a data protection dispute lands in, which matters if your customers ask for a processor in Europe.
- Model choice instead of one vendor A gateway such as Cortecs or Melious AI routes one key across many providers, and the cloud providers each serve a rotating set of open models. Models are retired on a schedule (IONOS lists Llama 3.1 405B as retired on 15 September 2026), so the benefit is being able to move to the next model without changing contract, as long as you keep your own evaluation set.
What you have to replace, not just match
Start by listing what you actually call. Most applications use chat completions, streaming, function calling and embeddings, and almost every provider here covers those through the OpenAI format. Features specific to OpenAI, such as its newest GPT models, the Responses API features or hosted tools, usually have no equivalent, so check each one before you commit.
Run the new provider beside OpenAI on a sample of real prompts before you move traffic. Quality is the real difference, not the endpoint: an open-weight model can match GPT on your task or fall short, and only your own test set will tell you. Move one low-risk workload first, keep OpenAI as a fallback during the trial, and then decide.
The alternatives compared
| Position | Tool | Headquarters | Pricing | Jurisdiction |
|---|---|---|---|---|
| #1 | Berget AI | Sweden | Trial €0 with €5 starting credits / Starter €25 per month / Team €50 per month / Dedicated endpoints via sales | EU (Sweden) |
| #2 | Cortecs | Vienna, Austria | Pay as you go at the provider price plus a 5% flat service fee, billed in euro by card or invoice | EU (Austria) |
| #3 | 2BA.AI | Denmark | €20 per month, up to 4,500 requests in a rolling 5-hour window | EU (Denmark, as stated by the vendor; the terms name no entity) |
| #4 | Infomaniak | Geneva, Switzerland | From CHF 5.75/month (kSuite); kDrive Solo free | Switzerland (adequacy decision, outside the EEA) |
| #5 | Scaleway | Paris, France | From about €0.0025/hour (DEV1-S) | EU (France) |
| #6 | Mistral AI | Paris, France | Free tier / pay-per-use API | EU (France) |
| #7 | IONOS | Montabaur, Germany | PostgreSQL from about €0.06/core/hour plus €0.007/GB RAM/hour, plus storage from about €0.04-0.15/GB per 30 days depending on tier; MongoDB and In-Memory DB priced separately | EU (Germany) |
| #8 | OVHcloud | Roubaix, France | PostgreSQL essential from about €0.16/hour (2 vCPU / 8GB RAM); MySQL, MongoDB, Kafka, Cassandra, OpenSearch, M3DB and Valkey priced separately by engine and tier | EU (France) |
| #9 | STACKIT | Bad Friedrichshall, Germany | EU (Germany) | |
| #11 | Nebius AI Studio | Amsterdam, Netherlands | Pay per generation | EU (Netherlands) |
| #12 | Heabsy | Bratislava, Slovakia | Per million tokens in US dollars, for example gpt-oss-20b at $0.05 input and $0.20 output / no minimum spend or subscription | EU (Slovakia) |
| #13 | GreenPT | Utrecht, Netherlands | API needs a Pro, Teams or API-only subscription / chat plans from €4.50 to €17.50 per month / DeepSeek V4.1 Flash at €0.22 input and €1.10 output per 1M tokens / glm-5.3 at €1.10 and €4.40 | EU (Netherlands; Dutch law and Amsterdam courts in its terms) |
| #14 | Melious AI | Saarbrücken, Germany | Pay as you go, for example DeepSeek V4.1 Flash at €0.20 input and €1 output per 1M tokens / coding plans Basic €49, Pro €149, Unlimited 3x €499, Unlimited 6x €999 per month, businesses only, plus VAT | EU (Germany) |
How each alternative compares to OpenAI API
- Which law reaches it. EU (Sweden). OpenAI Group PBC is incorporated in Delaware, United States, and run from San Francisco, California, so the CLOUD Act obliges the provider to hand over data on a valid order regardless of which country the servers are in.
- Where the data sits. The home page says models run under Swedish jurisdiction and data stays in the EU; the DPA lists Obenet AB for colocation and AI Iron AB for GPU bare-metal infrastructure, both Swedish affiliates; no city is named on the pages read.
- Source code. Closed source, as OpenAI API is.
- Independently checked. GDPR, NIS2 and Schrems II alignment claimed; DPA included with all plans.
Best for: Developer teams that want a drop-in OpenAI-format endpoint under Swedish law, with a documented DPA and a predictable monthly plan
Berget AI AB is a Swedish limited company, and its terms name its address in Stockholm and a Swedish organisation number. Its home page offers serverless inference through an OpenAI-compatible API plus dedicated endpoints for your own or fine-tuned models, and says it runs open models under Swedish jurisdiction with zero data retention. The code sample points the OpenAI SDK at api.berget.ai.
The DPA says all its data centres are within the EEA and that the customer chooses where data is processed. The terms are governed by Swedish law, with arbitration in Stockholm for customers outside Sweden.
Pricing is by plan: a Trial with EUR 5 of credits, Starter at EUR 25, Team at EUR 50 and Enterprise at EUR 500 a month, and Berget says it does not offer pure pay-as-you-go without a plan. Berget Code is a separate seat plan from EUR 150 a month.
What Berget AI does better than OpenAI API
- Small, independent and Swedish, with a DPA included in all plans
- States on its home page that it never stores prompt or output content or trains on them
- Published plans from EUR 25 a month and a trial with credits
- Dedicated endpoints for custom or fine-tuned models
Where Berget AI is a step down from OpenAI API
- Requires an active plan after the trial, so no pure pay-as-you-go
- Serves open models only, so no GPT
- The exact data centre city is not named on the pages we read
Standout against OpenAI API. It is the small, independent option that pairs a stated zero-retention position with a DPA and a plan you can budget.
- Which law reaches it. EU (Austria). OpenAI Group PBC is incorporated in Delaware, United States, and run from San Francisco, California, so the CLOUD Act obliges the provider to hand over data on a valid order regardless of which country the servers are in.
- Where the data sits. Depends on the provider chosen: the home page lists 16 active providers including IONOS, Scaleway, OVH, Nebius and Berget alongside Amazon, Google and Microsoft, and a Sovereign Cloud mode restricts routing to providers established in the EU.
- Source code. Closed source, as OpenAI API is.
- Independently checked. ISO certified, EU Sovereign, AI Act and GDPR claimed on its home page.
Best for: Teams that want one key, one euro invoice and EU-only routing across many models and providers
Cortecs GmbH is registered in Vienna, and its imprint gives a company register number and the Commercial Court of Vienna. It is a gateway: one API, 150+ models, 10+ EU providers, accepting OpenAI and Anthropic clients. Its Sovereign Cloud mode limits routing to providers established in the EU. The provider list on its home page also names Amazon, Google and Microsoft, so choose the mode on purpose.
The terms say Cortecs does not retain or log payload data within its own infrastructure, which is processed in volatile memory. Pricing is the upstream provider price plus a flat 5 percent service fee, billed in euro by card or invoice. The contract is under Austrian law with the Vienna courts. Account data uses Auth0, a US company, for sign-in, and the web platform is hosted at OVH in France, per its privacy policy.
What Cortecs does better than OpenAI API
- One key across many providers, with budgets, fallbacks and routing policies
- EU-only routing mode and euro billing by card or invoice
- Accepts OpenAI and Anthropic clients
- Austrian law and Vienna courts in the terms
Where Cortecs is a step down from OpenAI API
- A gateway adds another company to the data path, so the sub-processor list matters
- Outside Sovereign Cloud mode requests can reach non-EU providers
- Sign-in uses a US vendor (Auth0) for account data
Standout against OpenAI API. It is the gateway to pick when you want one contract and an EU-only routing switch over many models.
2BA.AI
a Danish flat-rate OpenAI-compatible API for coding agents, EU-hosted, with unverified company details
- Which law reaches it. EU (Denmark, as stated by the vendor; the terms name no entity). OpenAI Group PBC is incorporated in Delaware, United States, and run from San Francisco, California, so the CLOUD Act obliges the provider to hand over data on a valid order regardless of which country the servers are in.
- Where the data sits. European Union: the privacy policy says API inference, including processing of inputs and outputs, is hosted within the EU; payment and account services have separate providers (Stripe processes subscription payments).
- Source code. Closed source, as OpenAI API is.
Best for: Developers who use coding agents such as Cursor, Claude Code, Aider or Continue and want a flat monthly price with EU-hosted inference
2BA.AI sells API access for coding agents at EUR 20 a month for 4,500 requests per rolling 5 hours. It offers an OpenAI-compatible Chat Completions endpoint at api.2ba.ai/v1 and an Anthropic Messages endpoint, with an installer that configures common coding tools.
Its terms and privacy policy say API inference, including input and output, is hosted within the European Union, that request content is processed transiently and not stored as prompt history, and that inputs and outputs are not used for training. Zero logging applies to request content, and token counts and timestamps are kept as operational metadata.
Read the limits. The terms and privacy policy do not name the company, its address or a registration number: they say the provider is the one identified in your order or billing documentation. The Danish origin comes from the vendor and cannot be checked on its pages.
The quality figures on its site, 88 percent for its Amber model and 86 percent for GPT-5.6 Sol, come from its own evaluation of 50 SWE-bench Verified tasks, so treat them as the vendor's test, not an independent benchmark. It is a flat quota plan, not a per-token API for production workloads.
What 2BA.AI does better than OpenAI API
- Flat EUR 20 a month price with a clear rolling request quota
- States EU-hosted inference and no storage of request content
- OpenAI and Anthropic-style endpoints plus an installer for coding tools
- Cancel renewal any time through the billing portal
Where 2BA.AI is a step down from OpenAI API
- Company name, address and registration are not published on its own pages
- Quota plan for coding agents, not a metered API for general workloads
- Quality figures come from the vendor's own 50-task evaluation
Standout against OpenAI API. It is the only option here built as a flat monthly plan for coding agents, with EU inference stated.
Infomaniak
a Geneva cloud company with an OpenAI-compatible AI service run from its own Swiss data centres
- Which law reaches it. Switzerland (adequacy decision, outside the EEA). OpenAI Group PBC is incorporated in Delaware, United States, and run from San Francisco, California, so the CLOUD Act obliges the provider to hand over data on a valid order regardless of which country the servers are in.
- Where the data sits. Switzerland (Geneva, Winterthur).
- Source code. Closed source, as OpenAI API is.
- Independently checked. ISO 27001.
Best for: Teams that accept Switzerland as the hosting country and want pay-per-use pricing in Swiss francs
Infomaniak Network SA is registered in Geneva, and its legal notice gives the commercial register and VAT numbers. Its AI Services page offers open-source models through OpenAI-compatible APIs, says no queries are stored, and says its AI and GPU services are powered from its own data centres in Switzerland. The page also lists embeddings, audio and image models.
Pricing is pay per use with no commitment and a card for identity, billed monthly on tokens, and new accounts get 1M free credits. Examples on the price page are CHF 0.20 per 1M input tokens and CHF 0.40 output on Gemma 4 31B, and CHF 0.40 input and CHF 3.20 output on Qwen3.5-122B. The terms of the AI service were not readable at the address we tried, so the governing law is not confirmed here.
What Infomaniak does better than OpenAI API
- Own Swiss data centres and a published price list in Swiss francs
- 1M free credits to try it without a call
- Says no queries are stored
- Covers LLMs, embeddings, audio and image models in one account
Where Infomaniak is a step down from OpenAI API
- Switzerland is outside the EU, so check that it meets your legal test
- Open models only, so no GPT
- The AI service terms were not readable, so governing law is unconfirmed here
Standout against OpenAI API. It is the one with Swiss data centres and a franc price list that you can budget from.
Scaleway
a Paris cloud provider whose Generative APIs serve open models from Paris with zero data retention by default
- Which law reaches it. EU (France). OpenAI Group PBC is incorporated in Delaware, United States, and run from San Francisco, California, so the CLOUD Act obliges the provider to hand over data on a valid order regardless of which country the servers are in.
- Where the data sits. Paris, Amsterdam, Warsaw — EU only.
- Source code. Closed source, as OpenAI API is.
Best for: Teams that want a European cloud provider with an OpenAI-compatible API, a named region and a free token allowance
Scaleway documents an OpenAI API compatibility page for its Generative APIs, and its data privacy page names Scaleway at an address in Paris as the data processor.
By default it applies a Zero Data Retention policy and says it does not collect, read, reuse or analyse the content of your inputs, prompts or outputs, with an exception when traffic harms the service. Aggregated anonymised usage data is kept for up to 6 months, and the page names Paris as the storage region. It also says that as a European company its AI services are not subject to the US CLOUD Act, which is its own claim.
Serverless pricing is per million tokens in euro. Examples on the pricing page are EUR 0.25 input and EUR 0.50 output for gemma-4-26b-a4b-it, and EUR 1.50 input and EUR 7.50 output for mistral-medium-3.5-128b. New customers get 1 million free tokens and 60 minutes of audio transcription, and the Batches API is priced 50 percent lower.
What Scaleway does better than OpenAI API
- Names the region (Paris) and applies zero data retention by default
- Free tier of 1 million tokens and 60 minutes of audio
- Euro price list per million tokens and 50 percent off batch requests
- Dedicated deployments available when shared capacity is not enough
Where Scaleway is a step down from OpenAI API
- Open and Mistral models only, so no GPT
- The CLOUD Act statement is the vendor's own view, not a legal opinion
- Larger and less specialised than the small inference providers
Standout against OpenAI API. It combines a named region, zero retention by default and a free token allowance, from a large European cloud.
Mistral AI
a Paris model maker with its own API, EU hosting by default and an optional US endpoint
- Which law reaches it. EU (France). OpenAI Group PBC is incorporated in Delaware, United States, and run from San Francisco, California, so the CLOUD Act obliges the provider to hand over data on a valid order regardless of which country the servers are in.
- Where the data sits. EU (France).
- Source code. Open source, where OpenAI API is not: you can read what it does rather than take the description on trust.
- Independently checked. GDPR.
Best for: Teams that want the model itself to be European, not just the hosting
Mistral AI's commercial terms say its products are provided by Mistral AI, a French limited joint-stock corporation incorporated in Paris under number 952 418 325, at 15 rue des Halles.
Its Help Center says data is hosted in the European Union by default, that a US API endpoint can be used on purpose and then data is hosted in the United States, and that depending on the feature, data can be temporarily transferred outside the EU to sub-processors listed in its Trust Center, under Standard Contractual Clauses.
Mistral prices its API per million tokens, with input and output counted separately and Mistral Large given as an example at USD 0.5 input and USD 1.5 output, with batch processing 50 percent lower. The Free plan lists USD 10 a month in API credits.
Mistral publishes its own API and SDKs, and the pages we read make no OpenAI-compatibility claim, so test your client before you assume a base-URL swap works. Its Help Center has an article on activating zero data retention, so check the terms for your account and product.
What Mistral AI does better than OpenAI API
- The only entry that builds the models itself, in France
- EU hosting by default, stated in its Help Center
- Per-token pricing with batch discounts and free monthly API credits
- French company number and address in its terms and privacy policy
Where Mistral AI is a step down from OpenAI API
- No OpenAI-compatibility claim on the pages read, so a drop-in swap is not assured
- Data can be transferred outside the EU to listed sub-processors
- Has a US endpoint, so you must choose the region on purpose
Standout against OpenAI API. It is the one where the model, the company and the default hosting are all European.
IONOS
a German cloud whose AI Model Hub performs all inference in Germany, never logs prompts and names no third-party sub-processors
- Which law reaches it. EU (Germany). OpenAI Group PBC is incorporated in Delaware, United States, and run from San Francisco, California, so the CLOUD Act obliges the provider to hand over data on a valid order regardless of which country the servers are in.
- Where the data sits. ISO 27001-certified European data centres.
- Source code. Closed source, as OpenAI API is.
- Independently checked. ISO 27001-certified data centres (stated).
Best for: Teams with strict data location rules that need a German provider with an OpenAI-compatible API
IONOS documents an OpenAI-compatible API for its AI Model Hub. Its data handling page says all inference is performed in IONOS CLOUD data centers in Germany, that prompts and outputs are never logged, never accessed by personnel, not used for training and not disclosed to third parties, including the developers of the models. It says no third-party sub-processors are engaged and that operational access is limited to personnel within the European Union.
It names ISO 27001 for the data centers and says the AI Model Hub falls within the BSI C5 scope. The contracting company depends on your country: IONOS CLOUD GmbH in Germany, IONOS CLOUD Solutions Spain SLU, IONOS CLOUD France SAS or IONOS CLOUD UK Ltd. The model catalogue changes: Llama 3.1 405B was retired on 15 September 2026 and its replacement is named. We did not verify IONOS prices, so check its price list.
What IONOS does better than OpenAI API
- States that all inference happens in German data centers
- Says prompts and outputs are never logged or used for training
- No third-party sub-processors for the AI Model Hub
- Publishes a model retirement schedule with a named replacement for each model
Where IONOS is a step down from OpenAI API
- Models are retired on a schedule, so you must re-test periodically
- Prices were not verified here, so check before you budget
- Contracting company varies by country, including a UK entity
Standout against OpenAI API. It makes the strongest stated data handling claim of the cloud providers: German-only inference, no logs and no sub-processors.
OVHcloud
a French cloud whose AI Endpoints is a serverless OpenAI-compatible API with a public price list
- Which law reaches it. EU (France). OpenAI Group PBC is incorporated in Delaware, United States, and run from San Francisco, California, so the CLOUD Act obliges the provider to hand over data on a valid order regardless of which country the servers are in.
- Where the data sits. 30+ data centres, majority in Europe.
- Source code. Closed source, as OpenAI API is.
- Independently checked. SecNumCloud (ANSSI), GAIA-X founding member.
Best for: Teams that want a large French provider, a published euro price list and a free credit to test
OVHcloud's AI Endpoints product page lists an OpenAI-compatible API, and says it keeps only the data required for billing, as zero data retention.
Its documentation says the infrastructure is located in Gravelines, France, that it is designed to be compatible with the OpenAI API, and that data is not stored or shared during or after model use. The product page also lists "Worldwide availability (non-EU)" next to the plan, which is ambiguous, so ask OVHcloud to confirm in writing where inference runs for your account.
The catalogue publishes prices: gpt-oss-120b at EUR 0.08 per million input tokens and EUR 0.40 output, Qwen3.8-27B at EUR 0.4 and EUR 2.7, and Qwen3.5-397B-A17B at EUR 0.6 and EUR 3.6. New Public Cloud projects get USD 200 of free credit valid for a month. OVH Groupe SA, Roubaix, is named in its voucher terms.
What OVHcloud does better than OpenAI API
- Public euro price per million tokens for every model
- USD 200 free credit to test
- Documentation names Gravelines, France as the infrastructure location
- Says only billing data is kept
Where OVHcloud is a step down from OpenAI API
- The "Worldwide availability (non-EU)" label on the product page is unclear
- Open models only, so no GPT
- A flow limit of 400 requests per minute per project on the base plan
Standout against OpenAI API. It is the one with a full public price list per model, backed by documentation that names Gravelines.
STACKIT
a Schwarz Group cloud from Germany with an OpenAI-compatible AI Model Serving and a public price table
- Which law reaches it. EU (Germany). OpenAI Group PBC is incorporated in Delaware, United States, and run from San Francisco, California, so the CLOUD Act obliges the provider to hand over data on a valid order regardless of which country the servers are in.
- Where the data sits. Germany: the AI Model Serving pricing table names the region Germany South.
- Source code. Closed source, as OpenAI API is.
Best for: Teams that want a German hyperscaler, a stated region and tiered per-token prices
STACKIT AI Model Serving offers an inference API that is OpenAI compatible, with shared models used by all clients and no data exchange between projects or customers, according to its documentation. The documentation lists endpoints under api.openai-compat.model-serving.eu01.onstackit.cloud/v1, with rate limits such as 30 to 45 requests per minute and 350,000 tokens per minute on the models we read. Context windows of 200K to 262K tokens are listed.
The price table on its product page names the region Germany South and prices by tier: the standard tier at EUR 0.15 per million input tokens and EUR 0.25 output, the plus tier at EUR 0.65 output, and the premium tier at EUR 1.50 per million input tokens. The company contact is Schwarz Digits Cloud GmbH & Co. KG, Bad Friedrichshall. We did not read its retention terms, so we claim nothing about prompt storage.
What STACKIT does better than OpenAI API
- Names the region (Germany South) in a public price table
- Tiered per-token prices in euro
- Documented OpenAI-compatible inference API with listed rate limits
- Shared models with no data exchange between customers, per its docs
Where STACKIT is a step down from OpenAI API
- Modest default rate limits on the shared models
- Prompt retention was not read, so it cannot be compared with the zero-retention claims above
- Open models only, so no GPT
Standout against OpenAI API. It is the German hyperscaler with a public tiered price table that names the region.
T-Systems AI Foundation Services
an enterprise LLM platform that runs open models on T-Cloud in Germany, with a high monthly minimum
Best for: Large organisations that want a German enterprise vendor and can commit EUR 1,000 or more a month
T-Systems AI Foundation Services offers an OpenAI-compatible API at llm-server.llmhub.t-systems.net, with 38 models listed. For open models such as GLM, Gemma, Mistral Small 4 and GPT-OSS 120B the catalogue says "Server Location: Germany" and "Data Processing: Germany" on the T-Cloud. Closed models are forwarded to their upstream provider on Azure or Google Cloud, and some are listed with a worldwide location, so check each model's row.
Plans are by monthly minimum commitment: Essential at EUR 1,000, Professional at EUR 3,000 and Agentic at EUR 5,000, with example prices of EUR 0.20 input and EUR 0.65 output per million tokens on T-Cloud hosted models. The minimum is the floor of the monthly bill, not an extra fee. A free trial key is available, and the service says it is ISO 27001 certified.
What T-Systems AI Foundation Services does better than OpenAI API
- Lists server location and data processing country for each model
- Enterprise plans with rate limits and a free trial key
- Open models on a German cloud with an OpenAI-compatible endpoint
Where T-Systems AI Foundation Services is a step down from OpenAI API
- EUR 1,000 monthly minimum puts it out of reach for small teams
- Closed models run on Azure or Google, some with a worldwide location
- The parent company is not stated on the pages we read
Standout against OpenAI API. It is the enterprise-priced option that states a location for every model, including the ones that are not in Germany.
Nebius AI Studio
an Amsterdam AI cloud with an OpenAI-compatible Token Factory, but public endpoints have no fixed region
- Which law reaches it. EU (Netherlands). OpenAI Group PBC is incorporated in Delaware, United States, and run from San Francisco, California, so the CLOUD Act obliges the provider to hand over data on a valid order regardless of which country the servers are in.
- Where the data sits. Public endpoints: no region commitment (global routing); EU region only on dedicated endpoints.
- Source code. Closed source, as OpenAI API is.
Best for: Teams that will use dedicated endpoints in a chosen region, or that need scale beyond smaller providers
Nebius Token Factory serves open models through an OpenAI-compatible API at api.tokenfactory.nebius.com/v1, with dedicated endpoints, batch inference and post-training on the same platform. Its legal quick guide is direct about location: for public model endpoints the region field shows Global, processing location is decided dynamically and can change without notice, and public endpoints are not intended for use cases that need a predictable region. Dedicated endpoints can be deployed in a region you choose.
Zero Data Retention is an account setting that is off by default. Without it, inputs and outputs are kept for speculative decoding and stored in Finland, and with it nothing is retained. The contracting party is the applicable Nebius entity within the group headed by Nebius Group N.V., which is listed on Nasdaq. We did not verify its prices.
What Nebius AI Studio does better than OpenAI API
- Clear documentation of where data is processed and stored
- Dedicated endpoints in a chosen region, plus batch inference and post-training
- ZDR available at organisation level, covering all projects
Where Nebius AI Studio is a step down from OpenAI API
- Public endpoints show region Global, so they do not meet an EU-only test
- Zero Data Retention is off by default and may reduce inference speed
- Contracting entity is not named on the pages we read, and prices were not verified
Standout against OpenAI API. It is the entry that tells you plainly its shared endpoints are not region-bound.
- Which law reaches it. EU (Slovakia). OpenAI Group PBC is incorporated in Delaware, United States, and run from San Francisco, California, so the CLOUD Act obliges the provider to hand over data on a valid order regardless of which country the servers are in.
- Where the data sits. Two tiers per its catalog: models on its own GPUs inside the EEA with zero data retention, and routed models fulfilled through global providers whose compute may run outside the EEA.
- Source code. Closed source, as OpenAI API is.
- Independently checked. SOC 2 listed as a roadmap item, not claimed as held.
Best for: Teams that want an OpenAI- and Anthropic-compatible endpoint with one flagship model on EEA GPUs and no subscription
Heabsy is operated by FEYA, s.r.o. in Bratislava, Slovakia, and is an OpenAI- and Anthropic-compatible inference API. Its flagship model, Qwen3.8 27B, runs on dedicated GPUs in EEA data centres with zero data retention, and its DPA says inference on Heabsy-served models is provisioned entirely within the EEA. Pricing is pay as you go with no minimum or subscription: USD 0.40 per million input tokens, USD 0.05 cached input and USD 3.00 output on the flagship.
The rest of the catalogue is fulfilled through other providers, whose compute may run outside the EEA, and the catalogue marks per model whether it runs on Heabsy's own hardware. It is a young platform with no published founding date or team page.
What Heabsy does better than OpenAI API
- Own EEA GPUs and zero retention on the flagship model
- Pay as you go with no subscription or minimum spend
- OpenAI and Anthropic compatible, with per-model location marking
Where Heabsy is a step down from OpenAI API
- Only one model runs on its own EEA hardware
- Routed models may run outside the EEA
- Young platform with little public history and prices listed in US dollars
Standout against OpenAI API. It separates which models run on its own EEA hardware from which are routed, model by model.
GreenPT
a Dutch provider of an OpenAI-compatible API for open models, hosted by Scaleway with AI sub-processors in Finland
- Which law reaches it. EU (Netherlands; Dutch law and Amsterdam courts in its terms). OpenAI Group PBC is incorporated in Delaware, United States, and run from San Francisco, California, so the CLOUD Act obliges the provider to hand over data on a valid order regardless of which country the servers are in.
- Where the data sits. The privacy policy says personal data is processed in EU data centres in France; its sub-processor table names Scaleway SAS (France) for hosting and Inceptron, Tensorix and Melious on Verda data centres in Finland for AI hosting.
- Source code. Closed source, as OpenAI API is.
- Independently checked. ISO 27001 and GDPR compliance claimed on its site.
Best for: Teams that want an API with a sustainability angle and a Dutch contract
GreenPT BV is based in Utrecht and registered at the Chamber of Commerce, and its terms are governed by Dutch law with the Amsterdam courts. Its API is OpenAI compatible, runs open-weight models and offers embeddings, reranking, OCR and speech to text. Its privacy policy says API payloads are processed in memory and not stored, and lists Scaleway in France for hosting and Inceptron, Tensorix and Melious on Verda data centres in Finland for AI hosting.
The API needs an active subscription, including an API-only account, and there is a 14-day free trial without a card. Its site announces EU and US regions, so choose the endpoint on purpose.
What GreenPT does better than OpenAI API
- Dutch company, Dutch law and a named sub-processor table
- API payloads processed in memory only, per its privacy policy
- 14-day free trial without a card
Where GreenPT is a step down from OpenAI API
- AI hosting is split across several sub-processors, so the data path is longer
- Needs a subscription before API use
- A US region now exists, so the endpoint choice matters
Standout against OpenAI API. It is the Dutch option that publishes a full sub-processor table with locations.
- Which law reaches it. EU (Germany). OpenAI Group PBC is incorporated in Delaware, United States, and run from San Francisco, California, so the CLOUD Act obliges the provider to hand over data on a valid order regardless of which country the servers are in.
- Where the data sits. The home page says it runs entirely on EU-based infrastructure across 11 providers in 8 countries; the DPA lists IONOS SE (Germany), Scaleway SAS (France), DataCrunch Oy (Verda, Finland) and others as inference sub-processors.
- Source code. Closed source, as OpenAI API is.
- Independently checked. Described as independently checked by staysin.eu on its home page.
Best for: Businesses that want one OpenAI- and Anthropic-compatible key across many EU providers, with coding plans
Melious AI GmbH is registered in Saarbrucken, Germany, with a commercial register number in its legal notice. It offers more than 60 open-weight models through an OpenAI- and Anthropic-compatible API, routed across 11 providers in 8 countries, and says it runs entirely on EU-based infrastructure. Its DPA lists IONOS, Scaleway, Nebius Group N.V. and Verda among its inference sub-processors.
Pricing is pay as you go, for example EUR 0.2 per million input tokens and EUR 1 output on DeepSeek V4.1 Flash, and coding plans from EUR 49 a month, offered to businesses.
What Melious AI does better than OpenAI API
- One key across 60+ open-weight models and 11 EU providers
- OpenAI and Anthropic compatible with failover built in
- Coding plans from EUR 49 a month and pay-as-you-go pricing
Where Melious AI is a step down from OpenAI API
- A gateway adds sub-processors, and one of them (Nebius) has region-free public endpoints
- Plans are for businesses only
- Prompt retention is not described on the pages we read
Standout against OpenAI API. It is the German gateway that routes one key across the most open models, with its sub-processors listed in its DPA.
What actually breaks when you switch
The hard part is quality, not the code. A new base URL takes minutes, but an open-weight model can answer differently from GPT, call tools less reliably or handle your language less well. Build a small evaluation set from your real prompts before you move traffic, and compare on that, not on a leaderboard.
Check each provider's compatibility notes. Chat completions and streaming usually work. Newer OpenAI features, certain parameters, structured output modes and hosted tools may not, and Mistral AI describes its own API. Expect to change some code even where the endpoint looks identical.
Models are not permanent. IONOS publishes a retirement schedule and retired Llama 3.1 405B on 15 September 2026, and every provider rotates its catalogue. Pin a model version where you can, keep your evaluation set, and plan to re-test when a model is retired.
Can I switch from the OpenAI API by changing the base URL?
For most providers on this page, yes. Scaleway, IONOS, OVHcloud, STACKIT, Infomaniak, Berget AI, GreenPT, 2BA.AI, Nebius AI Studio, T-Systems, Heabsy and Melious AI describe an OpenAI-compatible endpoint, and Cortecs accepts OpenAI and Anthropic clients. You change the base URL, the key and the model name, and the official OpenAI SDKs keep working.
Mistral AI is the exception on the pages we read: it publishes its own API and SDKs without an OpenAI-compatibility claim, so test your client first. Compatibility also stops at the common endpoints. Do not assume every parameter or newer OpenAI feature is supported, and read each provider's compatibility notes.
Doesn't OpenAI already offer data residency in Europe?
Yes. The OpenAI documentation describes European data residency for the API (EEA plus Switzerland) with eu.api.openai.com, covering storage and inference for endpoints such as chat completions. It is subject to eligibility checks by OpenAI sales, requires approved abuse monitoring controls and a Modified Retention amendment, and adds a 10 percent uplift on models released on or after 5 March 2026.
It also excludes system data, such as billing, analytics and support data, from the region. The case for a European provider is therefore not that OpenAI has nothing, but that the location and retention are the default, the contract is European, and there is no approval step. If OpenAI's residency meets your legal team's test, it is a valid answer and keeps you on GPT models.
What does "EU-hosted" mean for a gateway such as Cortecs or Melious AI?
A gateway sends each request to one of several inference providers, so the data path depends on the route. Cortecs offers a Sovereign Cloud mode that uses only providers established in the EU, and lists Amazon, Google and Microsoft among its providers outside that mode. Melious AI says it runs on EU-based infrastructure across 11 providers in 8 countries, and its DPA lists the sub-processors.
The sub-processor list therefore matters more than the head office. Ask for it, check where each provider hosts, and pin the routing policy in your own configuration.
Which models can I run instead of GPT?
Mostly open-weight models. The catalogues we read include Qwen, GLM, Gemma, Llama, gpt-oss (OpenAI's own open-weight model), Kimi, DeepSeek and Mistral models, in sizes from small to several hundred billion parameters. Mistral AI serves its own models, and Cortecs and Melious AI also route to larger closed models in some modes.
None of these is GPT, so measure quality and tool-calling reliability on your task. Model lists change often; IONOS, for example, publishes a retirement schedule with a recommended replacement for each model.
Which one to pick
If you want the shortest path off the OpenAI API, start with Berget AI or Cortecs: small, European, with data handling stated in their own terms, and OpenAI-format endpoints. Cortecs is the one to take if you want one key over many providers, as long as you switch on its EU-only routing.
If you need the location named, IONOS (Germany), Scaleway (Paris) or Infomaniak (Switzerland) run their own data centres and say where. Mistral AI is the pick when the model itself should be European, with the caveat that its own API is not described as OpenAI-compatible. 2BA.AI is a flat-rate coding API at EUR 20 a month, with the caveat that its company details are not published on its own pages.
Be realistic about the trade-off. None of these is GPT, OpenAI's own European residency is a valid option if your legal team accepts it, and the best choice depends on your task. Test on your own prompts, keep a fallback during the trial, and decide on the data path and the contract, not on the endpoint.
Frequently Asked Questions
It depends on what you need to know about the data path. Berget AI (Sweden) and Cortecs (Austria) are small independents that state zero retention and routing policy in their own terms. IONOS, Scaleway and Infomaniak run their own data centres and name the country. Mistral AI is the choice when the model itself should be European. None of them is GPT, so test your prompts first.
OpenAI publishes a data processing framework and a European data residency option, and whether your use is lawful depends on your own legal basis and your contract.
The points European buyers raise are that residency needs OpenAI approval and a retention amendment, that abuse monitoring logs with prompts are kept up to 30 days by default, and that OpenAI is a US company. Choosing a European provider does not make you compliant by itself, but it removes some of those questions.
Infomaniak lists Swiss francs, for example CHF 0.20 per 1M input tokens and CHF 0.40 output on Gemma 4 31B. OVHcloud lists EUR 0.08 input and EUR 0.40 output on gpt-oss-120b.
Scaleway lists EUR 0.25 and EUR 0.50 on gemma-4-26b-a4b-it, and gives new customers 1,000,000 free tokens. STACKIT lists EUR 0.15 input and EUR 0.25 output on its standard tier. Berget AI has plans from EUR 25 a month, 2BA.AI is EUR 20 a month, and Cortecs charges the provider price plus 5 percent. Prices change, so check each site.
Yes, for coding agents. 2BA.AI sells EUR 20 a month for 4,500 requests per rolling 5 hours, and Melious AI lists business coding plans from EUR 49 a month with models such as GLM 5.3 and DeepSeek V4.1 Flash. Berget AI has Berget Code seats from EUR 150 a month. Flat plans have request or token limits, so they suit interactive coding better than a large unattended batch job.
Berget AI is Berget AI AB (Stockholm), Cortecs is Cortecs GmbH (Vienna), Infomaniak is Infomaniak Network SA (Geneva), Scaleway is a Paris company, Mistral AI is registered in Paris, IONOS is a German company, OVHcloud is OVH Groupe SA (Roubaix), STACKIT is Schwarz Digits Cloud GmbH & Co.
KG (Bad Friedrichshall), Heabsy is FEYA, s.r.o. (Bratislava), GreenPT is GreenPT BV (Utrecht) and Melious AI is Melious AI GmbH (Saarbrucken). Nebius AI Studio belongs to Nebius Group N.V., which is listed on Nasdaq. 2BA.AI does not name its company on its own pages.
Infomaniak is Swiss, so it is outside the EU. Its data centres are in Switzerland, which has an EU adequacy decision for personal data transfers, but its terms are not those of an EU member state. It is on this page because many European buyers accept Swiss hosting, and the page says so rather than hiding it.
Nebius AI Studio public endpoints show region Global and can change processing location without notice, while dedicated endpoints stay in the chosen region. Cortecs lists Amazon, Google and Microsoft among its providers outside its Sovereign Cloud mode.
Heabsy runs one model on its own EEA hardware and routes the rest of its catalogue through providers whose compute may sit outside the EEA. T-Systems forwards its closed models to Azure or Google. Mistral AI says data can be temporarily transferred outside the EU to the sub-processors in its Trust Center and also offers a US endpoint. GreenPT has announced EU and US regions, so choose the endpoint on purpose.
Scaleway gives new customers 1,000,000 free tokens and 60 minutes of audio transcription, and Infomaniak gives 1M free credits. OVHcloud gives USD 200 of credit to new Public Cloud projects for one month. Berget AI has a trial plan with EUR 5 of credits, Mistral AI lists USD 10 a month of API credits on its Free plan, T-Systems offers a free trial key, and GreenPT has a 14-day free trial. The others bill from the first use.
gpt-oss-120b and gpt-oss-20b are listed by OVHcloud and others, which means you can run an OpenAI-released open-weight model on a European provider. They are separate from the GPT models behind the OpenAI API, so evaluate them like any other open model.
Related comparisons
Who worked on this review
Explore More European Alternatives
Discover privacy-focused European alternatives to other popular US tech services.