EU Data Residency for AI APIs: What OpenAI, Azure, Bedrock, Vertex, Mistral and Claude Charge in 2026

Most EU buyers asking "can we keep this in the EU?" get a yes from every big AI vendor. What they rarely get is the price. It sits in a dozen places: a one-line uplift on OpenAI's pricing page, meter names in Microsoft's price feed, a second price table on Google's page, and per-region SKUs in AWS's price list.

So we read them all on 27 September 2026. For AWS and Microsoft, whose price pages load in JavaScript, we pulled the numbers from their own official price APIs. We didn't call these endpoints or deploy anything, and none of these vendors pay us. This page is the full residency table. For what a private LLM costs a company per month across every route, including running your own GPU, see our private LLM cost guide for EU companies.

EU vs global price, model by model

For frontier models the EU tier is a flat 10% on top of the global price, except GPT-6 on Azure, which is 20%. Cached input rises by the same share. On open models the EU price depends on the region you pin, not on "the EU".

Four notes on reading the table:

• OpenAI publishes the uplift, not the EU prices. Its pricing page says regional processing endpoints "are charged a 10% uplift for models released on or after March 5, 2026". We applied 10% to every token class. Azure's, AWS's and Google's EU prices are printed in their own feeds and pages.
• In euros, gpt-6-sol on OpenAI's EU endpoint is about €1.93 in and €9.65 out per million tokens; on Azure's Data Zone it is about €2.10 and €10.52.
• Gemini 3.8 Flash's lower price ends on 31 December 2026. Google marks it as promotional. Budget a year on the 2027 row.
• Mistral's Enterprise APIs, which add "regional data processing controls", cost "75% above list pricing on select APIs", Mistral says. It doesn't list which APIs, so we don't print per-model Enterprise prices.

The EU premium isn't one number

The uplift ranges from 0% to 75% depending on vendor, model generation and region. Azure charges 10% on GPT-5.6 and gpt-oss-120b but 20% on GPT-6. AWS charges the US price for gpt-oss-120b in Stockholm and a third more in Frankfurt.

What that means for a buyer:

• For GPT-6, OpenAI's own EU endpoint is cheaper than Azure's. gpt-6-sol costs $2.20 in on OpenAI's EU endpoint and $2.40 on Azure's EU Data Zone. Azure's advantage is that it's self-serve.
• For Claude, Bedrock and Vertex cost the same in the EU: +10% on every token class. The difference is batch pricing and region choice, covered below.
• For open models, the region is the price. If gpt-oss-120b is good enough, Bedrock in Stockholm charges the US East price. Our cost to run page for gpt-oss-120b shows what the same model takes on your own hardware.

How to turn EU residency on

Azure, AWS Bedrock and Google Cloud let you choose EU residency yourself when you deploy. Mistral is EU-hosted by default. OpenAI's EU residency needs a sales approval first. Anthropic's own API has no EU setting.

• OpenAI. Its docs say to "Contact our sales team to see if you're eligible for using data residency controls." For any region other than the US, you must also "be approved for abuse monitoring controls, and execute a Modified Retention amendment." Once approved, you pick the Europe region when you create a project, or send single requests to eu.api.openai.com with a key from a Global project.
• Azure (Microsoft Foundry). Deploy the model with the Data Zone Standard deployment type. Microsoft's price feed lists GPT-6 Data Zone meters in six EU regions: Sweden Central, Germany West Central, France Central, West Europe, Poland Central and Spain Central. Norway East, Switzerland North and UK South list Global meters only.
• AWS Bedrock. Call the model with its eu. inference profile. For a single country, AWS's model cards say to use the bedrock-mantle endpoint instead, which offers Claude Sonnet 5 in-region in Stockholm and Ireland only.
• Google Cloud. Claude's newer models run on the eu multi-region endpoint, not on single-country regions, according to Anthropic's docs. Gemini uses any non-global endpoint and is billed at the "Non-global" price.
• Mistral. Its help center says "By default, your data is hosted in the European Union." Just don't switch to its US endpoint.
• Anthropic API. Its data residency docs list two inference locations, global and us, and "Currently, "us" is the only available workspace geo." For Claude in the EU, go through Bedrock or Google Cloud.

Which regions count as "EU"

Each vendor's "EU" is a slightly different map. OpenAI's European region covers the EEA plus Switzerland. Azure's EU Data Zone and Bedrock's EU profile stay in EU member states if you deploy or call from an EU member-state region. If your contract says "EU-27 only", check the map, not the label.

• OpenAI names the region "Europe (EEA + Switzerland)". That includes Norway, Iceland, Liechtenstein and Switzerland. The UK is a separate region at gb.api.openai.com.
• AWS Bedrock publishes the destinations for Claude Haiku 4.5's EU profile. Called from Frankfurt, Stockholm, Milan, Spain, Ireland or Paris, it routes only to those six regions. Called from Zurich it adds Zurich, and called from London it adds London. The Sonnet 5 and Opus 5.5 cards say only that the EU profile "Keeps data within EU regions", without a destination list.
• AWS also warns that profile requests can reach opt-in regions you never enabled, and that "Your input prompts and output results may be stored in the opt-in Regions for abuse detection purposes."
• Google Cloud calls its EU multi-region eu. Anthropic's docs say specific single regions support "Claude Sonnet 4.6 and earlier; newer models use the global or multi-region endpoints."
• Mistral says data "can be temporarily transferred outside of the European Union" for some features, to listed sub-processors.

What EU residency doesn't cover

The EU tiers keep your prompts and outputs in the region, not everything. They exclude account and usage data, often drop the batch discount, and some models have no EU route at all.

• System data. OpenAI's docs say data residency "does not apply to system data", meaning account data, metadata and usage data "that do not contain Customer Content", which "may be processed and stored outside the selected region." Its list of limitations also excludes transfers caused by where your own users or infrastructure are.
• No batch discount for GPT-6 in the EU. OpenAI says "For GPT-6 Sol and Luna, EU data residency is available only with Standard processing", so the 50% Batch and Flex discounts are gone. Microsoft's price feed has no Batch meter for GPT-6 in any region, though older models have Data Zone Batch prices.
• No batch for Claude's newest models on Bedrock. AWS's price feed has no regional Batch price for Sonnet 5 or Opus 5.5, and the Opus 5.5 model card lists Batch as not supported. Google Cloud does list EU batch prices: Sonnet 5 at $1.10 in and $5.50 out, Opus 5.5 at $2.20 and $11. If you batch Claude in the EU, that's the route.
• Opus 5.5 can't be pinned to one country on Bedrock. Its card lists no European in-region option, only the multi-country EU profile.
• GPT-6 on Bedrock has no EU route. From every European region, AWS's card for GPT-6 Sol lists only global cross-region inference.
• Gemini 3.1 Pro Preview is global only on Google's price page. Gemini 3.8 Flash keeps an EU batch price: $0.825 in and $4.125 out from 2027.
• Mistral's zero retention excludes batch. Mistral says zero data retention is "available only with pay-as-you-go and only for stateless API calls", not for batch processing files.

Retention and training terms, vendor by vendor

Every vendor here says in its docs that it doesn't train on API data by default, except Mistral, whose pay-as-you-go wording doesn't say which way the setting starts. Zero data retention almost always needs an application or a sales conversation.

We couldn't read OpenAI's Data Processing Addendum: its page returned an access error to our reader on 27 September 2026. Ask for it through your account rather than relying on a summary.

What it adds to a monthly bill

At 10 million tokens a workday, the EU tier of gpt-6-sol costs about $63 to $127 more a month than the global price. That's $636 (€558) globally, $699 (€613) on OpenAI's EU endpoint and $763 (€669) on Azure's EU Data Zone, over 22 workdays with eight input tokens per output token.

Try your own volume. The calculator covers the EU routes above plus open models and your own GPU:

If the EU premium on a frontier model looks large next to an open model's price, our private LLM cost guide compares every route at three volumes, including when a GPU of your own pays off. More guides for companies putting AI to work are on our AI automation hub.

Does EU residency make you GDPR-compliant?

No, and GDPR doesn't require it. GDPR requires a written processor contract with the AI vendor and a lawful basis for any transfer outside the EU. EU residency removes the transfer question for prompts and outputs. It doesn't remove the contract, and it doesn't cover system data.

• A processor contract. Article 28 says you may use only processors giving "sufficient guarantees", under a written contract covering what's processed, for how long and on whose instructions. That's the vendor's Data Processing Agreement, needed whether inference runs in Paris or Oregon.
• A transfer basis. For US companies that participate in the EU–US Data Privacy Framework, the Commission's adequacy decision of 10 July 2023 lets data flow freely. Check each vendor on the official Data Privacy Framework list.
• Why buyers pay for residency anyway. The framework faces an appeal, case C-703/25 P, lodged on 31 October 2025, and we found no ruling as of September 2026. EU residency means a ruling against it wouldn't touch your prompts.

This is a summary of the law and official guidance, not legal advice.

Security reviews and compliance timelines for companies selling AI features are on our security and compliance hub.

The practical steps