EU data residency is no longer a vendor question. It is a per-model, per-tier matrix, and on three of the four largest vendors the flagship is the model a European buyer cannot fully purchase. GPT-6 Astra runs in the EU but cannot be escalated to Fast or Ultrafast at any price. Google’s residency tables list no Pro-tier Gemini. Anthropic’s own API has no EU option — Claude in Europe means Amazon Bedrock or Google Vertex. Mistral is the inversion: its flagship is the EU one, with the only published uptime SLA in the set. The surcharge for residency is a known market price — 10% nearly everywhere, 20% on Microsoft’s EU Data Zone — and it is the least important number here.
The matrix: what an EU-resident buyer can actually purchase
Availability now varies within a single vendor’s lineup, which is why a vendor-by-vendor list is the wrong shape. Read this by row.
| Route | EU endpoint | Best model EU-purchasable | Speed / priority tier in EU | Published SLA | EU uplift |
|---|---|---|---|---|---|
| OpenAI direct | eu.api.openai.com | GPT-6 Astra (Standard only) | No — Astra Fast and Ultrafast are US/global only | None for Fast or Ultrafast | +10% |
| OpenAI direct, mid-tier | eu.api.openai.com | GPT-6.1 Sol | Yes — Fast and Ultrafast, from 8 Oct 2026 | None published | +10% |
| OpenAI via Microsoft Foundry | EU Data Zone | GPT-6 Astra (Standard + Provisioned) | No — Priority Processing is Global + US Data Zone | Provisioned deployments only | +20% |
| Anthropic direct | — | None. geo accepts us and global | n/a | n/a | n/a |
| Claude via Amazon Bedrock | 9 EU regions + EU profile | Claude Opus 5.5, Fable 5 | n/a — no speed tier exists | AWS Bedrock service terms | +10% |
| Claude via Microsoft Foundry | No EU Data Zone | None | n/a | n/a | n/a |
| Google Vertex AI | eu multi-region | Flash only — 3.8 / 3.7 / 3.6 / 3.5 Flash, 3.5 Flash-Lite. No Pro tier listed | n/a | n/a | not published |
| xAI direct | None — Global and US only | None self-serve; enterprise contact | n/a | n/a | US endpoint +10% |
| Grok via Amazon Bedrock | EU inference profile | Grok 4.7 | n/a | AWS Bedrock service terms | +10% |
| Mistral | api.eu.mistral.ai | Mistral Large 4 | Priority Tier at 1.75x list | 99.5% uptime SLA | +10% (1.1x) |
| AWS European Sovereign Cloud | eusc-de-east-1 | No frontier model — Nova Lite/Pro, Gemma 4 | n/a | AWS | n/a |
Verified 9 October 2026 against OpenAI’s data-residency guide and Ultrafast mode docs, Anthropic’s Claude in Amazon Bedrock page, Google’s Vertex AI data residency tables, Microsoft’s Foundry deployment types doc, xAI’s regional endpoints page and Mistral’s regional inference docs. Three of these cells moved inside five weeks; the dates are in What changed.
Two rows deserve reading twice.
xAI has no EU endpoint. Its own regional-endpoints page lists exactly two: https://api.x.ai (global) and https://us.api.x.ai/v1, with token usage on the US endpoint costing “10% more than on the global endpoint”. Self-serve EU residency on Grok does not exist; the only European path is Bedrock’s EU inference profile, which is the split this desk documented when Grok reached Bedrock in August. If your shortlist has Grok on it and your requirement is EU processing, the shortlist is shorter than you think.
Anthropic direct has no EU option at all. The geo parameter takes us or global. There is no eu. Claude under EU residency is a cloud-marketplace product — Bedrock in Frankfurt, Zurich, Stockholm, Milan, Spain, Ireland, London or Paris, or Vertex AI’s European regions — and never a first-party purchase. On Microsoft Foundry it is not available in Europe at all, because no EU Data Zone exists for Claude models. That is a meaningful constraint for anyone who picked Claude on capability and discovered the procurement path afterwards.
Residency has a market price. It is 10%, and it is not the decision
Four of the five vendors who publish a number publish the same number.
| Vendor | What carries the uplift | Uplift | Exact wording / scope |
|---|---|---|---|
| OpenAI | Regional processing endpoints | 10% | “charged a 10% uplift for models released on or after March 5, 2026”; FedRAMP endpoints too |
| Mistral | Regional endpoints | 10% | “billed at 1.1× standard list pricing” — input, output, cached reads and cache writes |
| Anthropic on Bedrock | Regional endpoints vs Global | 10% | “Regional endpoints carry a 10% pricing premium over global endpoints” |
| xAI | US endpoint vs Global | 10% | “Token usage on the US endpoint costs 10% more” (no EU endpoint exists) |
| Microsoft Foundry | EU Data Zone vs Global | 20% | US Data Zone at 10%, EU at 20%, per Microsoft’s own GPT-6 Foundry post; changed 1 Sep 2026 |
| Google Vertex | — | not published | Residency is enforced as a model-availability gate, not a line item |
So residency is a commodity with a going rate, and that rate has one documented exception. Microsoft’s Foundry model deployment pricing update took effect 1 September 2026, raising EU Data Zone and non-US Regional deployment prices and launching an APAC Data Zone. Be careful reading it: the announcement and the current rate card describe the EU step differently — the update is reported as a single-digit increase, while Microsoft’s GPT-6 in Foundry announcement (22 September 2026) prices Data Zone deployments at “US at 10% premium and EU at 20% premium to Global rates”.
Worse, you cannot settle it on Azure’s own pricing page. As of 9 October it renders the GPT-6 rows as $- with the notice: “GPT-6 Sol and Luna prices are currently in processing for publishing on this page. Please find the model pricing in the blog.” Azure is the one route where the residency price is not readable from the vendor’s price list. Price it in the Azure portal calculator against your own subscription before committing, and treat 20% as the planning figure.
The arithmetic: one workload, every EU-resident route
Percentages hide the real spread. Here is a single workload priced across every route that can legally serve a European request, denominated in characters rather than tokens so the tokenisers do not distort it: 200 million characters of input and 40 million characters of output per month — roughly a mid-size production assistant.
Converted at each vendor’s own ratio (4.0 characters per token for OpenAI, Google, Mistral and xAI; 2.692 for Anthropic’s current tokeniser, which turns the same text into about 49% more billable tokens):
| Route | Model | Tokens billed (in / out) | Global list | EU-resident monthly |
|---|---|---|---|---|
Mistral api.eu.mistral.ai | Large 4, current preview rate | 50M / 10M | $54.90 | $60.39 |
Google Vertex eu | Gemini 3.8 Flash | 50M / 10M | $75.00 | $75.00 |
Mistral api.eu.mistral.ai | Large 4, list | 50M / 10M | $109.80 | $120.78 |
Mistral api.eu.mistral.ai | Large 4 + Priority Tier | 50M / 10M | $192.15 | $211.37 |
OpenAI eu.api.openai.com | GPT-6.1 Sol | 50M / 10M | $200.00 | $220.00 |
| OpenAI via Foundry EU Data Zone | GPT-6.1 Sol | 50M / 10M | $200.00 | $240.00 |
| Claude via Bedrock EU | Sonnet 5.5 | 74.3M / 14.9M | $297.18 | $326.90 |
| Claude via Bedrock EU | Opus 5.5 | 74.3M / 14.9M | $594.36 | $653.79 |
OpenAI eu.api.openai.com | GPT-6 Astra | 50M / 10M | $1,000.00 | $1,100.00 |
| OpenAI via Foundry EU Data Zone | GPT-6 Astra | 50M / 10M | $1,000.00 | $1,200.00 |
OpenAI eu.api.openai.com | GPT-6.1 Sol Ultrafast | 50M / 10M | $1,200.00 | $1,320.00 |
Rates from our model price tracker, re-verified 9 October 2026. Mistral’s Large 4 preview rate of $0.68 in / $2.09 out is a launch sale with no published end date against a $1.36 / $4.18 list — still in force today, so budget the list row. Mistral’s docs price the regional uplift and the Priority Tier separately and do not say whether they compound; the Priority row assumes they do.
Three things fall out of that table.
The residency surcharge is noise. It moves every row by 10–20%. The spread between the cheapest and most expensive EU-resident option is roughly 22x. What residency costs you is not the uplift — it is the set of models you are allowed to choose from, and on OpenAI the EU-allowed set at the top excludes the speed tiers entirely.
Anthropic’s tokeniser is a bigger line than anyone’s residency fee. Claude Sonnet 5.5 and GPT-6.1 Sol have identical per-token rates — $2 in, $10 out. On the same text, EU-resident Sonnet costs $326.90 against Sol’s $220.00, a 49% gap created entirely by tokenisation. A buyer comparing the two rate cards side by side would conclude they cost the same. Our API cost calculator applies these ratios.
The cheapest credible EU-resident frontier option is Mistral, and it is not close — $120.78 at list against $220 for GPT-6.1 Sol and $653.79 for Claude Opus 5.5. That is the honest version of the European pitch, and the next section is the dishonest version’s correction.
The SLA column, which nobody else prints
A European buyer comparing speed tiers is mostly comparing marketing claims. The exceptions matter.
- OpenAI publishes no latency SLA for Ultrafast on any model, and none for Astra Fast. The “up to 8x faster” figure is a performance claim; a request that cannot be served Ultrafast falls back rather than failing. For EU-resident traffic it is also only available one rung below the flagship.
- Microsoft’s only EU-resident commitment is Provisioned. Per its deployment-types doc, “Provisioned types provide guaranteed throughput and lower latency variance. Standard types offer best-effort service.” So Data Zone Provisioned is the only Azure configuration that buys residency and a throughput guarantee together — and it means reserving PTUs rather than paying per token.
- Mistral’s Priority Tier is the only EU-jurisdiction frontier-adjacent option with a contractual uptime number: 1.75x list, 99.5% uptime SLA, shipped 11 August 2026 alongside the regional endpoints, as this desk covered when Mistral raised its Series D. One caveat worth carrying into the contract: Mistral’s Compute product page advertises 99.9% elsewhere, so get the number you are buying written down.
- Amazon and Google sell platform SLAs, not model latency SLAs. Bedrock’s commitment is AWS’s; there is no per-model latency tier to escalate to.
Set against that, note what OpenAI’s 8 October change actually did for Europe: it made latency a model choice rather than a price choice. You can buy Ultrafast in the EU — on GPT-6.1 Sol and only there. The flagship gap this desk first documented on 6 September is still open.
Residency is not jurisdiction, and the vendors say so
This is the distinction most buyer guides collapse, and it is the one an auditor will not.
Residency means the inference ran in a given geography. Jurisdiction means which country’s authorities can compel the operator. An EU endpoint on a US provider gives you the first and not the second, and the documentation is candid about the seams:
- Mistral: “Regional inference does not provide regional storage for all control-plane data. For example, account configuration, API keys, billing, access management, usage analytics, and other operational metadata may still be handled by Mistral systems outside the selected inference geography.”
- Google: global endpoints “don’t provide any data residency guarantees” — and an SDK that quietly defaults to the global endpoint voids the control without erroring.
- xAI: “A regional endpoint is not, by itself, a comprehensive data-residency guarantee.”
- OpenAI: EU residency is gated on being “approved for abuse monitoring controls” and executing “a Modified Retention amendment” — not a dashboard toggle.
If your requirement is genuine EU jurisdiction, three options exist and each costs capability:
- Mistral — EU-headquartered, EU-hosted, frontier-adjacent, with the only published SLA. The honest limit is agentic coding: Mistral Large 4 scores 28.3% on Terminal-Bench 4.0 against Claude Opus 5.5’s reported 66.4%. If your workload is an autonomous coding agent, this is not a swap — see best AI coding tools.
- Cohere — sovereign and on-premises deployments in region-specific environments, plus the SAP-partnered EU AI Cloud. No public per-token price; Model Vault dedicated inference starts around $4.00/hour (about $2,500/month), so it is a commitment purchase, not a metered one.
- AWS European Sovereign Cloud — the most complete legal separation available and the thinnest model menu. Open since January 2026 from Brandenburg, transitioning to operation “exclusively by EU citizens located in the EU”. On 17 September 2026 Bedrock there gained Gemma 4 — 31B dense, 26B-A4B MoE and E2B — as “the first open weight model family”. Alongside Amazon Nova Lite and Pro, that is the whole frontier-model list: there is none. Global cross-Region inference “isn’t available” in the partition at all.
The trade is explicit, so make it explicitly: you can have the best model, or you can have the strongest jurisdictional guarantee. As of today, nobody sells both.
What a regional endpoint costs you in features
Budget engineering time, not only the surcharge. Every route drops something.
| Route | What you lose |
|---|---|
| Mistral regional | ”Stateful features, including Agents, Batch, and the Files API, are not available on regional endpoints.” Function calling is the only supported regional tool. Regional endpoints “only serve models hosted in that region” — verify per model |
| OpenAI EU | Eligibility approval + Modified Retention amendment; extended prompt caching “may require that OpenAI process and temporarily store Customer Content outside of the Region” |
| Claude on Bedrock | Structured outputs, Files/URL inputs, server-side tools, Message Batches, Agent Skills and MCP connector unavailable (regardless of region); regional endpoints for Claude Fable 5.1 exist in us-east-1 only |
| Microsoft Data Zone | Models arrive Global first, Data Zone second, geography-based last with “no guaranteed availability date”; Priority Processing not offered in the EU zone |
Google Vertex eu | No Pro-tier Gemini in the residency tables; the global endpoint is the default in several SDK paths |
| AWS Sovereign Cloud | No global cross-Region inference, no frontier models, no RAG or fine-tuning, no GPU instance families |
On Mistral specifically, log the endpoint hostname, SDK server value, model ID, timestamp and response identifier per request — its docs ask for exactly that, and it is the only cheap way to prove residency after the fact.
Decision table: pick by constraint
| Your binding constraint | Buy this | Monthly on the reference workload |
|---|---|---|
| EU residency, lowest cost, frontier-class | Mistral Large 4 on api.eu.mistral.ai | $120.78 at list |
| EU residency + a contractual uptime SLA | Mistral Priority Tier | $211.37 |
| EU residency + sub-second interactive latency | GPT-6.1 Sol Ultrafast on eu.api.openai.com | $1,320 |
| EU residency + the strongest coding agent | Claude Opus 5.5 via Bedrock EU | $653.79 |
| EU residency + cheapest acceptable quality | Gemini 3.8 Flash on Vertex eu | $75.00 |
| EU residency + Azure-native billing and policy | Foundry EU Data Zone (Standard or Provisioned) | $240 Sol / $1,200 Astra |
| EU jurisdiction, not just location | Mistral, Cohere sovereign, or AWS European Sovereign Cloud | varies; no frontier model on the last |
| OpenAI’s flagship with a speed tier | Not available in the EU at any price | — |
| Grok, self-serve, EU-resident | Not available — Bedrock EU profile only | $176 US endpoint (not EU) |
Two notes for anyone buying subscriptions rather than tokens: none of this applies to consumer plans. ChatGPT, Claude and Gemini subscriptions carry their own regional terms and no residency choice, and the plan tables on those pages are the right reference. And if your evaluation is really “which assistant”, the Claude vs ChatGPT comparison is the capability axis; this page is only the procurement one.
Before you commit: a five-point check
- Name the model, not the vendor. “Can we use OpenAI in the EU?” has no answer. “Can we use GPT-6 Astra Ultrafast in the EU?” has one, and it is no.
- Check the tier separately from the model. Residency support for Standard tells you nothing about Fast, Priority or Ultrafast — on OpenAI and Microsoft they differ.
- Read the residency price off the vendor’s own page. Four of six publish it. If you are on Azure, you will have to use the portal calculator instead.
- Normalise to characters before comparing. Identical per-token rates are not identical bills; Anthropic’s tokeniser is a 49% line item.
- Separate residency from jurisdiction in writing, and decide which one the requirement actually is before shortlisting. One is a 10% surcharge. The other eliminates every frontier model on the market.
Prices and retirement dates on this page are the ones in our pricing tracker and model retirement tracker, re-verified 9 October 2026.
What changed
- 9 Oct 2026: Page created. Every cell verified against vendor documentation; three cells had moved within the prior five weeks.
- 8 Oct 2026: OpenAI shipped Ultrafast for GPT-6.1 Sol with US and EU data residency, and added EU residency to GPT-6.1 Sol Fast and GPT-6 Luna Fast — the first OpenAI speed tiers purchasable in Europe. GPT-6 Astra’s Fast and Ultrafast remain US/global only. See the analysis.
- 22 Sep 2026: Microsoft brought GPT-6 Astra, Sol and Luna to the Foundry EU Data Zone on Standard and Provisioned deployments, closing the gap reported on 6 September — but Priority Processing stayed Global and US-only.
- 17 Sep 2026: Amazon Bedrock in the AWS European Sovereign Cloud gained Gemma 4 as its first open-weight model family. Still no frontier model in the partition.
- 8 Sep 2026: Mistral’s €3B Series D financed the compute behind its EU endpoints; its Priority Tier (1.75x list, 99.5% SLA) remains the only EU-jurisdiction option with a published uptime commitment.
- 1 Sep 2026: Microsoft’s Foundry deployment pricing update raised EU Data Zone and non-US Regional prices and launched an APAC Data Zone. The current EU Data Zone premium over Global is 20%.
- 6 Sep 2026: GPT-6 Astra confirmed to support neither Fast nor Priority under EU data residency, at a 10% residency premium.
- 11 Aug 2026: Mistral’s regional endpoints reached GA at 1.1x list, with Agents, Batch and the Files API excluded.
Frequently asked questions
Which frontier models can I actually call from an EU-resident endpoint today?
Four routes, and they do not offer the same models. OpenAI's eu.api.openai.com serves GPT-6 Astra, GPT-6.1 Sol, GPT-6 Sol and GPT-6 Luna at Standard processing. Microsoft Foundry's EU Data Zone serves GPT-6 Astra, Sol and Luna on Standard and Provisioned deployments. Amazon Bedrock serves Claude — including Opus 5.5 — through EU inference profiles and nine European regions, which is the only way to get Claude under EU residency, because Anthropic's own API has no EU option at all. Mistral's api.eu.mistral.ai serves its own lineup including Mistral Large 4. Google Vertex AI is the gap: its published data-residency tables cover Gemini Flash and Flash-Lite models only, with no Pro-tier model listed for the EU multi-region. And xAI publishes no EU endpoint of any kind — only Global and a US endpoint at a 10% premium.
How much does EU data residency cost on top of list price?
The going rate is 10%. OpenAI charges 'a 10% uplift for models released on or after March 5, 2026' on regional processing endpoints. Mistral bills regional inference at 1.1x standard list on input, output, cached reads and cache writes. Anthropic's Bedrock documentation puts regional endpoints at 'a 10% pricing premium over global endpoints.' Microsoft is the outlier: its own GPT-6 Foundry announcement prices the US Data Zone at a 10% premium and the EU Data Zone at 20% over Global, following a pricing change effective 1 September 2026. Google publishes no residency uplift line at all. But the uplift is the smallest number in this decision — on one reference workload the spread between EU-resident options is roughly 22x, driven by which model residency lets you buy, not by the surcharge.
Can I buy a fast or priority latency tier with EU data residency?
Only on two routes, and not on any flagship. On OpenAI, GPT-6.1 Sol 'supports US and EU data residency, including with Fast and Ultrafast modes' as of 8 October 2026, and GPT-6 Luna supports Fast in the EU — but GPT-6 Astra's Ultrafast 'supports US data residency and global processing only', and Astra's Fast tier has never been EU-available at any price. On Microsoft Foundry, Priority Processing is available for Sol across Global regions and the US Data Zone, not the EU Data Zone. Mistral's Priority Tier at 1.75x list is the only EU-jurisdiction option with a published uptime SLA attached. Everywhere else, escalating latency means leaving the EU endpoint.
Does an EU endpoint make a US vendor GDPR-safe?
It moves the inference, not the jurisdiction, and the vendors say so. Mistral's own regional-inference docs warn that 'account configuration, API keys, billing, access management, usage analytics, and other operational metadata may still be handled by Mistral systems outside the selected inference geography.' xAI states plainly that 'a regional endpoint is not, by itself, a comprehensive data-residency guarantee.' Google notes that global endpoints 'don't provide any data residency guarantees' — so a misconfigured client that falls back to the global endpoint silently voids the control. If your requirement is EU jurisdiction rather than EU location, the options narrow to Mistral, to Cohere's sovereign deployments, or to the AWS European Sovereign Cloud — where the frontier models are not available.
What do I give up by moving to a regional endpoint besides money?
Features, consistently, and the losses are specific. Mistral: 'Stateful features, including Agents, Batch, and the Files API, are not available on regional endpoints', and function calling is the only supported regional tool. OpenAI: EU residency requires that you 'be approved for abuse monitoring controls, and execute a Modified Retention amendment', and extended prompt caching may process content outside the region. Claude in Amazon Bedrock drops structured outputs, server-side tools, Message Batches and Agent Skills regardless of region, and regional endpoints for Claude Fable 5.1 exist in us-east-1 only. Microsoft ships Global first, Data Zone second and geography-based deployments last, with 'no guaranteed availability date'. Budget engineering time, not just the surcharge.
Is the EU residency premium the same across vendors once you normalise for tokenisers?
No, and this is where the percentages mislead. A 10% uplift applies to a rate denominated in that vendor's own tokens, and the tokenisers are not comparable. Anthropic's current tokeniser runs about 2.69 characters per token against the 4.0 rule of thumb used for OpenAI, Google and Mistral, so the same text becomes roughly 49% more Anthropic tokens. Comparing EU-resident routes per million characters rather than per million tokens is the only apples-to-apples view, and it widens the gap between Claude on Bedrock and Mistral on its own EU endpoint well beyond what the rate cards suggest. Our api cost calculator applies the per-vendor tokeniser ratios used here.