OpenRouter Data Residency 2026: EU Routing + ZDR

Q: Does the eu.openrouter.ai endpoint add latency?

No. Your request is decrypted and routed from EU infrastructure, and the upstream provider is also called in-region. End-to-end latency is comparable to the main endpoint for the same model.

Q: Can I pin a specific upstream (e.g. Azure Sweden vs AWS Frankfurt) for the same model?

Indirectly, via provider.order with allow_fallbacks: false. The provider slug for Azure is typically 'azure' and for AWS Bedrock is 'amazon-bedrock'; OpenRouter routes to whichever the order specifies. Direct region pinning within a provider is not exposed and must be configured on the provider side.

On June 22, 2026, OpenRouter published a blog post titled "How to Enforce AI Data Residency Without Building Local Infrastructure". The post packaged an existing toolkit under a new brand name — "Sovereign AI" — and added one new operational primitive: an enterprise-only eu.openrouter.ai base URL that hard-locks decryption and processing to the EU.

This article walks through every layer of the stack, from the per-request provider object that any account can use, to the enterprise EU endpoint that 26 models currently serve. You'll see real code, live model counts (verified June 24), and a side-by-side comparison with the alternatives — Cloudflare AI Gateway, LiteLLM, and Portkey — that compliance teams typically evaluate.

The short version: OpenRouter is the only one of the four that exposes a single managed region URL. Everyone else requires you to wire the region pin yourself at the provider layer.

What "Sovereign AI" actually means

OpenRouter did not ship a single named feature. The June 2026 push bundles three pre-existing primitives under a marketing brand:

Per-request provider object — filter and pin upstream providers in the request body
Zero Data Retention (ZDR) toggle — account-wide, per-model-group, and per-API-key guardrails
eu.openrouter.ai base URL — enterprise-only, hard EU jurisdiction for decryption and processing

The first two work for any account. The third requires an enterprise contract (sales contact, not a SKU). Together they cover GDPR, EU AI Act enforcement, and country-of-origin procurement — but not Chinese PIPL/DSL data localization. There is no China region, no Chinese providers on the EU list, and no documented CN endpoint.

The operational reality: OpenRouter's residency is enforced at the router layer. It decides which upstream provider to call, then calls that provider's existing endpoint. The residency controls therefore work for any provider that OpenRouter routes to and that publishes a regional endpoint — Anthropic, OpenAI, Google, Amazon Bedrock, and 20+ others.

The `provider` object: per-request routing DSL

The most useful control is also the simplest. Any OpenRouter account can add a provider object to a chat completions request to filter and pin upstream routing.

curl https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "anthropic/claude-sonnet-4.6",
    "messages": [{"role": "user", "content": "Summarize this compliance report."}],
    "provider": {
      "order": ["anthropic", "amazon-bedrock"],
      "allow_fallbacks": false,
      "data_collection": "deny",
      "zdr": true
    }
  }'

The four most important fields for residency use cases:

Field	Type	Default	Residency effect
`order`	string[]	–	Try providers in this priority order
`only`	string[]	–	Allow only these provider slugs (hard restriction)
`data_collection`	"allow" / "deny"	"allow"	"deny" excludes providers that store or train on inputs
`zdr`	boolean	–	Restrict to Zero-Data-Retention endpoints only

Two more fields matter for compliance audit trails: allow_fallbacks: false returns an error instead of silently falling through to a non-listed provider (useful when you need a written guarantee that a specific provider served the request), and the response carries a Router Metadata header that records which provider actually handled the call. That audit header is the bread-and-butter evidence you cite in a SOC 2 or ISO 27001 control review.

A common pattern: pin to Bedrock-hosted models in Frankfurt for EU compliance:

import os
import requests

response = requests.post(
    "https://openrouter.ai/api/v1/chat/completions",
    headers={
        "Authorization": f"Bearer {os.environ['OPENROUTER_API_KEY']}",
        "Content-Type": "application/json",
    },
    json={
        "model": "anthropic/claude-sonnet-4.6",
        "messages": [{"role": "user", "content": "Extract vendor names from this contract."}],
        "provider": {
            "order": ["amazon-bedrock"],
            "only": ["amazon-bedrock"],
            "data_collection": "deny",
            "zdr": True,
            "allow_fallbacks": False,
        },
    },
    timeout=30,
)
data = response.json()
# Audit trail: which provider served this request?
provider_used = response.headers.get("x-openrouter-provider")
print(f"Provider: {provider_used}")

The only + allow_fallbacks: false combination is the safest. Without it, a transient outage on Bedrock would silently fall through to a non-EU provider.

ZDR: per-model-group and per-key

The Zero Data Retention toggle is the second layer. It's a per-model-group setting in /settings/privacy:

Anthropic — first-party endpoints removed; Amazon Bedrock and Google Vertex remain
OpenAI — first-party endpoints removed; Azure remains
Google — AI Studio removed; Vertex remains
Non-frontier — applies to providers that don't fit the above buckets

The behavior is consistent: when ZDR is on for a group, OpenRouter filters out any provider in that group that retains data. For a compliance buyer this is straightforward — turn on ZDR for every group that contains a model you might use, and you've removed the worst offenders from the fallback pool.

For team-level enforcement, OpenRouter has guardrails — per-API-key objects that can enforce ZDR by group:

const openRouter = new OpenRouter({
  apiKey: process.env.OPENROUTER_API_KEY!,
});

await openRouter.guardrails.create({
  api_key_id: '',
  enforce_zdr_anthropic: true,
  enforce_zdr_openai: true,
  enforce_zdr_google: true,
  enforce_zdr_other: true,
});

These OR with the per-request flags — you can only strengthen, not weaken. That's the right semantic: a developer can't accidentally route a compliance workload to a non-ZDR provider if the key is locked.

The `eu.openrouter.ai` base URL: hard EU jurisdiction

This is the new piece. Enterprise accounts can switch the base URL from https://openrouter.ai/api/v1 to https://eu.openrouter.ai/api/v1, and the request is decrypted and processed entirely within the EU before being forwarded to an EU-region upstream provider.

The integration is a one-line change in the SDK:

import { OpenRouter } from '@openrouter/sdk';

const openRouter = new OpenRouter({
  apiKey: '',
  serverURL: 'https://eu.openrouter.ai/api/v1',  // <-- the only change
});

const completion = await openRouter.chat.send({
  model: 'meta-llama/llama-3.3-70b-instruct',
  messages: [{ role: 'user', content: 'Hello' }],
  stream: false,
});

Or with raw fetch:

const res = await fetch('https://eu.openrouter.ai/api/v1/chat/completions', {
  method: 'POST',
  headers: {
    'Authorization': 'Bearer ',
    'Content-Type': 'application/json',
  },
  body: JSON.stringify({
    model: 'anthropic/claude-sonnet-4.6',
    messages: [{ role: 'user', content: 'GDPR audit summary' }],
  }),
});

To see what models the EU endpoint serves, query its model list directly:

curl https://eu.openrouter.ai/api/v1/models | jq '.data | length'
# 26 (as of 2026-06-24)

The 26 models break down roughly as: 8 Anthropic (Claude Opus 4.5–4.8, Sonnet 4–4.6, Haiku 4.5), 11 OpenAI (GPT-4o-mini, GPT-4.1 family, GPT-5 family, GPT-5.5, GPT-OSS 20B/120B), 3 Google (Gemini 2.5 Pro/Flash/Flash-Lite), and 4 Amazon (Nova Micro/Lite/Pro and Nova 2 Lite). The 26 are the subset of OpenRouter's 338-model main catalog that are routed through EU-region upstream providers.

One nuance worth flagging: the OpenRouter blog frames Mistral as an "EU-jurisdiction anchor" you can pin via provider.only: ["mistralai"]. That's true for the main endpoint — Mistral is in the main catalog with 20+ models. But Mistral is not on the eu.openrouter.ai endpoint list. If you need a hard EU-jurisdiction guarantee, Mistral alone is not enough; you need an EU-region upstream for the model you're using (e.g. Anthropic via Bedrock Frankfurt, or OpenAI via Azure Sweden).

Pricing: 0% markup, 5.5% platform fee

No residency surcharge appears in public materials. The pricing model is the same as standard OpenRouter:

0% markup on underlying provider prices
5.5% platform fee on credit purchases, $0.80 minimum
5% with Bring Your Own Key (BYOK)
First 1M requests/month waived (on BYOK)
Failed requests not billed
20+ free models for evaluation

The EU in-region endpoint itself is gated behind enterprise sales contact (https://openrouter.ai/enterprise/form) rather than a SKU. For a serious compliance buyer, expect an enterprise contract, not pay-as-you-go credits.

For comparison: a typical EU-pinned workload at 1M requests/month on BYOK is roughly 5% on top of upstream provider list price. The "no markup + 5% platform fee" framing is competitive against going direct — direct Anthropic, OpenAI, and Google all charge standard list price, but you'd need to wire three separate accounts, three separate SSO integrations, and three separate vendor-risk reviews.

How the alternatives stack up

The four approaches compliance teams typically evaluate:

Approach	Residency mechanism	How it's configured	Pricing	Caveats
OpenRouter `provider` object	Filter providers by data policy / ZDR	`provider.only`, `provider.zdr`, `provider.data_collection` in body	Included in 5.5% platform fee	"Region" = provider-headquartered; not hard jurisdiction
OpenRouter `eu.openrouter.ai`	Hard EU jurisdiction	Switch base URL to `https://eu.openrouter.ai/api/v1`	Enterprise, contact sales	EU only; 26 models; Mistral not on the EU list
Cloudflare AI Gateway	No native region-routing	Caching, rate limits, retry/fallback only	Free tier + Workers paid plan	No residency primitive in the gateway itself
LiteLLM (self-hosted)	Your network, your jurisdiction	`model_list` with `openrouter/...` upstream	Free OSS; infra ~$200–$500/mo	No SOC 2 / ISO 27001 / HIPAA certs on the OSS build
Portkey	Control-plane over your provider keys	BYOK + guardrails + PII redaction	$49/mo Production; Enterprise for HIPAA	Acquired by Palo Alto Networks in 2026; HIPAA + BAAs on Enterprise
Direct provider region-pinned endpoints	Provider's own regional endpoints	`region: eu` on Vertex, Azure EU, AWS Bedrock in `eu-central-1`	Provider list price	Each provider defines "region" differently

The key takeaway: OpenRouter is the only one of the four managed options that exposes a single managed EU-only base URL. Cloudflare AI Gateway (the infrastructure under OpenRouter itself) does not expose residency as a first-class control — caching, rate limiting, retry/fallback only, as of the last doc update on April 20, 2026. LiteLLM gives you full sovereignty because you host it, but you own the SOC 2 / ISO 27001 / HIPAA certification. Portkey adds HIPAA + BAAs on its Enterprise plan but no managed EU base URL.

If you need a BAA for healthcare, the answer is Portkey Enterprise plus your own provider BAAs (Azure OpenAI, AWS Bedrock with a BAA). If you need a managed EU endpoint and don't need HIPAA, OpenRouter is the cleanest answer.

Compliance use cases, mapped

Regime	How OpenRouter helps	What's missing
GDPR (Articles 44–50)	`data_collection: "deny"` + `zdr: true` for Schrems-II-style data minimization; `eu.openrouter.ai` for hard EU jurisdiction	Article 30 records of processing — must be at the controller level, not the API
EU AI Act (2026 enforcement)	Same primitives; Sovereign AI doc calls it out by name	Risk classification documentation must be at the deployer level
Deloitte "country of origin" procurement	77% of 3,235 surveyed leaders now factor country of origin into vendor selection	Vendor risk questionnaires still need to be filled out per-provider
Financial services (PCI-DSS, FINRA, OCC)	ZDR per-model-group toggle, guardrails per API key for segregation of duties, Router Metadata header for audit trail	SOC 2 Type 2 report at `https://trust.openrouter.ai` is sufficient for most, but not all, third-party risk teams
Healthcare (HIPAA — US-side)	Partial: ZDR + `data_collection: "deny"` + EU/US-only base URL	OpenRouter is not HIPAA-certified; a BAA must be at the provider layer (Azure OpenAI, AWS Bedrock with a BAA)
China data localization (PIPL, DSL)	No coverage	No China region, no Chinese providers on the EU list, no documented CN endpoint. This is a real gap.
Defense / air-gapped	Not the right answer	"The infrastructure framing fits governments, defense contractors, and air-gapped environments" — for those, self-hosted LiteLLM is the only option.

The China gap is worth repeating. If you're serving a Chinese user base and need PIPL compliance, OpenRouter is not the answer. You'd need a China-region upstream (Alibaba Qwen, DeepSeek, Zhipu GLM, Moonshot Kimi, etc.) configured at the provider layer, plus your own data localization controls. That's outside the scope of what OpenRouter offers today.

A 30-line implementation recipe

Putting it all together for a typical EU compliance workload:

# requirements: requests, python-dotenv
import os
import requests
from dotenv import load_dotenv

load_dotenv()

# 1. Use the EU endpoint if your contract covers it
# 2. Pin to a single EU-region upstream
# 3. Enforce ZDR + data_collection: deny
# 4. Disable fallbacks so a transient outage surfaces as an error
# 5. Capture the router metadata for audit logs

def compliance_chat(prompt: str, model: str = "anthropic/claude-sonnet-4.6") -> dict:
    base = os.environ.get("OPENROUTER_BASE", "https://openrouter.ai/api/v1")
    # Switch to EU endpoint for hard jurisdiction (enterprise only)
    if os.environ.get("EU_RESIDENCY") == "true":
        base = "https://eu.openrouter.ai/api/v1"

    response = requests.post(
        f"{base}/chat/completions",
        headers={
            "Authorization": f"Bearer {os.environ['OPENROUTER_API_KEY']}",
            "Content-Type": "application/json",
        },
        json={
            "model": model,
            "messages": [{"role": "user", "content": prompt}],
            "provider": {
                "order": ["amazon-bedrock"],
                "only": ["amazon-bedrock"],
                "data_collection": "deny",
                "zdr": True,
                "allow_fallbacks": False,
            },
        },
        timeout=30,
    )
    response.raise_for_status()

    # Audit trail
    return {
        "provider": response.headers.get("x-openrouter-provider"),
        "content": response.json()["choices"][0]["message"]["content"],
    }


if __name__ == "__main__":
    result = compliance_chat("Summarize this vendor risk report in 3 bullets.")
    print(f"Served by: {result['provider']}")
    print(result["content"])

Two patterns worth keeping in mind: the EU_RESIDENCY=true env-var pattern lets you flip between the main and EU endpoints without code changes, and the per-request provider object ORs with the account-wide ZDR setting — both have to be on for the strictest guarantee.

FAQ

Q: Does the eu.openrouter.ai endpoint add latency?

A: The endpoint itself does not. Your request is decrypted and routed from EU infrastructure, and the upstream provider is also called in-region. End-to-end latency is comparable to the main endpoint for the same model.

Q: Can I pin a specific upstream (e.g. Azure Sweden vs AWS Frankfurt) for the same model?

A: Indirectly, via provider.order with allow_fallbacks: false. The provider slug for Azure is typically azure and for AWS Bedrock is amazon-bedrock; OpenRouter routes to whichever the order specifies. Direct region pinning within a provider (e.g. forcing AWS Frankfurt over AWS Ireland) is not exposed — that has to be configured on the provider side.

Q: Is there a free tier for the EU endpoint?

A: The EU endpoint is enterprise-gated. There is no public free SKU; you need to contact sales for access.

Q: Does ZDR cover fine-tuning data?

A: No. ZDR covers inference only. Fine-tuning is a separate workflow that uses provider-specific endpoints and is not currently exposed through OpenRouter's residency controls.

Q: How does OpenRouter handle the new EU AI Act requirements?

A: The Sovereign AI doc explicitly calls out EU AI Act compliance. The combination of data_collection: "deny", ZDR, and the eu.openrouter.ai endpoint is positioned as the operational primitive. Risk classification and conformity assessments remain the deployer's responsibility under Article 17.

Q: What's the difference between ZDR and data_collection: "deny"?

A: ZDR means the provider does not retain the prompt or response at rest. data_collection: "deny" means the provider does not use the data to train future models. A provider can have one without the other — for example, a provider might retain logs for 30 days (violating ZDR) but never train on inputs (satisfying data_collection: "deny"). For the strictest guarantee, set both.

Verdict

OpenRouter's June 2026 push is best understood as: a blog post (data-residency framing) + a new enterprise-only EU endpoint that hardens an existing toolkit. The toolkit is mature — the provider object, ZDR toggles, and guardrails have been there for months. What changed is that you can now point at a single managed URL and tell your compliance team "all inference happens in the EU."

If you're building an EU product and don't need a BAA, OpenRouter is the cleanest managed option today. If you need HIPAA, you need Portkey Enterprise plus provider BAAs. If you need PIPL data localization, you need a China-region provider stack — OpenRouter does not cover it. If you need full sovereignty, self-hosted LiteLLM is the only path.

The one thing not to do: assume that the OpenRouter brand name "Sovereign AI" alone is a compliance artifact. It's a marketing label. The actual controls are the provider object, the ZDR toggles, and the eu.openrouter.ai endpoint — and you need to wire all three correctly to get the guarantee you're promising your auditor.

If you ship to production and want to consolidate multi-provider traffic (OpenAI, Anthropic, Google) through one OpenAI-compatible endpoint with built-in ZDR and provider routing, FreeModel is a China-direct aggregator that mirrors the OpenRouter pattern with DeepSeek, Qwen, GLM, and Kimi under the hood. It runs a similar provider object shape on its enterprise tier and is the closest domestic equivalent for teams that need China access alongside EU/US routing.