OpenRouter ZDR vs No-Training vs In-Region Routing

01

Overview

OpenRouter exposes several privacy and routing controls that sound similar but solve different problems. The three most important are Zero Data Retention, data-collection restrictions and in-region routing.

Using the wrong one can create a false sense of privacy. A no-training setting does not necessarily mean zero retention, and Zero Data Retention does not mean the request stays inside your country or inside your own infrastructure.

02

Quick comparison

ZDR: controls whether an eligible inference provider retains prompts and responses after processing.

data_collection: "deny": excludes endpoints that store user data non-transiently and may train on it.

US/EU in-region routing: controls where eligible inference traffic is decrypted and processed.

These can be combined because they address separate policy dimensions.

03

What ZDR actually protects

OpenRouter defines Zero Data Retention as a provider policy under which the provider processes a request and returns the response without persisting the prompt or response afterward.

That is stronger than a simple “not used for training” promise on the retention dimension.

However, the request still reaches the provider. ZDR does not make OpenRouter an on-prem system and does not automatically govern your application's own logs, tools, caches or analytics.

04

What no-training/data-collection controls protect

A provider can retain data for reasons such as abuse monitoring or legal requirements while still promising not to use it for model training.

That is why OpenRouter exposes data_collection separately from ZDR.

When data_collection is set to deny, OpenRouter says routing excludes endpoints that store user data non-transiently and may train on it. This is useful when organizational policy is broader than prompt-retention alone.

05

What in-region routing protects

Data residency is a geographic requirement.

OpenRouter's US in-region endpoint is designed so requests are decrypted inside the United States and routed only to US provider endpoints. Its EU endpoint applies the same idea to EU processing.

This does not automatically mean ZDR. A provider can process data in the required region while still having a retention policy you do not want. Region and retention must therefore be evaluated separately.

06

Which control should you use?

Use ZDR when the requirement is: “The inference provider must not retain prompts or responses after processing.”

Use data_collection deny when the requirement is broader: “Do not route to endpoints that collect/store user data in ways that may include training.”

Use in-region routing when the requirement is: “Processing must occur in the US or EU.”

Use all three when the workload needs all three guarantees simultaneously.

07

Example policy combinations

Internal prototype with ordinary business data

You may not need strict routing controls, but should still review provider policies and avoid logging sensitive information unnecessarily.

Customer-support agent with confidential tickets

ZDR plus data_collection: "deny" can be appropriate if company policy prohibits retention and training use at the provider layer.

EU customer data with residency requirements

Use the EU base URL together with ZDR and data-collection controls if both geography and retention matter.

Highly regulated workload

Gateway controls may still be insufficient. If policy requires customer-controlled infrastructure or prohibits sending data to external inference providers altogether, self-hosted or contractually dedicated infrastructure may be necessary.

08

Why model name alone is not enough

OpenRouter routes models across provider endpoints, and endpoint policies can differ.

A privacy assessment should therefore ask: “Which endpoint will serve this request under these routing controls?” rather than only “Which model am I using?”

This is one reason OpenRouter's policy-aware routing is useful. The gateway can enforce constraints as provider availability changes, instead of forcing the application to hard-code one vendor forever.

09

The hidden layer: your own application

None of these controls automatically fixes data handling in your stack.

A request can be routed to a zero-retention provider and still be stored by:

• server logs;
• APM or observability tools;
• agent traces;
• connected tools;
• plugin databases;
• browser telemetry;
• caches;
• support systems.

A real privacy architecture maps the full data path from the user's input to every system that sees it.

10

Is ZDR always more private than no-training?

On provider retention specifically, ZDR is the stricter concept because it says the prompt and response are not persisted after inference.

But privacy is multidimensional. A ZDR endpoint in the wrong geographic region may violate residency policy, while an in-region endpoint with weak retention rules may violate storage policy.

There is no single switch that replaces the policy matrix.

11

Does in-region routing guarantee compliance?

No. Geography is one compliance input, not a complete compliance program.

OpenRouter's routing features can help enforce technical constraints, but organizations still need to evaluate contracts, endpoint policy, application logging, access control, security posture and legal requirements relevant to their workload.

12

Bottom line

OpenRouter ZDR, no-training/data-collection controls and in-region routing should be thought of as three separate axes:

• retention;
• collection/training;
• geography.

The strongest configuration is not automatically the right one for every workload. Start with the actual policy requirement, then enforce only the controls needed to satisfy it without silently weakening them when routing becomes constrained.