Your prompt may pass through more than one service before a model answers it. An AI model aggregator receives your request, routes it to a model provider, and returns the result; the aggregator’s settings, the provider’s policies, and any app or tools in the workflow can each affect what is logged, retained, or used for model improvement. OpenRouter is a documented example of these layers, not a guarantee that every aggregator works the same way.
How a prompt travels through an AI aggregator
A typical request follows this path: you → aggregator → model provider → aggregator or app response path. The aggregator needs to receive the request to fulfill it. It then sends the prompt to the provider serving the selected model. Depending on the service and configuration, you may choose a provider or the aggregator may route automatically. The selected model and route therefore affect which provider receives the content.
As an Amazon Associate I earn from qualifying purchases.
The prompt and generated response are content, but they are not the only data involved. A service may also process account details, uploaded files, and request metadata. OpenRouter, for example, says it collects metadata such as token counts and latency; that is distinct from the text or other content of a prompt and response. See its privacy documentation.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Who may keep or use the data?
The aggregator
Do not assume that a statement about prompt logging describes every kind of storage. OpenRouter documents private input/output logging as an opt-in setting, off by default, that makes prompts and completions visible in logs. Its terms also describe temporary processing-related storage for certain features, including batch or large-volume requests that cannot be handled in memory. Storage can depend on the feature, settings, and applicable agreement.
#1 Best Overall
OpenRouter also says it samples a small number of prompts for categorization used in reporting and model rankings. Its documentation says that if a user has not opted into OpenRouter’s use of inputs and outputs, categorization is stored anonymously and is not associated with an account or user ID; the categorization model is described as having a zero-data-retention policy. Separately, OpenRouter stores request metadata such as token counts and latency. These are distinct practices, not one blanket rule about whether prompts are logged.
The underlying model provider
The provider that processes the request has its own data practices. OpenRouter’s privacy policy states: “Different Model Providers have different data practices, including with respect to whether they retain or use your Inputs and Outputs to train, fine-tune, evaluate, or improve their Models.” Some providers may retain inputs or outputs or use them for training or improvement. An aggregator’s opt-out does not necessarily control a provider’s independent handling; check the terms for the specific provider and any organization-specific agreement.
Your app and other services
The application calling the aggregator can create its own copies in databases, analytics, error tracking, or application logs. External search tools, plugins, and other enabled services may also receive request data under separate terms. OpenRouter’s ZDR explainer treats these as layers outside its provider-routing control. Provider-side ZDR does not remove copies held elsewhere: “Provider-side ZDR doesn’t delete any of those copies.”
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Fix the driver behind crashes, sound loss and screen glitches3Clear out junk files and repair common Windows errorsWhat no-training, ZDR, and routing controls mean
Privacy labels can refer to different protections. Treat them as separate questions rather than a single “private” setting.
- No training: whether inputs or outputs may be used to train, evaluate, or otherwise improve a model.
- Zero data retention (ZDR): whether an inference provider persistently stores prompts after responding.
- Data residency or region pinning: where processing occurs.
- Application logging and external tools: other copies or recipients, which are not automatically covered by provider-side ZDR.
- Caching: a separate behavior to check. OpenRouter says provider-side in-memory prompt caching can be compatible with ZDR, while its response-caching feature temporarily stores generated responses.
OpenRouter documents data_collection: deny as a routing filter that excludes providers classified as collecting user data, and zdr: true as a request-level constraint for endpoints designated as Zero Data Retention. The first filters routes according to a data-collection classification; the second constrains eligible inference endpoints. OpenRouter cautions that its provider-policy tags are not definitive third-party policy statements: “This is not a definitive source of third party data policies, but represents our best knowledge.” Check the current provider-routing documentation and the underlying provider’s own terms.
ZDR is not a universal promise that nothing is stored anywhere. It concerns inference-provider routing and does not by itself cover your application’s logs, external tools, plugins, or processing region. No-training and ZDR address different issues, so both may need to be enforced. Nor does a provider-side ZDR setting answer whether response caching or other service features retain data.
How images, audio, video, and files differ
Multimodal inputs can follow the same provider route as text, but persistent file storage is a separate feature. OpenRouter’s privacy policy says image, audio, and video inputs are sent to the applicable model provider and are not persisted by OpenRouter beyond the time needed to route them, except for abuse detection, security, billing, or legal compliance. By contrast, files uploaded through a separate Files API or persistent file-storage feature are retained until the user deletes them or closes the account, subject to stated exceptions. Files submitted with inference requests are sent to the selected provider under that provider’s terms. Check the current privacy policy for the feature you use, and distinguish an attachment sent with one inference request from a file kept in persistent storage.
Recommended Free Tools
What to check before sending sensitive information
- Identify the exact aggregator feature and the downstream model provider used for the request, including whether routing is automatic.
- Read the aggregator’s current privacy and terms documentation, then read the downstream provider’s data terms.
- Check prompt and response logging, file storage, retention, training, and request metadata separately.
- If available, configure no-training and ZDR controls independently. Verify which providers or endpoints each setting allows.
- Inspect the calling application’s logs and analytics, plus any external tools, plugins, caching, and region settings.
- For organizational use, review the applicable data-processing agreement and admin configuration rather than assuming consumer defaults apply.
- Do not submit secrets or personal data unless the full data path and protections are acceptable for your use case.
These checks help identify where data may go; they are not a legal conclusion or a guarantee about another aggregator. Policies, routes, and settings can change, so confirm the current terms for the actual service, provider, and feature.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
