# Data policy

https://modelportal.ai/docs/data-policy

A request passes through three parties: the platform itself, the content moderation service, and the provider that serves it. Where each processes data, and whether it keeps it, is set out separately below.

## Data processing regions of the three parties

| Party | What it does | Where to find its processing region |
| --- | --- | --- |
| The platform | Authentication, billing, routing, relaying requests and responses | The [privacy policy](/legal/privacy) states where the platform is deployed |
| Content moderation | Risk review of requests | The [privacy policy](/legal/privacy) names the moderation provider and its processing region |
| Provider | Runs the model and generates the answer | The data processing region in each endpoint's data policy, shown in the endpoint list on the model page and in `GET /api/v1/models/<model ID>` |

`data_regions` and `exclude_regions` constrain only the provider's processing region. They do not mean every hop of the request stays in that region: the platform's and the moderation service's regions are not affected by these parameters.

## Endpoint data policy declarations

Every endpoint declares whether it retains request data, for how many days, whether data is used for training, its data processing region, and a link to the provider's privacy policy. Declarations are public on the model page.

## data_collection

Put `"data_collection": "deny"` in `provider` to use only endpoints that explicitly declare no retention and no training. Missing or unclear declarations count as not meeting it. If no endpoint qualifies, you get 422 `no_eligible_endpoint` with a message naming `data_collection`.

```json
{"provider": {"data_collection": "deny", "data_regions": ["US", "SG"]}}
```

`data_regions` keeps only endpoints whose processing region is listed and `exclude_regions` drops the listed regions, both as two-letter uppercase country or region codes. An endpoint that declares no processing region does not meet either parameter.

## What the platform forwards

- Only protocol-required headers are forwarded (such as `Content-Type`, `anthropic-version` and `anthropic-beta`); your IP, User-Agent, cookies and platform key are not.
- Body fields that could link a request to you (`user`, `metadata.user_id`, `safety_identifier`, `prompt_cache_key`) are removed before forwarding.
- The platform does not store request or response content, except for risk review. See [Privacy](/docs/privacy).