Data handling
What Echo Token Station records and does not record, how generated video files are handled, what is passed to upstream APIs, the upstream and processing region of each model, and what happens during outages.
This page describes how data that passes through Echo Token Station is handled.
Summary
| Item | Details |
|---|---|
| Request and answer text | The text of successful requests and the model's answers are not stored on Echo's servers. The exceptions are described in "Recording of error text" and "Video generation" below |
| Use for training | Echo does not use the data you send to train AI models |
| Upstream and processing region | For each model, the API it goes through and the region where it is processed are shown on the model page |
| Switching during outages | If an upstream has an outage, Echo does not switch to an upstream in another country or region |
| What we record | The items needed to calculate usage and prices (table below) |
What Echo records and does not record
| Recorded | Not recorded |
|---|---|
| Time, account, key name, model ID, input and output token counts (including cached tokens), amount, processing time, whether the call was streamed, the upstream used, request ID, the path called | Your input such as text and images, the model's answer |
- Some of the recorded items are shown in 「利用状況」 (Usage) and 「呼び出し履歴」 (Call history) in the console
- Source IP address: not kept in the usage records (the items in the table above). It is kept in the relay server's access log (one line per request with the time, request ID, HTTP status, processing time, source IP address and the path called)
- Retention period: not decided yet. As of 2026-09-28, the usage records and the access log are not deleted automatically
- Location of Echo's relay server: Japan (Tokyo)
Recording of error text
When an upstream API returns an error, the error text sometimes contains part of the input you sent. In that case, Echo records that error text. Retention period: not decided yet (as of 2026-09-28, it is not deleted automatically).
Video generation
- Generated video files are not stored on Echo's servers. Each time you download a file, Echo fetches it from the provider (the company that actually runs the model) and passes it to you as is
- You can download a file only while the provider keeps it: 24 hours after generation finishes for Seedance 2.0; the period is not published for MiniMax-H3. Once that period has passed, the file cannot be downloaded through Echo either (Video generation)
- For tasks (jobs you submit and whose result you collect after generation finishes), Echo records the task ID, state, reason for failure, the amounts used to calculate the price (length, resolution, token count, number of reference images and so on), the provider's task ID and the provider's file URL, in addition to the items in the table above
- The prompt and the images, videos and audio used as references are not recorded
Information passed to the upstream
Echo passes your request to the upstream API (the API of the provider that actually runs the model) and receives the answer. The following information is also passed to the upstream.
| Information | Handling |
|---|---|
| The request body | Passed to the upstream API so that the model can produce the answer |
The value of metadata (Anthropic format) | Passed on unchanged. Claude Code puts IDs that identify the device and the session here |
The value of store (Responses API) | Passed on unchanged. In the OpenAI API it defaults to true, and the upstream stores the response. If you do not want it stored, specify store: false |
| URLs of images, videos and audio used as references in video generation | Passed to the provider's API. The provider fetches the files at those URLs |
How each upstream provider handles data (whether it is used for training, how many days it is kept) follows that provider's terms. Links to each provider's terms: [ ]
Upstream APIs and processing regions
Models actually run on each provider's API (the upstream). Each model's page shows the following two items.
| Shown | Example |
|---|---|
| The API it goes through | Anthropic API, AWS Bedrock (Tokyo region) |
| Processing region | One of the 4 in the table below |
| Processing region | Meaning |
|---|---|
| Japan | Processed in Japan |
| Overseas | Processed in a country or region other than Japan and China. The specific country is written on the model page |
| China | Processed in China |
| Global | The processing region is not fixed. It may be processed in any country or region |
The same applies to video generation models: the processing region is shown on each model's page.
If a model has more than one upstream, only upstreams in the same processing region are used. When an upstream returns an error, the request may be sent again to another upstream, but that upstream is also in the same processing region.
Parameters in Anthropic-format requests that set the processing region and similar options (inference_geo, service_tier, speed) are removed by Echo before the request is passed to the upstream. The processing region is the region fixed for each model.
Switching during outages
If an upstream API has an outage, Echo does not automatically switch to an upstream in another country or region. Requests to that model return an error. The processing region never changes without your knowledge.
Going through a tool
Some tools, such as Cursor, send requests to Echo through the tool company's servers. In that case, the tool's terms also apply to what you send.
Personal information
Personal information in your account is handled under the Echo AI Hub privacy policy (https://echohub.io/legal#privacy ).
Sources
| What | Source | Checked |
|---|---|---|
| The recorded items, the text of successful requests is not stored, error text is recorded | Checked in Echo's test environment: we called the API with a marker string in the input and searched the whole database and log files | 2026-09-27 |
| The access log keeps the source IP address; there is no setting that deletes records automatically | Measured in Echo's test environment (same relay server software and settings as production) | 2026-09-28 |
| Location of the relay server | Echo relay server settings (as of 2026-09-28) | 2026-09-28 |
| Video generation not storing generated files, what is recorded for tasks | Checked in Echo's test environment: submitted tasks with a marker string in the prompt and searched the whole database and log files (with a server that returns test responses in place of the provider's API) | 2026-10-02 |
| How long providers keep files | https://www.volcengine.com/docs/82379/1521309 , https://platform.minimax.io/docs/api-reference/video-generation-v2-query | 2026-10-01 |
| Upstreams of the same model are limited to one processing region; no switching to another country or region | Echo relay server settings, and a measurement in the test environment (with upstreams in Japan and the US, a model assigned only to the Japanese upstream used only the Japanese upstream, also when retried) | 2026-09-27 |
metadata and store are passed on unchanged; inference_geo and the others are removed | Measured in Echo's test environment | 2026-09-27 |
store defaults to true in the OpenAI Responses API | https://developers.openai.com/api/reference/resources/responses/methods/create | 2026-09-27 |
| Cursor's requests go through Cursor's servers | https://cursor.com/docs/settings/api-keys | 2026-09-26 |
| Where the Echo AI Hub privacy policy is | https://echohub.io/legal | 2026-09-28 |