Processed in memory
The request body is read in memory to pick a channel, count tokens and calculate the price, and is released when the response finishes.
What zero data retention covers
Zero data retention (ZDR) means the content of a request, the prompt you send and the output the model returns, is not kept once the request completes. At TokenRa this applies to synchronous API requests for every model in the catalog, including chat completions, image generation and typed decisions.
Never logged
Production logging records request metadata only. Request and response bodies are not written to the database, the application logs or disk.
Not used for training
TokenRa is an API gateway, not a model developer. We do not train, fine-tune or evaluate models on your prompts or outputs.
What TokenRa does not store
For synchronous requests, none of the following is written to storage at any point.
- Prompts and conversation history: system prompts, user and assistant messages, and tool or function definitions.
- Attachments sent with a request: images, audio and files, whether sent inline or as URLs.
- Model outputs: generated text, reasoning content, tool calls and generated images.
- Streamed responses: tokens are relayed to your client as they arrive and are not collected for storage.
- Response caches: TokenRa keeps no gateway-side cache of prompts or responses. When a model supports prompt caching, that cache lives at the provider.
What TokenRa keeps: request metadata
To bill your usage, show it in your dashboard and stop abuse, TokenRa records metadata about each request. None of these fields contains prompt or output text.
| Field | Why we keep it |
|---|---|
| Time, account and API key name | Billing, usage history and per-key limits. |
| Model and endpoint | Pricing and usage reports. |
| Input, output and cached token counts, and the charge | Billing and your usage dashboard. |
| Latency and streaming flag | Performance monitoring. |
| Request ID and upstream request ID | Support and resolving disputes with providers. |
| Upstream channel ID | Routing diagnostics. |
| Client IP address | Security and abuse prevention. |
Retention follows the Privacy Policy: request metadata and operational logs are typically kept for no more than 12 months, then deleted or anonymized. Account and payment records are covered there as well.
Two exceptions
Two kinds of records can contain request content. Both are described here so you can decide what to send.
- Asynchronous video jobs. Video generation, such as Seedance, runs as a job that you poll. To return the result, TokenRa stores a task record: the task ID, status, model and the provider's job response, which can include the parameters you submitted, such as the prompt, and a link to the generated video. Task records stay in your task history; email support@tokenra.io to have them deleted.
- Provider error messages. When a provider rejects a request, TokenRa writes its error message, truncated to 2 KB, to the operational logs for troubleshooting. Some providers quote part of the request in their errors. These logs follow the same retention as request metadata.
Retention at the model provider
TokenRa forwards each request to the provider that serves the model you call, and most models in the catalog are served through the model developer's official API. What happens to your data on the provider's side is governed by that provider's policy, not by TokenRa. Providers can keep API data for a limited time, for example for abuse monitoring, and their terms on training use differ, so review the policy for the models you use.
| Provider | Models on TokenRa | Provider policy |
|---|---|---|
| DeepSeek | DeepSeek V4.1 Flash, DeepSeek V4.1 Flash 1017 | DeepSeek Privacy Policy |
| MiniMax | MiniMax M3.1 Flash, Space Bunny Alpha | MiniMax API Platform privacy policy |
| Moonshot AI | Kimi K3, Kimi K2.8 Preview | Kimi Open Platform privacy policy |
| Zhipu AI (Z.ai) | GLM-5.3-Flash, GLM-5.3-FlashX, Ox Alpha | Z.ai privacy policy |
| Alibaba Cloud | Qwen3.8-Omni-Flash | Alibaba Cloud privacy policy |
| ByteDance (BytePlus) | Seedance 2.0, Seedance 2.5 | BytePlus privacy policy |
| OpenAI | GPT Image 2.5 | OpenAI platform data controls |
| TypeSafe AI | Jev | TypeSafe privacy policy |
| Not disclosed | Union Alpha, Omen Alpha | No policy published |
Links point to each developer's international API policy. The upstream for a model can change; email support@tokenra.io to confirm the current upstream for a specific model.
Anonymous preview models
Union Alpha and Omen Alpha are offered by providers that have not disclosed their identity or data policy. Assume their prompts and outputs may be retained, and do not send personal or confidential data to them.
Recommendations for sensitive workloads
- Use synchronous endpoints for content you do not want stored anywhere at TokenRa.
- Leave personal data, credentials and secrets out of prompts unless your use case needs them.
- Choose models whose provider policy meets your requirements, and keep confidential data away from anonymous preview models.
- Keep API keys on your server and rotate any key that may have been exposed.
- Email support@tokenra.io to delete your task history or to ask how a specific model handles data.
Zero data retention FAQ
Does TokenRa store my prompts or model outputs?
Not for synchronous requests. The prompts and outputs of chat completion, image generation and typed decision requests are processed in memory and are not written to TokenRa's database, logs or disk. The two exceptions are asynchronous video jobs, whose task records can include the submitted prompt, and provider error messages, which TokenRa logs truncated to 2 KB.
Does TokenRa use my data to train AI models?
No. TokenRa is an API gateway and does not train, fine-tune or evaluate models on API data. Whether a model provider may use API data is set by that provider's own terms.
What does TokenRa keep about my API requests?
Request metadata only: the time, account and API key name, model and endpoint, token counts and charge, latency, request IDs, upstream channel ID and client IP address. It is used for billing, your usage dashboard, support and abuse prevention, and is typically kept for no more than 12 months.
Is TokenRa zero data retention end to end?
TokenRa's commitment covers the TokenRa gateway. Once a request is forwarded, the model provider's own data policy applies, and providers can keep API data for a limited time, for example for abuse monitoring. Most models on TokenRa are served through the model developer's official API, and the provider table links to each policy.
How long does TokenRa keep logs?
Request metadata and operational logs are typically kept for no more than 12 months and then deleted or anonymized. Longer retention applies only where the law, a dispute or a security investigation requires it, as described in the Privacy Policy.
How do I delete my data?
Email support@tokenra.io from the email address on your account. We can delete your task history and account data, except billing records that the law requires us to keep. Your data rights are described in the Privacy Policy.
This page describes how TokenRa handles API data as of October 22, 2026, and supplements the Privacy Policy, which governs if the two differ. Provider names are trademarks of their respective owners. TokenRa is an independent API gateway and is not affiliated with the providers listed.