Zero Data Retention
Zero data retention (ZDR) means that once a request finishes, no prompt or output content is kept anywhere in the system.
1. Moyi API retains no content of its own
As a gateway, Moyi API does not retain your prompts or model outputs anywhere along the request path. Once a request completes, the content is released from memory — never written to disk, never stored in a database, and never used to train any model.
For billing, troubleshooting, and uptime reporting we do keep content-free call metadata: timestamp, model, provider and channel, token counts, cost, HTTP status, time to first token and generation time, plus request and session IDs. None of these fields contain prompt text, model output, or any fragment from which a conversation could be reconstructed.
| Data | Retained | Notes |
|---|---|---|
| Prompt text | No | Target state; currently still retained |
| Model output text | No | Already not retained |
| Upstream error bodies | No | Target state: keep only the error code and error type |
| Call metadata (timestamp / model / token counts / cost / latency) | Yes | Required for billing and uptime reporting; contains no content |
Nothing needs to be configured on your side — this is the default behaviour for every request.
2. Providers that support ZDR
Once Moyi API forwards a request upstream, how the data is handled is governed by that provider's own policy. The table below lists every provider we currently run in production along with a link to their official data policy. Their own terms are authoritative.
| Provider | ZDR | Official policy |
|---|---|---|
| Anthropic | Yes | API and data retention |
| AWS Bedrock | Yes | Data protection |
| Google Vertex | Yes | Vertex AI ZDR |
| Azure OpenAI | Yes | Data privacy |
| OpenAI | Yes | Your data |
| Google AI | No | Gemini API additional terms |
| DeepSeek | No | Privacy policy |
| Z.AI (Zhipu) | No | Privacy |
| Moonshot | No | Privacy policy |
| xAI | No | Privacy policy |
| Volcengine | No | Data security notes |
| Tencent Cloud | No | Privacy policy |
| Baidu AI Cloud | No | Privacy policy |
| OpenRouter | Yes | Privacy and logging |
| Third-party providers | No | No uniform ZDR commitment; requests with a ZDR requirement should not be routed to this tier |
If a provider's ZDR position cannot be confirmed, or we have no agreement in place with them, we mark them "No". So "No" in the table covers both providers that have explicitly ruled ZDR out and providers we have not yet confirmed. Some providers marked "Yes" require you to request ZDR from them separately.
If you need ZDR and want to pick providers yourself, contact sales.
Known exception: Claude Fable 5
Anthropic classifies claude-fable-5 (and Anthropic's Mythos 5) as a Covered Model with mandatory 30-day data retention and no ZDR on any provider — including Anthropic direct, Google Vertex, and Amazon Bedrock. Anthropic's stated reason is that some abuse patterns are only detectable across multiple accumulated requests; this data is not used to train Claude.
So if you have a ZDR requirement, do not use claude-fable-5.