Zero Data Retention for AI APIs: Which Providers Store Your Prompts, and How to Turn It Off
Default retention, training and zero data retention (ZDR) at OpenAI, Anthropic, Google, Azure, Bedrock, xAI and 20 inference hosts, what ZDR doesn't cover, the frontier models that now force retention, and what ZDR costs on OpenRouter.
Zero data retention (ZDR) means the provider processes your request and then keeps nothing: no prompt, no output, no logs of either, once the response is sent. It is not the default at most of the big labs:
- The model makers keep API traffic for about 30 days for abuse monitoring, Google 55 days on the paid Gemini API. OpenAI, Anthropic, Mistral, Cohere, Moonshot and Alibaba offer ZDR only to approved or enterprise customers. xAI, Amazon Bedrock and Google’s Vertex AI let you switch it on yourself.
- Many open-model hosts store nothing by default: DeepInfra, Fireworks, Baseten, Parasail, Cloudflare and Novita among them. A few store by default until you turn it off: Together AI, Nebius, and fal.ai for 90 days.
- Training is a separate question. No major provider trains on paid API traffic by default. The exceptions are free tiers (Google’s free Gemini API, Mistral’s free mode) and DeepSeek, whose policy says inputs may be used for training unless you opt out by email.
The big change this year: the newest frontier models require retention even for ZDR customers. Anthropic’s Claude Fable and Mythos models keep every request for at least 30 days on every platform, and Azure is ending its no-logging option for models released from November 2026.
Everything below was checked on each provider’s own policy pages on 22 November 2026. Policies change often; check the linked page before relying on it.
The model makers and big clouds
| Provider | Default retention of API prompts and outputs | Trains on API data? | Zero data retention |
|---|---|---|---|
| OpenAI | Abuse logs up to 30 days | No (since March 2023) | By approval via sales (ZDR or “modified abuse monitoring”) |
| Anthropic | Deleted within 30 days; flagged content up to 2 years | No | By approval via sales, per organization. Not for Fable/Mythos models |
| Google Gemini API (AI Studio) | Paid: 55 days for abuse monitoring | Free tier: yes, with human review. Paid: no | None. Google points ZDR customers to Vertex AI |
| Google Vertex AI | Cache up to 24h; flagged prompts up to 90 days | No | Self-serve for Google models (turn off caching and storage). Partner frontier models need approval |
| Microsoft Foundry / Azure OpenAI | Flagged content kept for human review in your geography | No | ”Modified abuse monitoring” by application; ending for models released from Nov 1, 2026 |
| Amazon Bedrock | None for most models; OpenAI models 30 days if flagged; Claude Fable all traffic 30 days | No | Self-serve: set data retention mode to none per region |
| xAI | 30 days | No | Self-serve, team-wide switch in the console |
| Mistral | 30 days | Free mode: yes (opt-out). Paid: opt-out available | Pay-as-you-go only, on request, at Mistral’s discretion |
| Cohere | 30 days | Opt-out toggle | Enterprise only, on approval |
| DeepSeek | Not stated; data stored in China | May be used, opt out by email | None found |
| Moonshot (Kimi) | Not stated; servers in Singapore | Docs say no; privacy policy says it may | Enterprise, via sales |
| Z.ai (GLM) | Its API data agreement says content isn’t stored; cache period unannounced | Not stated | No formal program |
| Alibaba Cloud (Qwen) | Up to 30 days | No | Enterprise on the international site, per workspace |
Sources: OpenAI, Anthropic, Gemini API terms, Vertex AI, Azure, Bedrock, xAI, Mistral, Cohere, DeepSeek, Kimi, Z.ai, Alibaba.
Inference hosts and routers
These run open-weight models (and sometimes resell closed ones). Their defaults vary much more:
| Provider | Default retention | Trains on your data? | ZDR |
|---|---|---|---|
| OpenRouter | None, unless you opt in to logging (metadata always kept) | No (opt-in for a 1% discount) | Routing filter you switch on; see below |
| DeepInfra | None (in memory only) | No | Default |
| Fireworks | None for open models | No | Default; account-wide enforcement on Enterprise |
| Baseten | None for synchronous calls; async inputs up to 72h | No | Default |
| Parasail | None for serverless and dedicated; batch about 30 days | No | Default |
| Cloudflare Workers AI | None once no longer needed | No | Default |
| Novita | None per its terms; metadata 6 months | No | Default |
| Cerebras | Says it doesn’t retain inputs or outputs | No | Default, but a thin policy |
| Groq | Not retained, but may log up to 30 days for troubleshooting | No | Opt-in, self-serve (Data Controls) |
| SambaNova | Up to 30 days for abuse monitoring (its data agreement) | No | No setting |
| Together AI | Stored by default | No | Opt-in: an org admin turns off prompt storage |
| Nebius | Stored by default (in Finland) | No | Opt-in, on your profile page |
| Friendli | Model pages say nothing is stored | No | Shown per model; no setting |
| Replicate | API predictions deleted after 1 hour | No clear statement | None |
| fal.ai | 90 days by default; opt out per request with X-Fal-Store-IO: 0 | Not stated | Through sales |
| kie.ai | Media 14 days, logs 2 months | Not stated | None |
| WaveSpeed | Media up to 7 days | No | None |
Two things to watch: a provider’s marketing page and its legal documents sometimes disagree (SambaNova’s site says “zero retention” while its data agreement allows 30 days), and OpenRouter’s own records of providers don’t always match what the providers claim. When it matters, read the privacy policy or data-processing agreement, not the landing page.
What ZDR doesn’t cover
Even with ZDR switched on, some data is still kept:
- Features that need storage are excluded. At OpenAI that means Files, Assistants and threads, vector stores, Batch, fine-tuning and stored conversations. At Anthropic: Batch, Files, code execution and the MCP connector. At xAI: stored responses, Files and Batch. Under Anthropic’s ZDR, using an excluded feature isn’t blocked; it just isn’t covered.
store: falseon a request is not ZDR. It skips one kind of storage, but abuse-monitoring logs can still be kept. Only an account- or organization-level setting guarantees it.- Web grounding leaks. Grounding Gemini with Google Search keeps data 30 days on the Gemini API (up to 3 days on Vertex) and can’t be turned off; Maps grounding keeps 30 days.
- Child-safety screening applies everywhere. OpenAI, OpenRouter and others keep material flagged as CSAM even for ZDR traffic.
- Metadata stays. Token counts, timestamps, model names and billing records are kept at every provider.
Frontier models now force retention
The most capable models of late 2026 are being treated differently:
- Anthropic’s “Covered Models” (Fable 5 to 5.2, Mythos 5 to 5.2) keep prompts and outputs for at least 30 days, with automated review, on Anthropic’s API, Bedrock and Google Cloud. A request from a ZDR organization returns an error; you need a separate workspace with retention enabled.
- Amazon Bedrock requires its review retention mode for Claude Fable, and blocks models that need retention when your account is set to
none. - Vertex AI treats OpenAI’s GPT Astra, Sol, Terra and Luna and Claude Fable and Mythos as “Advanced AI” models with up to 30 days of abuse logging.
- Azure announced on 17 November 2026 that modified abuse monitoring will end for models released on or after 1 November 2026, even for customers already approved (with at least 30 days’ notice).
If your contract or regulator requires ZDR, you may be limited to the previous generation of frontier models, or to open models.
Zero data retention on OpenRouter
OpenRouter doesn’t log prompts unless you opt in, but the provider it routes your request to might. To route only to providers that don’t keep data:
- turn on ZDR in your privacy settings (separately for Anthropic, OpenAI, Google, xAI and all other models), or
- add
"provider": {"zdr": true}to a request, or"data_collection": "deny"to exclude providers that may train on your data.
OpenRouter publishes the list of ZDR endpoints at https://openrouter.ai/api/v1/endpoints/zdr. On 22 November it had 937 endpoints from 61 providers. 53 of them, all Google Vertex endpoints, are marked “safety retention”: they keep flagged content and are only used if you allow that. Since October, Google’s Vertex models need safety retention, so under strict ZDR most Gemini models aren’t available through OpenRouter at all.
What does ZDR cost? For most open models, nothing. We compared the cheapest endpoint for each model with the cheapest ZDR endpoint (blended 3:1 input:output price, 22 November):
| Model | Cheapest endpoint | Cheapest ZDR endpoint | ZDR premium |
|---|---|---|---|
| GLM-5.3 | $0.07 / $1.40 | same (40 of 47 endpoints are ZDR) | none |
| Kimi K3.1 Pro | $0.24 / $8.40 | same (24 of 28) | none |
| gpt-oss-120b | $0.03 / $0.15 | same (22 of 23) | none |
| Qwen3.8 27B | $0.04 / $1.35 | same (16 of 21) | none |
| Llama 3.3 70B | $0.10 / $0.32 | same (7 of 7) | none |
| DeepSeek V4 Pro | $0.66 / $1.32 | $0.90 / $1.50 | +27% |
| Claude Opus 5.5 | $4.00 / $20.00 | same, via Amazon Bedrock or Vertex only | none |
| GPT-6 Sol | $2.00 / $10.00 | same, via Azure only | none |
| Gemini 4 Carbon | $2.00 / $10.00 | none without safety retention | not available |
| Claude Fable 5.2 | $8.00 / $40.00 | none | not available |
Prices per million input / output tokens. Many of the cheapest open-model endpoints run at reduced precision (fp4 or fp8), ZDR or not; each model page lists the providers and their notes.
So on OpenRouter, ZDR mostly costs choice rather than money: OpenAI and Anthropic models can only be served with ZDR through Azure, Bedrock or Vertex, never directly from the lab, and the newest Gemini and Claude Fable models can’t be served with ZDR at all.
How to get zero data retention
- Open models: pick a host that doesn’t store by default (DeepInfra, Fireworks, Baseten, Parasail, Cloudflare), or switch off storage at Together, Nebius or Groq. Or use OpenRouter with ZDR on.
- Claude: apply to Anthropic for ZDR, or use Amazon Bedrock with retention mode
none(self-serve). Not possible for Fable or Mythos. - GPT models: apply to OpenAI or Azure. Azure’s option is closing for new models.
- Gemini: use Vertex AI, turn off caching, don’t store responses and avoid Search grounding. The Gemini API (AI Studio) has no ZDR, and its free tier trains on your data.
- Grok: a team admin can switch ZDR on in the xAI console.
- Keep data in one country: US-only processing costs about 10% more at Anthropic, OpenAI (for newer models) and xAI.
Whichever route you choose, confirm it in writing: most providers show ZDR status in the console or in a response header (xAI sends x-zero-data-retention), and your data-processing agreement is what actually binds them.