Data Retention

When you send a message, Pentos passes it to the AI provider that runs the model you picked. The Data Retention row in the model selector’s details shows what that provider keeps of your request after it answers, for chat and image models.

Your chats and files are saved in Pentos, encrypted, so you can come back to them. For how long Pentos keeps deleted chats, see Download, Archive & Delete. Pentos never uses your data to train AI models. See Security.

What Each Value Means

ValueProvidersWhat the provider keeps
NoneAnthropic, SpaceXAI, MistralNothing, under Zero Data Retention agreements
Not loggedModels hosted by CloudflareNo logs, without a formal Zero Data Retention commitment
Up to 30 daysOpenAIYour request, for up to 30 days, only to check for abuse
Up to 55 daysGoogleYour request, for up to 55 days, only to check for abuse
Your API keyAny model running on your own keyWhatever your own account’s terms with that provider allow
VariesSmart RouterDepends on the model it picks for each message

How Requests Reach Providers

Every chat, image, read aloud, and dictation request passes through Cloudflare AI Gateway on its way to the provider. We’ve set it not to keep your requests:

  • No request logs: log collection and log classification are off.
  • No cached responses: response caching is off, so every answer comes straight from the provider.

None: Zero Data Retention

Zero Data Retention (ZDR) is an agreement with an AI provider that it won’t store your requests or its responses after it answers. Pentos has ZDR agreements with Anthropic, xAI, and Mistral:

  • Anthropic: Claude Opus 5.5, Claude Sonnet 5.5, and Claude Haiku 5.5.
  • xAI: Grok 4.7, Grok Imagine 2.0, and the voices that read messages aloud.
  • Mistral: Mistral Large 4, Mistral Medium 3.5, Mistral Small 4, and Ministral 3 14B.

Claude Fable 5.1 runs only on your own API key, so it shows Your API key. See Using Your Own API Key.

Not Logged: Models Hosted by Cloudflare

GPT OSS 120B and 20B, Gemma 4, Kimi K2.7 Code, Kimi K2.6, GLM 5.3, GLM 5.3 Flash, GLM 4.7 Flash, DeepSeek V4 Pro, DeepSeek V4 Flash, Llama 4 Scout, Nemotron 3 120B, Granite 4.0 Micro, FLUX.2 Dev, and FLUX.1 Schnell run on Cloudflare’s own servers, and so does Whisper, which turns your dictation into text. Your request never goes to the company that made the model.

  • Not logged: our AI Gateway doesn’t log these requests. See How Requests Reach Providers.
  • No training: Cloudflare’s terms say it doesn’t use your requests or responses to train AI models or to improve its services, and doesn’t share them with other Cloudflare customers.
  • Not labeled ZDR: Cloudflare doesn’t offer a formal ZDR commitment for the models it hosts, so we don’t call it ZDR.

Up to 30 or 55 Days: OpenAI and Google

GPT 6.1 Sol, GPT 6 Luna, GPT 5.3 Codex, the Gemini models, and the GPT Image and Nano Banana image models are sent to OpenAI and Google.

  • Kept only for abuse checks: OpenAI keeps requests for up to 30 days and Google for up to 55 days, only to check for misuse of their services, then deletes them.
  • Logging is off: we’ve turned off request logging in our OpenAI and Google accounts, so neither keeps your requests for us to review.
  • No training: neither uses these requests to train its models.

Using Your Own API Key

When a message runs on your own API key, it goes to the provider under your own account, so that account’s data terms apply instead of Pentos’s agreements. The model selector shows Your API key in that case. The same applies to images made with your key, and the image model selector shows Your API key too. Claude Fable 5.1 and GPT 6 Astra always run on your own key.

If your provider account has its own ZDR agreement, it covers these messages.

Smart Router

The Smart Router shows Varies, because it picks a model for each message and can switch to a backup model from another provider if the first one fails. To keep a message with a provider that keeps nothing, pick a model that shows None yourself. A model you pick never switches providers.

When Web Search or Connectors are on, the AI can use tools run by our partner Composio, which handles web searches and the calls to your connected apps.

  • What Composio keeps: a record of each tool call, including what was sent and what came back, for up to one year under its standard policy.
  • What you can see: every tool a response used. Open the response’s Message Info and click View next to Connectors.