Privacy and data
This page describes what the service does with your data, as implemented. The legal privacy policy and the data processing addendum are separate documents.
What is stored
Section titled “What is stored”| Data | When | For how long |
|---|---|---|
| Call metadata: time, requested and served model, vendor, token counts, cost, routing decision, outcome, latency, key and organization ids | always | call records 24 months; routing decisions 180 days; the billing ledger and audit events for as long as the account exists and for the period the law requires afterwards |
The app that sent a call, when your software names it: the origin of the URL in the HTTP-Referer header (for example https://myapp.com, never the path or anything after it) and the name in the X-OpenRouter-Title or X-Title header. Zabio does not publish it; an app is listed publicly only if its owner asks |
only when your software sends these headers | with the call records, 24 months |
Prompts and completions of /v1/chat/completions and /v1/embeddings |
only when the API key has body logging on (off by default) | the organization’s retention: 30 days by default, adjustable from 1 to 365 |
Stored conversations (/v1/responses, /v1/conversations) |
when you use them | the organization’s retention, or the shorter or longer retention you set on the conversation; the clock restarts on every turn |
A sealed copy of the request of a response created with background: true (the same content the conversation already holds) |
while the response is queued or running | until the turn ends, then dropped; deleting the response drops it too |
| Routing session state: line, the job (task type, difficulty, stakes, subject, language), model, confidence, a few counters and a truncated copy of the last user message | for calls that pass a session id or use a conversation | 1 hour after the last call, 24 hours at most |
| A fingerprint of a long prompt: a keyed hash of its system prompt, tools and messages, never the text, so the prompt is written to the vendor’s cache only when it comes back | prompts of about 1,024 tokens or more that do not ask for caching themselves | until the vendor’s cache would have expired: 5 minutes or 1 hour after the prompt was last seen |
A template fingerprint of each auto request: a keyed hash of your organization, the system prompt, the tool names, the response format and whether images are attached, never the text, plus whether the request had none of these and whether it carried an earlier answer. It measures how often the same template comes back; routing and billing never read it |
auto requests |
with the routing decision: 180 days |
| Prompt-injection guard events: for a request in which the guard found something, where it was (message, field, character offsets), the rule, the action taken, the key and the call, and a keyed hash of each detected span; never the text. The patterns run on Zabio’s own servers; no text is sent anywhere to be checked | when the guard finds something (it is on by default) | 90 days |
| The guard’s allowlist: the rule and the keyed hash of each phrase your organization reported as not an injection or typed in the AI Firewall; never the text. An entry may carry a note your owner or admin writes; Zabio stores the note as written. | when an owner or admin adds an entry | until it is removed |
Batch files: the input you upload to /v1/files, the results file and the error file of a batch (/v1/batches) |
when you use batches | the organization’s retention period, counted from the upload for inputs and from the end of the batch for results, or sooner if you delete them or set an expiry on upload; never while a batch is using them |
Background image generations (zabio.background: true on /v1/images/generations): the request, prompt included, sealed while it is queued or running; then the answer with its images, as a file |
when you use them | the request until the generation ends, then removed; the answer for the organization’s retention period, or sooner with the request’s zabio.expires_after; deleted at once with DELETE /v1/images/generations/{id} |
Images returned as links (response_format: url on /v1/images/generations and /v1/images/edits): each image made, as a file. Anyone holding the link can open the image until it expires |
when you ask for links | 60 minutes from when the image is made, then removed; deleted at once with DELETE /v1/files/{id} |
Video generations (POST /v1/videos): the request, prompt and reference images included, sealed while the video is made |
when you generate a video | until the video ends, then removed; deleted with the organization |
| Generated videos: each video made, as a file. Anyone holding its link can open the video for 60 minutes from the read that made the link, without an API key | when you generate a video | the organization’s retention period, or sooner with the request’s zabio.expires_after; deleted at once with DELETE /v1/videos/{id} or DELETE /v1/files/{id} |
| Your own vendor keys | when you add them | until you remove them |
| Connected AI apps (MCP): the app’s name, its client id (the address of its metadata document, or the id Zabio gave it when it registered), the host it returns to, the organization and project, what it may do, its daily limit, and when it was connected, last used, expires and was disconnected. Tokens are kept only as a one-way hash | when you connect an app | while the organization, the project and your account exist |
Feedback sent with the feedback tool: the call’s id, good or bad, and your reason (500 characters at most) |
when an app sends it | while the organization exists |
| Shared calls: a de-identified copy of the request and the answer of a share of your calls, with the routing decision and the cost of each attempt | when your organization’s share setting is above 0 (100 by default during the beta) | 90 days, or sooner if you delete them |
| Account data: email, name, language, password hash, sessions, organization memberships, the IP address the account was created from | while the account exists | deleted within 30 days after you delete the account; an account closed for fraud or abuse keeps its records for 2 years and then only a scrambled fingerprint of its email, so the same address cannot sign up again |
| Sign-ins: the time, how you signed in (password, email link, Google, GitHub) and whether the second step was used | every sign-in | 90 days, then deleted; sooner when you delete your account |
| Server logs: request path, status code, timing, request id and the client IP address; never prompts, completions, keys or passwords. Security events are part of them (sign-up, login, sign-in link and password reset attempts with their outcome, unknown paths, rate-limit refusals, Turnstile failures, welcome credits refused by the abuse caps), read by a CrowdSec engine on our own server that can block an abusive address at Cloudflare on the web app’s auth pages, never on the website or the API | always | 30 days |
A connected app’s ask calls run on the connection’s API key like any call: metadata as above, never the request or answer text, and the task goes to the vendor that answers.
Prompts and completions are sent to the vendor the call is routed to, under that vendor’s API terms. On auto requests, the routing classifier also receives an excerpt of the request so it can name the job: the first 1,500 characters of your system prompt, the last user message, the names of your tools, the number of attached images, the input size and a short digest of earlier turns; it never receives images, tool results, earlier turns in full or any output. The classifier is TypeSafe AI’s Jev (United States; TypeSafe does not train on API requests). When Jev cannot answer, the same excerpt goes to Google’s Gemini Decision (United States), and then to Anthropic’s Haiku (United States) if Gemini cannot answer either. Neither is asked for a request whose model access leaves its provider out. Every response says which one decided in zabio.routing.classifier. The one sentence of reasoning the classifier stores with each routing decision is capped at 200 characters. When you name a Jev model, TypeSafe AI receives the whole request, system prompt and every message, and returns the answer, like any vendor whose model you name. The same happens when auto answers a classification job with Jev (a request whose JSON schema is made only of fixed choices, booleans and bounded scores, see Classification jobs on auto): the whole request goes to TypeSafe AI, and zabio.model_served says so; naming a text model keeps every request away from TypeSafe except the classifier’s excerpt. When Zabio splits a big auto request (see Split and route), TypeSafe AI also receives the excerpt of your source that the cheap model copied, up to 40 percent of it, so Jev can check that it holds what your question needs; the extractor’s vendor (OpenAI or Google) receives the whole message, as the vendor of a routed call does. Zabio does not use your content to train anything.
If your organization turns on the firewall’s paid tier, or when you send text, images or documents to the check endpoint, the tool results in your requests, with their images and documents, each with the first 2,000 characters of the user message that led to it, and your users’ new messages, with the images they mark as not their own and their documents, or what you send there, are also sent to TypeSafe AI in the United States for security screening; Zabio keeps none of that content, only a hash of each screened text or image and its result for 24 hours.
With each call on Zabio’s vendor keys, the vendor also receives an opaque identifier in the field it documents for one (OpenAI, Anthropic, Google, DeepSeek, DeepInfra and Moonshot AI today, and xAI on images): a hash of your organization’s id and, when you send safety_identifier or user, of your end user’s id. It is never your organization’s id, your end user’s value, a name or an email; it lets a vendor act on one end user instead of on Zabio’s account. On your own vendor key, the vendor receives the value you sent, as you sent it. See What the vendor receives.
| Vendor | Where it processes requests |
|---|---|
| Anthropic, OpenAI and Google (all their models), xAI (chat models) | United States |
xAI image and video models (grok-imagine-image, grok-imagine-image-2.0, grok-imagine-video-1.5, grok-imagine-video-1.5-lite) |
not guaranteed: xAI serves images and video only on its global endpoint, which may route requests between its regions. The prompt, the images you send to edit and the reference images of a video go to xAI for the call and are not stored by Zabio beyond the sealed request of a video generation described above |
TypeSafe AI (requests that name a Jev model, and classification jobs auto answers with Jev: the whole request; the routing classifier on auto requests: prompt excerpts only, never full conversations or outputs; on a split auto request, also the excerpt of the source copied for the check; for organizations that turn on the firewall’s paid tier: tool results and new user messages, with their images and documents, for security screening, or when you send text, images or documents to the check endpoint, what you send there) |
United States |
Together AI (open-weight models and the DeepSeek models with ids ending in @together, shown as “processed in North America”) |
North America |
| Mistral AI | not guaranteed: Mistral AI does not commit to an inference location on its global API |
Moonshot AI (kimi-k3.1-pro, kimi-k3.1-pro-highspeed and kimi-k3@moonshot, served by Moonshot AI’s international API) |
not guaranteed: Moonshot AI commits to no inference location; its privacy policy says its servers are in Singapore. Its terms allow it to use inputs and outputs to improve its services unless agreed otherwise in writing; Zabio has no such agreement, so turn the vendor off if that matters to you |
DeepInfra (the models with ids ending in @deepinfra, copies of open-weight models served by DeepInfra at a stated precision) |
not guaranteed: DeepInfra commits to no inference location; its privacy policy says data is processed in the United States. Its terms say it does not use inputs or outputs to train, fine-tune or improve any model and does not retain, store or log them beyond the request |
Parasail (the models with ids ending in @parasail, copies of open-weight models served by Parasail at a stated precision) |
not guaranteed: Parasail commits to no inference location; its privacy policy says data may be stored and processed in the United States, the European Union or any other country where it or its providers have facilities. Its terms say it does not use inputs or outputs to train its models or improve its platform and retains them only for the time it takes to answer |
ElevenLabs (scribe-v2 and scribe-v2-medical, transcription; eleven-v4, eleven-v4-turbo, eleven-v3, eleven-v3-conversational, eleven-multilingual-v2, eleven-flash-v2.5 and eleven-flash-v2, speech) |
United States: ElevenLabs hosts customer data in the U.S. as a standard. The audio to transcribe and the text to speak are sent for the call and not stored by Zabio. Its terms allow it to use inputs and outputs to improve its services and models unless the account opts out in its Data use settings; Zabio’s account opts out, and turn the vendor off if that matters to you |
Deepgram (nova-3, nova-3-multilingual and whisper-large, transcription; aura-2, aura-1 and flux-tts, speech) |
not guaranteed: Deepgram’s default endpoint commits to no location (its regional endpoints are in the EU, Australia and India). The audio to transcribe and the text to speak are sent for the call and not stored by Zabio. Its terms allow training on inputs unless a request opts out; Zabio sends the opt-out on every request |
| DeepSeek, direct API (“DeepSeek (China)”) | China |
Alibaba Cloud (US): the models with ids ending in @alibaba-us and the video models wan2.6-t2v and wan2.6-i2v, served by Alibaba Cloud Model Studio’s Virginia workspace with the US deployment scope |
United States (Virginia): data stored and processed only in the United States. Alibaba Cloud does not use your inputs or outputs to develop or improve its models unless you separately consent (Alibaba Cloud International Website Product Terms of Service, section 4.48, Alibaba Cloud Model Studio). Alibaba’s data retention policy for Model Studio states that prompts and responses are “retained for no longer than 30 days, unless longer retention is required by applicable laws or is reasonably necessary to ensure the security” of the service |
Alibaba Cloud (International): the Qwen models with plain ids and the copies with ids ending in @alibaba-intl, served by Alibaba Cloud Model Studio’s Singapore workspace |
Data stored in Singapore and processed anywhere outside mainland China. The same terms apply: no use of your inputs or outputs to develop or improve its models without your separate consent, and prompts and responses retained for no longer than 30 days as stated above |
Batch files. When you use the batch API, we store the file you upload, the results file and the error file, encrypted, for your organization’s retention period, counted from the upload for inputs and from the end of the batch for results. You can delete a file at any time, or set a shorter expiry when you upload it. We do not keep a copy of batch requests or results outside those files, except the working copies a running batch needs on each line (a result until it is written to the results file, and the shortened request of a line Smart Splitting handled), which are encrypted the same way and deleted when the line ends; and we delete the copies held by the model vendor as soon as we have collected the results. Each line of a batch is sent to the model vendor it is routed to, under that vendor’s API terms, like a normal request.
Every vendor is enabled by default. Owners and admins can turn any vendor off for the organization on the Models page, and the account, a project or an API key can be limited to some models; a vendor that is off never receives your content.
All storage is in the United States. Processing is in the United States too, except for requests routed to Together AI (North America), to DeepSeek’s own API (China), to Mistral AI (no guaranteed region), to Moonshot AI (no guaranteed region), to DeepInfra (no guaranteed region), to Parasail (no guaranteed region), to Deepgram (no guaranteed region) or to Alibaba Cloud (International) (stored in Singapore, processed anywhere outside mainland China), as listed above; Alibaba Cloud (US) stores and processes only in the United States. No vendor in the catalog processes only in the European Union yet, so an organization that requires EU processing cannot use any model today. Customers anywhere may use the service; content sent from the EU crosses to the US. The transfer basis is described in the data processing addendum.
How stored bodies are protected
Section titled “How stored bodies are protected”- Encrypted with AES-256-GCM under a data key unique to your organization. The data key is wrapped by a master key that lives only in the server’s configuration, with restricted permissions, and can be rotated without re-encrypting your data.
- Each stored item is bound to its own record, so a payload cannot be moved or replayed as another item.
- The web app never displays conversation bodies. They are readable only through the API with a key that has the
conversationsscope. Logged bodies of stateless calls are readable by owners and admins through an audited endpoint. - Logs, audit events, metrics and error messages never contain prompts or completions.
- Your own vendor keys are encrypted too, each under its own data key wrapped by the master key; only the last four characters are ever shown.
How data is deleted
Section titled “How data is deleted”- On request.
DELETE /v1/conversations/{id}andDELETE /v1/responses/{id}are immediate and permanent. Owners and admins can delete conversations from the web app. Vendor keys are erased when removed. - By retention. A nightly job removes expired conversations, logged bodies and batch files, routing decisions older than 180 days, and delivered emails older than 30 days.
- Batch files.
DELETE /v1/files/{id}is immediate and permanent, except for a file a running batch still needs, which is deleted when the batch ends and you ask again. The vendor’s copies of a batch’s inputs and results are deleted as soon as the results are collected. - On deleting an organization. An owner deletes it from Settings. Its API keys stop at once; its conversations, logged bodies, files and vendor keys are deleted at once; its members and invitations are removed. Its data key is destroyed 30 days later, after which any remaining ciphertext is unreadable, and everything else about it is deleted at the same point, except the billing records (ledger, payments, the billing details on receipts), which we keep for the period accounting law requires.
- On deleting your account. You delete it from Settings, with your password or right after signing in again, once you no longer own an organization that has other members or a balance; the organizations you own alone are deleted with it. Your account is hidden at once and everything about it is deleted within 30 days (name, email, password, Google identity, tokens, sessions, memberships, the address you signed up from), except billing records and, for an account closed for abuse, the evidence we keep for 2 years.
- Your data, downloaded. Settings has two downloads: your account data (profile, organizations, token and session records, your recent sign-ins, referral names) and, for owners and admins, the organization’s data (members, invitations, projects, key records without the keys, the IP allow and deny lists, the ledger, the audit log, billing details). Both are JSON files; the usage page exports call records as CSV.
Call data export
Section titled “Call data export”If your organization sets up a call data export, Zabio sends a record of each call (model, vendor, tokens, cost, timing, status, the routing line, the firewall’s tag and charge, and the ids of the key, project, session, conversation and response; every field is listed on the Call data export page) to the destination your organization configures. Prompts and answers are sent only when you switch on Include bodies and the call’s key has body logging on. Once sent, that data is held by the destination under its own terms. See Call data export.
Web search
Section titled “Web search”When a model’s own provider runs a web search, the search and its results stay with that provider, and its terms apply. Zabio stores how many searches a call ran and what they cost. The search queries and the cited addresses are kept only where the answer itself is kept (a stored conversation, body logging), and never the text of a result.
What is never stored
Section titled “What is never stored”- Payment card numbers. Payments run on Stripe; Zabio keeps your Stripe customer id only.
- API keys in clear. Zabio stores a hash of each key and its first 8 characters.
- Prompts or completions for calls without body logging, outside stored conversations and outside the share of calls your organization shares, de-identified.
- Videos you send into a chat request (
video_urlparts on/v1/chat/completions,input_videoparts on/v1/responses). They go to Google or Alibaba, whichever answers, for the call and are dropped when it ends; Zabio writes them to no database, log or file. - Spoken answers from chat models (
modalitieswith audio on/v1/chat/completions). The audio goes from the model’s vendor to you, and Zabio writes it to no database, log or file and never shares the call; with body logging on, only the answer’s transcript, its audio id and its size are kept. The vendor (today OpenAI) keeps the audio on its side until the answer’sexpires_at, so a later turn can refer to it by id.
Shared calls. Your organization shares a percentage of its calls with Zabio, 100 percent by default during the beta, to improve how Zabio routes requests; the owner sets it from 0 to 100 in Settings, and a project or an API key can set its own. A shared call is a de-identified copy of the request and the answer: the organization, project, key, user and IP identifiers are removed and emails, phone numbers, keys, card numbers and IP addresses are scrubbed before it is stored, in a separate store on RemarkableCloud’s object storage in the United States, for 90 days. De-identified is not anonymous: a prompt can still contain personal data the scrubber did not recognize. Shared calls are read only by Zabio’s operators and by a judge model. The judge sees a shared call to score the answer or to ask another model for its own answer to the same request, and it reaches only vendors whose published terms say they do not train on API inputs and outputs: Alibaba Cloud (International), Alibaba Cloud (US), Anthropic, Deepgram (with its per-request training opt-out on every call), DeepInfra, ElevenLabs (with its Data use setting switched off on Zabio’s account), Google (paid services), Mistral AI (with model improvement switched off on Zabio’s workspace), OpenAI, Parasail, Together AI, TypeSafe AI and xAI (under its enterprise terms). DeepSeek and Moonshot AI receive no shared call. A call sent with store: false, the header X-Zabio-Share: no or zabio.share: false in the body is never shared. Sharing earns up to 5 percent off Zabio’s fee (see Billing). The owner can delete what the organization shared from Settings; the store is rewritten without it within the hour.
Guest and free-plan chats. Guest and free-plan chats may be kept for up to 90 days, with personal details removed, and used to improve how Zabio picks models. They are never sold or used to train models. Paid usage is kept only as your organization’s sharing setting allows; you can turn it off in Settings.
Emails
Section titled “Emails”Zabio sends transactional emails only: verification, sign in links, password reset, invitations, payment confirmations and low balance warnings. The text of a sent email is deleted 24 hours after delivery; the record that it was sent (recipient, kind, subject, time) is kept for 30 days for support and then deleted.