What does zero data retention mean?

Zero data retention, usually shortened to ZDR, means the provider does not store your inputs or outputs at rest once the API response has been returned. Anthropic words it as not storing "customer prompts or responses at rest after the API response is returned". It is a contractual arrangement applied to your account, not a flag you set on a single request.

Does zero data retention cover document uploads?

Usually not, and this is the detail that catches document teams. ZDR is written for stateless inference. Anthropic marks the Files API "No" for ZDR because files are "retained until explicitly deleted", and Mistral names "/v1/files" directly in its exclusion list. If your pipeline uploads a PDF before reading it, the upload sits outside the arrangement even though the model call sits inside it.

Does OpenAI train on API data?

No, not by default. OpenAI states that data sent to the API "is not used to train or improve OpenAI models (unless you explicitly opt in)", a default in force since March 1, 2023. Training and retention are separate questions though: without ZDR, abuse monitoring logs containing prompts and responses are still retained for up to 30 days.

Does Anthropic train on API data?

Not without permission. Anthropic states that "retained data is never used for model training without your express permission" and that conversation content "is not retained by default". The exception it names is Covered Models, currently Claude Fable 5 and Claude Mythos 5, which require 30-day retention.

How do I get a zero data retention agreement?

You ask for it, and on two of the three vendors a human has to approve it. Anthropic says to "contact the Anthropic sales team" and enables ZDR per organization, so a second organization under the same account does not inherit it. Mistral gates ZDR to the Scale plan. OpenAI applies ZDR at the organization or project level for approved customers.

Can I use batch processing with zero data retention?

On Anthropic and Mistral, no. Anthropic marks batch processing "No" for ZDR with "29-day retention; async storage required", and Mistral names batch processing files in its exclusion list. That matters commercially because batch is where the 50 percent discount lives, so the cheapest way to run a large document job is also the way that steps outside the arrangement.

Are the newest models available under zero data retention?

Not all of them. Anthropic designates Claude Fable 5 and Claude Mythos 5 as Covered Models that "require 30-day data retention; ZDR is therefore not available for either model", and a ZDR organization calling one gets a 400 error. You can enable 30-day retention on a single workspace to unlock them while the rest of the organization stays at zero.

Is zero data retention the same as HIPAA compliance?

No, and on Anthropic they are alternatives rather than additions. Anthropic states that HIPAA-ready API access "is designed as an alternative to ZDR for organizations handling PHI" and that if you handle PHI, "HIPAA readiness is the arrangement to use; you do not also need ZDR". HIPAA readiness keeps data with safeguards around it; ZDR removes the data.

Does zero data retention mean literally nothing is kept?

No. Every vendor keeps a legal and safety carve-out. Anthropic says that even with ZDR in place, flagged inputs and outputs may be retained "for up to 2 years". OpenAI says image and file inputs scanned for CSAM may be retained "even if Zero Data Retention" is enabled. Zero is the default state, not an absolute guarantee.

// Verified from vendor docs, July 23, 2026

Zero Data Retention OCR: What ZDR Actually Covers on Claude, OpenAI and Mistral

A zero data retention arrangement covers the stateless inference call. Document processing is the one workload that is not stateless. Here is what each vendor's ZDR includes, what it names in the exclusion list, and why the file upload is the step that breaks it.

What ZDR means in each contract
The upload step that sits outside it
Why batch costs you the arrangement
The models ZDR cannot run

See the comparison Extract a document free

Upload a document, no signup

PDF, JPG, PNG, BMP, HEIC, TIFF

Upload a document to extract

Extract documents →

Use a redacted sample if the real document is sensitive.

SOC 2 Type II

256-bit encryption

US data handling

Fields, not just text

3 of 3: vendors exclude stateful file handling
29 days: Anthropic batch retention, outside ZDR
2 years: the flagged-content ceiling under ZDR
0: frontier models ZDR can run at Anthropic

// The short answer

What zero data retention means, and where it stops

Zero data retention means the provider does not store your inputs or outputs at rest once the API response has been returned. Anthropic's wording is that under ZDR it "does not store customer prompts or responses at rest after the API response is returned". It is an account-level arrangement, not a per-request flag, and on Anthropic and OpenAI a human has to approve it. Mistral gates it to the Scale plan.

The part that matters for document work is the scope. ZDR is written for stateless inference: you send a prompt, you get an answer, nothing is kept. A document pipeline usually is not shaped that way. It uploads a file, references it, sometimes queues it in a batch, and each of those steps is a different endpoint with its own policy. Anthropic marks the Files API "No" for ZDR because files are "retained until explicitly deleted". Mistral names "/v1/files" in its exclusion list in the same sentence as agents, batch and libraries. So a team can be entirely correct that its vendor offers ZDR and entirely wrong that its document pipeline runs under it.

Last updated July 2026. Every quote on this page is from the vendor's own current documentation, linked in each section below.

// The eight questions that decide it

Zero data retention on Claude, Mistral and OpenAI, question by question

Rows four and five are the ones document teams get wrong. All three vendors say yes to the model call and no to the stateful handling around it.

Question	Anthropic (Claude)	Mistral	OpenAI
Is ZDR offered at all?	Yes, by arrangement	Yes, on the Scale plan	Yes, for approved customers
Can you self-serve it?	No, sales enables it	Plan gated, then admin panel	No, approval required
Is the OCR or vision call covered?	Yes, inline PDFs and images	Yes, /v1/ocr is on the list	Yes, on the covered endpoints
Is uploading a file covered?	No, Files API is excluded	No, /v1/files is excluded	Stateful endpoints excluded
Is batch processing covered?	No, 29-day retention	No, named in exclusions	Not on stateful endpoints
Are all models available?	No, Fable 5 and Mythos 5 excluded	Not addressed by model	Not addressed by model
Trained on by default?	No, not without permission	Not on the API tier	No, since March 1, 2023
Retained anyway if flagged?	Yes, up to 2 years	Not addressed in the article	Yes, for CSAM-flagged files

Two cells deserve a caveat rather than a verdict. "Not addressed by model" for Mistral and OpenAI does not mean every model is available under ZDR; it means neither vendor publishes a model-level carve-out the way Anthropic does for its Covered Models, so the question belongs in your contract rather than in a table. And "not addressed in the article" for Mistral on flagged content means the ZDR help article does not discuss the legal carve-out, not that none exists. Read the data processing addendum for that.

// In the vendors' own words

What each vendor's zero data retention arrangement covers

Anthropic (Claude)

How you get it: Contractual arrangement, enabled per organization by the sales team
Covered: Messages API, including PDFs sent inline. Token counting, prompt caching, thinking, citations, web search.
Named exclusions: Files API (files kept until deleted), batch processing (29 days), code execution (30 days), Console and Workbench, Fable 5 and Mythos 5.
Default without ZDR: Conversation content not retained by default; Covered Models require 30 days.

Mistral

How you get it: Scale plan only, then activated from the admin panel
Covered: Stateless endpoints including /v1/ocr, chat completions, embeddings, moderations, classifications and audio.
Named exclusions: Agents, batch processing files, conversations, libraries, /v1/files, Vibe Work and Chat.
Default without ZDR: Retention applies for abuse monitoring unless ZDR is active.

OpenAI

How you get it: Applied to approved customers at organization or project level
Covered: Chat completions, responses, embeddings, moderations, audio transcriptions, image generations.
Named exclusions: Assistants, threads and conversations, which retain until deleted. Videos are blocked for ZDR requests.
Default without ZDR: Not used for training. Abuse monitoring logs kept up to 30 days.

// The finding

The file upload is the step that sits outside ZDR

Read the two Anthropic table rows next to each other. PDF support through the Messages API is ZDR eligible, with the note that eligibility "applies to PDFs sent inline through the Messages API, not through the Files API". The Files API row is marked "No", with the reason given as "files retained until explicitly deleted". Same document, same vendor, same account, two different retention outcomes, decided entirely by how the bytes reached the model.

Mistral draws the same line in one sentence: ZDR "does not apply to stateful APIs or products, including agents, batch processing files, conversations, libraries, /v1/files, Vibe Work, and Chat". The "/v1/ocr" endpoint is on the supported list. The file store next to it is not. OpenAI's version of the same boundary is that assistants, threads and conversations are ineligible and retain data until deleted.

None of this is hidden. It is written plainly in all three sets of documentation. It gets missed because the question a security reviewer asks is "does this vendor offer zero data retention", and the honest answer to that question is yes, while the answer to the question they meant to ask can be no. Anthropic is unusually direct about the consequence: using a feature marked "No" is not blocked, and "using one is a choice to step outside your ZDR arrangement for that specific data".

The practical fix, where the vendor supports it: send the document inline in the request body rather than uploading it first. On Anthropic that keeps a PDF inside both the ZDR and the HIPAA arrangement. It costs you the ability to reference the same file across many calls, and it puts the whole document inside your request size limit, so it works for a 10-page invoice and not for a 400-page loan file. Where inlining is not viable, the upload needs its own answer in your risk register rather than an assumption.

// Where privacy meets the invoice

The 50 percent batch discount costs you the arrangement

Batch processing is the standard way to cut the bill on a large document job. Mistral and Anthropic both discount batch by 50 percent, and OpenAI's batch tier roughly halves the rate too, which on our own LLM OCR pricing table is the difference between about $1.67 and about $0.84 per 1,000 pages on the cheapest model.

Batch is also excluded from ZDR. Anthropic marks it "No" and states the reason as "29-day retention; async storage required", which is inherent rather than a policy choice: an asynchronous job has to be stored somewhere while it waits. Mistral names batch processing files in its exclusion list. So the cheapest way to run a large extraction job is also the way that steps outside the zero-retention arrangement, and the two facts are almost never presented together because pricing pages and trust pages are written by different teams.

Price the job both ways before you commit. If the discount is worth more than the retention exposure on that particular document type, take it deliberately. If it is not, run the volume synchronously and put the real number in the business case rather than the batch number. Either answer is defensible. Discovering the trade-off after the security review is the expensive outcome.

// The model-level carve-out

Some models cannot run under zero data retention at all

Anthropic designates Claude Fable 5 and Claude Mythos 5 as Covered Models, and states that they "require 30-day data retention; ZDR is therefore not available for either model". A ZDR organization that calls one gets a 400 error telling it that the organization or workspace "must have data retention enabled". The 30-day window exists to support safety work and applies wherever those models are offered.

There is a documented middle path. An organization with a ZDR arrangement can enable 30-day retention on one workspace, which unlocks those models there while every other workspace keeps zero retention. That is a reasonable design if a small part of your workload genuinely needs the strongest model and the rest does not.

The reason this belongs on a document processing page is upgrade risk. A model choice is not usually treated as a compliance decision, so a team that signed a ZDR arrangement in January and moved to a newer model in June can break its own arrangement without anyone filing a change request. Neither Mistral nor OpenAI publishes an equivalent model-level carve-out today, which is worth confirming in your own contract rather than assuming it stays that way.

// Read the carve-out

Zero data retention does not mean nothing is ever kept

Every ZDR arrangement in this market carries a legal and safety ceiling, and the honest way to hold it is as a default state rather than a guarantee. Anthropic states that even with ZDR or HIPAA arrangements in place it may retain data where required by law or where content has been flagged by automated trust and safety systems, and that in that case it "may retain inputs and outputs for up to 2 years". OpenAI states that image and file inputs are scanned for CSAM content and may be retained "even if Zero Data Retention" is enabled.

This is not a reason to skip ZDR. It is a reason to write the ceiling into your risk register instead of a zero, because a control described to an auditor as absolute and then found to have a two year exception is a worse outcome than one described accurately from the start. It also affects document selection. A vendor's automated safety systems will look at whatever you send, so an internal decision about which document classes go to a third-party model at all is a more reliable control than the retention setting on top of it. That is the reasoning that pushes some workloads to on premise OCR software instead.

// Ask in writing

Six questions to send your vendor before the security review

Which endpoint does our document actually touch?

Trace the real call path, not the marketing one. If the pipeline uploads before it reads, the upload is a separate endpoint with a separate policy, and on two of these three vendors it is named in the exclusion list.

Does the arrangement cover the model we chose?

Ask by model name, not by vendor. Anthropic publishes two models that cannot run under ZDR at all, so a model upgrade six months from now can quietly break the arrangement you signed.

Does it survive the way we run volume?

Batch is where the discount is and where the exclusion is. Price the job both ways before you assume the cheaper number is the one you get.

Who has to approve it, and how long does that take?

Two of the three gate ZDR behind a human. If your security review has a date on it, start this conversation before you write the integration, not after.

What is kept regardless?

Every vendor keeps a safety and legal carve-out. Get the ceiling in writing so the number in your risk register is the real one.

Does it extend to a second account?

Anthropic states plainly that enablement "does not automatically extend to other organizations under the same account". If your company has a sandbox organization and a production one, that is two arrangements.

// Where we stand

How DocuOCR answers the same questions

DocuOCR is a cloud product. Your documents are processed on our infrastructure, they are encrypted in transit and at rest, and we do not use customer documents to train models. We are not going to publish a retention window on a marketing page and call it a commitment, because the document that binds us is the data processing addendum, and that is the one your counsel should read.

What we will say plainly is where we are the wrong answer. If your requirement is that the document never leaves your network, no cloud vendor on this page satisfies it, including us, and the honest options are a self-hosted model or a vendor-supplied container. We wrote up what that actually costs on self hosted OCR cost, including the utilization point below which a GPU is far more expensive than a per-page API. If your requirement is a specific contractual arrangement, ask us before you build the integration, which is the same advice this page gives about every other vendor on it.

Frequently asked questions

What does zero data retention mean?: Zero data retention, usually shortened to ZDR, means the provider does not store your inputs or outputs at rest once the API response has been returned. Anthropic words it as not storing "customer prompts or responses at rest after the API response is returned". It is a contractual arrangement applied to your account, not a flag you set on a single request.
Does zero data retention cover document uploads?: Usually not, and this is the detail that catches document teams. ZDR is written for stateless inference. Anthropic marks the Files API "No" for ZDR because files are "retained until explicitly deleted", and Mistral names "/v1/files" directly in its exclusion list. If your pipeline uploads a PDF before reading it, the upload sits outside the arrangement even though the model call sits inside it.
Is Mistral OCR covered by zero data retention?: The "/v1/ocr" endpoint does appear on Mistral's list of stateless APIs that support ZDR, which is unusual and worth knowing. The caveat is in the same article: ZDR "does not apply to stateful APIs or products, including agents, batch processing files, conversations, libraries, /v1/files, Vibe Work, and Chat". So the extraction call is covered and the file handling around it needs its own answer.
Does OpenAI train on API data?: No, not by default. OpenAI states that data sent to the API "is not used to train or improve OpenAI models (unless you explicitly opt in)", a default in force since March 1, 2023. Training and retention are separate questions though: without ZDR, abuse monitoring logs containing prompts and responses are still retained for up to 30 days.
Does Anthropic train on API data?: Not without permission. Anthropic states that "retained data is never used for model training without your express permission" and that conversation content "is not retained by default". The exception it names is Covered Models, currently Claude Fable 5 and Claude Mythos 5, which require 30-day retention.
How do I get a zero data retention agreement?: You ask for it, and on two of the three vendors a human has to approve it. Anthropic says to "contact the Anthropic sales team" and enables ZDR per organization, so a second organization under the same account does not inherit it. Mistral gates ZDR to the Scale plan. OpenAI applies ZDR at the organization or project level for approved customers.
Can I use batch processing with zero data retention?: On Anthropic and Mistral, no. Anthropic marks batch processing "No" for ZDR with "29-day retention; async storage required", and Mistral names batch processing files in its exclusion list. That matters commercially because batch is where the 50 percent discount lives, so the cheapest way to run a large document job is also the way that steps outside the arrangement.
Are the newest models available under zero data retention?: Not all of them. Anthropic designates Claude Fable 5 and Claude Mythos 5 as Covered Models that "require 30-day data retention; ZDR is therefore not available for either model", and a ZDR organization calling one gets a 400 error. You can enable 30-day retention on a single workspace to unlock them while the rest of the organization stays at zero.
Is zero data retention the same as HIPAA compliance?: No, and on Anthropic they are alternatives rather than additions. Anthropic states that HIPAA-ready API access "is designed as an alternative to ZDR for organizations handling PHI" and that if you handle PHI, "HIPAA readiness is the arrangement to use; you do not also need ZDR". HIPAA readiness keeps data with safeguards around it; ZDR removes the data.
Does zero data retention mean literally nothing is kept?: No. Every vendor keeps a legal and safety carve-out. Anthropic says that even with ZDR in place, flagged inputs and outputs may be retained "for up to 2 years". OpenAI says image and file inputs scanned for CSAM may be retained "even if Zero Data Retention" is enabled. Zero is the default state, not an absolute guarantee.

Settle the data question before you build the integration

Upload a document, look at the fields and the confidence scores, and see whether the output is worth the security review in the first place.

Extract a document free Compare the OCR APIs on retention

Related references

OCR API data retention

What AWS, Azure and Google keep, and for how long

LLM OCR pricing

What these same models cost per 1,000 pages

Claude OCR pricing

Visual tokens turned into a per-page rate

OpenAI OCR pricing

Patch counts, multipliers and the batch tier

Mistral OCR pricing

The endpoint that carries ZDR on the Scale plan

On premise OCR software

When the document cannot leave at all

From the same family of tools