When a company reviews an AI vendor, the question that gets asked is whether the vendor trains on its data. The answer, for the commercial APIs of Anthropic, OpenAI and Google, is no by default, and that settles very little. What a security or legal reviewer actually needs is AI vendor data retention: how long prompts and outputs sit on the vendor's systems, who can see them, which features change the answer, and what "zero data retention" does and does not cover.
This guide reads the three vendors' own documentation side by side and turns it into a checklist. It is a reading of published pages as of 19 October 2026, not legal advice, and these pages change; confirm against your contract.
What the three vendors document
The table compares what each page states. Where a page is silent, we say so rather than fill the gap.
| Spec | Anthropic | OpenAI | Google Cloud |
|---|---|---|---|
| Training on API customer content | Commercial terms: "Anthropic may not train models on Customer Content from Services."[2] | "As of March 1, 2023, data sent to the OpenAI API is not used to train or improve OpenAI models" unless you opt in.[3] | Will not train or fine-tune on your data "without your prior permission or instruction."[4] |
| Default retention of prompts and outputs | Conversation content "is not retained by default", except Covered Models, which require 30-day retention.[1] | Abuse-monitoring logs retained "for up to 30 days, unless longer retention is required by law."[3] | Cached data is deleted no later than 24 hours after it is written; the page ties other retention to configuration.[4] |
| Zero-retention route | Per-organisation arrangement requested through sales; each new organisation must be enabled separately.[1] | Zero Data Retention or Modified Abuse Monitoring, both requiring prior OpenAI approval.[3] | Assembled from settings: no request-response logging, caching off if needed, and an abuse-monitoring exception that customers can request.[4] |
| Examples outside the zero-retention route | Claude Console, Managed Agents (transcripts persist until deleted), consumer plans, Batch and Files APIs, code execution.[1] | Files, Conversations, Assistants, Threads, Vector Stores, fine-tuning, Batches and Evals are listed as not ZDR-eligible.[3] | Grounding with Google Search (logs up to 3 days), Grounding with Google Maps (30 days), the Deep Research agent (7 days).[4] |
| Retention that applies regardless | Flagged content may be retained for up to 2 years; legal holds apply.[1] | Longer retention where required by law.[3] | Prompts may be logged for abuse detection unless an exception is granted.[4] |
Three details are worth a closer look.
Anthropic's stricter models. The page says its Covered Models, which it names as Claude Fable 5.1, Claude Mythos 5.1, Claude Fable 5 and Claude Mythos 5, require 30-day retention and are not available under zero data retention unless Anthropic expressly authorises it. An organisation with a zero-retention arrangement can enable 30-day retention for a single workspace to use them while other workspaces stay at zero.[1] A team that signed a zero-retention agreement for one model may therefore find the newest model outside it.
OpenAI's endpoint table. OpenAI publishes retention per endpoint. Chat completions, Responses, embeddings and Realtime show 30-day abuse-monitoring retention and are ZDR-eligible with limitations; Files, Assistants, Threads, Vector Stores, fine-tuning, Batches and Evals are not. Under ZDR the store parameter on the Responses and Chat Completions endpoints is always treated as false.[3] The architecture decision (which endpoints your application calls) therefore changes the retention outcome as much as the contract does.
Google's configuration model. Google describes zero retention as something "customers must take specific actions" to achieve: leave request-response logging off, set store = false for the Interactions API (the default is true), avoid the Sandbox API, and avoid Live API session resumption.[4] This puts more responsibility on the customer's engineers and less on a contract clause, which suits some organisations and not others. The page also notes that some Advanced AI features may involve extra logging where zero retention "may not be possible."
Why the exceptions matter more than the headline
Our reading across the three pages: the headline promise (no training, short retention) holds for the plain request-response path. Retention returns when a feature has to remember something for you. File storage, batch jobs, threads and conversations, vector stores, agent sessions and search grounding all keep state, and each vendor documents that state separately.[1][3][4] A retrieval-augmented or agent product is therefore likely to touch at least one retained feature, even if a design review assumed "zero retention".
Anthropic's page says this directly: using a feature marked "No" for zero retention "is a choice to step outside your ZDR arrangement for that specific data."[1] That is a useful framing for a review. Each stateful feature is a separate data-handling decision, not a footnote to the main contract.
Contract and documentation are different layers
Anthropic's commercial terms add commitments the documentation page does not: customers retain rights to their inputs and own outputs, Customer Content is the customer's Confidential Information, and Anthropic defends and indemnifies customers against claims that paid use or outputs infringe third-party IP, subject to exclusions such as customer modifications and combinations with non-Anthropic technology.[2] The terms page, as we read it, sets no retention period of its own and says processing is governed by a separate data processing addendum, which we did not review.[2]
That is a general point for buyers. The product documentation tells you what the service does today. The contract and data processing addendum tell you what the vendor is bound to do, and for how long. Reconcile the two, and ask what notice you receive when the documentation changes. The Google page we read was last updated on 15 October 2026; the OpenAI and Anthropic retention pages we opened carried no date, which is itself something to record.
The deployer-side log question
If you deploy a high-risk AI system under the EU AI Act, Article 26(6) says deployers shall keep the logs automatically generated by the system "to the extent such logs are under their control" for a period appropriate to the intended purpose and "of at least six months, unless provided otherwise in applicable Union or national law."[5] Our interpretation is that vendor-side zero data retention and this duty address different copies of the data: the vendor's retention is about what the provider holds, while the duty concerns logs under your control. A zero-retention contract does not remove your need to keep your own logs, and it can make them the only record. Whether a given use is high-risk, and which dates apply to it, is a separate legal question; see our EU AI Act guide for general-purpose model providers for the dates the Commission lists and note that deployer obligations follow their own timeline.
A procurement checklist
Use these as questions for the vendor and for your own architecture review.
- Which exact products and endpoints will we call? Map each to the vendor's retention entry. Treat anything that stores state as a separate decision.
- What is the default retention for prompts and outputs, in days, and where is it written down? Record the page, its date and the retrieval date.
- Is zero data retention available for our models and features? Anthropic's Covered Models and OpenAI's non-eligible endpoints show it may not be.[1][3]
- Who enables it, and per what scope? Anthropic enables per organisation; a new organisation starts without it.[1]
- What is retained regardless? Ask about abuse monitoring, safety flags and legal holds, including maximum periods (Anthropic documents up to 2 years for flagged content).[1]
- Does caching, grounding or an agent feature add retention? Google documents 24-hour caches, 3-day and 30-day grounding logs and a 7-day agent session store.[4]
- Where is the data processed? OpenAI's residency controls need sales approval; Google distinguishes regional endpoints from the global one.[3][4] Neither page, as we read them, answers residency for every feature, so ask.
- What do we log ourselves, for how long, and who can read it? This is the record your own audit, incident response and any deployer duty will rely on.[5]
- How will we learn of changes? Subscribe to the vendor's change notices and re-check the pages quarterly.
- Is the answer in the contract? If a promise exists only in documentation, ask for it in the order form or data processing addendum.
What we did not verify
We read each page once on 19 October 2026 and did not test any configuration. We did not review any vendor's data processing addendum, trust centre or security attestations, and we did not read Amazon Bedrock or other reseller documentation, where, according to Anthropic's page, the cloud provider rather than Anthropic is the data processor.[1] Retention periods are the vendors' stated policies, not measured behaviour.
For the related cost side of the same design decisions, see tokens, context windows and AI cost, and for evaluation practice before you commit to a vendor, see how to build LLM evals for your application.
