What does zero data retention actually cover?

Three different promises get collapsed into one word. We do not train on your data is the weakest and is now the default at both large vendors: OpenAI has not trained on API data since March 2023 unless you opt in, and Anthropic says the same of its commercial products. We do not retain your data is a stronger and separate contractual claim, and a provider can honestly make the first while doing the opposite of the second. Your data does not leave this jurisdiction is a third thing again, orthogonal to both, and the one people usually mean when they say GDPR. Ask for the one you need by name, because a vendor answering a different question is not lying to you.

Zero data retention covers the stateless call. OpenAI publishes the eligibility list, and it is the clearest statement of the mechanism anyone has written: completions, responses, embeddings, audio and moderation are eligible; Assistants, conversations, threads, vector stores and fine-tuning jobs are not, and hold their state until you delete it. That is not a policy gap. A feature that remembers for you has to store something, so it cannot also not store anything. Anything on the API that spares you from resending your own history is, by construction, outside the retention promise.

Which means the compliant architecture is the expensive one. To stay inside the promise you keep the conversation in your own database and send the whole history back every turn. That is exactly the re-reading that the bill page prices, and on an agent-shaped workload it is 94% of every prompt token you are billed for. Statelessness is what ZDR requires and what makes the invoice grow. It is worth knowing that the two pull against each other before someone proposes server-side threads as a cost saving.

One correction, because we nearly published the opposite. Prompt caching stores your prefix between calls, so it looks like it should be incompatible with a zero-retention agreement. It is not: caching continues to work under ZDR at the major vendors. Every model in the bill page’s table has a cache-read price, and that discount survives. We assumed the tension was real, checked, and were wrong.

The carve-outs are where the data actually is. Abuse monitoring is the big one: OpenAI retains abuse logs for up to 30 days by default, and a zero-retention agreement is precisely the thing that excludes your content from them. Feedback is the surprising one: Anthropic’s policy says that when a user presses thumbs up or down, the entire related conversation is stored for up to five years. One button in your own interface, five years of retention, and it will not appear in any architecture diagram. Tools are the quiet one: OpenRouter’s zero-retention mode explicitly does not extend to third-party plugins such as web search, so the prompt is covered and the search query your model generated from it is somebody else’s policy.

Through a router, compliance becomes a routing constraint with a price. OpenRouter enforces zero retention by restricting your request to endpoints that have a zero-retention policy, and where a provider’s policy is unknown they assume it both retains and trains. That default is the right one. It also means turning the setting on removes routes from your pool, and which routes you can reach is what sets your price: the router page measured the same model across fourteen routes varying 6.8x in cost, 53% to 100% in cache hit share, and one route failed 55 calls in 100. A compliance switch is also a performance and cost decision, and nobody presents it that way.