Anthropic (Claude) Data Retention (2026): 30 Days on the API, Five Years If You Say Yes

Short answer - which Claude are you actually using?

• Claude Free / Pro / Max (the apps) → you choose whether your chats train models. Say yes and they're kept up to five years. Say no and the old 30-day period applies. Deleted chats leave the backend within 30 days.
• Claude for Work (Team / Enterprise), Gov, Education → no training by default; your organisation's admins set chat retention.
• The API with your own key → no training without express permission, backend deletion within 30 days, a stateless Messages API, and Zero Data Retention if Anthropic's sales team agrees.
• Nothing at all → possible for meeting notes, because summaries and transcription both run fine on-device.

I make Humla, a Mac meeting notetaker that can send summaries and chat to Anthropic with your own key, or to nobody at all, so I have a horse in this race. Everything below comes from Anthropic's own pages, checked 27 October 2026, linked so you can read them yourself instead of trusting me.

#Why this is the fourth one of these

I've now done this for OpenAI, Deepgram and Groq, and I honestly thought I was done. Then Humla grew an Anthropic provider in September (pick Claude for summaries, pick it for chat over your notes), and suddenly I was telling people "bring your Claude key!" without having read the boring document behind it. Which is a little hypocritical for a guy who has written three posts about reading the boring document.

So. The boring document. Several of them, actually, because Anthropic splits this across a privacy centre, a support site and the developer docs, and they don't all say it in the same words (more on that in a minute, because it's the most interesting thing I found).

#The products people mix up

Same trap as with OpenAI: "Claude" is an app and an API, and the rules differ completely.

Sources: Anthropic's consumer terms update (published 28 August 2025), the privacy centre's commercial retention and commercial training articles, and the API data retention docs. Checked 27 October 2026.

The consumer change is the one most people half-remember. In August 2025 Anthropic started asking Free, Pro and Max users whether their chats could be used to improve Claude, with a deadline of 8 October 2025 to choose. Saying yes extends retention to five years for new or resumed chats. Saying no keeps you on, in Anthropic's words, "our existing 30-day data retention period." Commercial Terms services, which includes the API and the Work plans, were explicitly left out of that change.

Two small asterisks on the app side. Incognito chats aren't used to improve Claude whichever way you answered. And the thumbs up / thumbs down button is its own little trapdoor: press it, and the whole related conversation is stored for up to five years and may be used for training, de-linked from your account ID. That one applies on the Work plans too, where an owner can switch the button off with the Rate chats setting. (I find it genuinely funny that the most retained thing in the whole Claude ecosystem is "I liked this answer.")

#The API: two pages, two sentences, one answer

Okay, here's the bit that had me squinting at my screen for a while.

The privacy centre's commercial retention article (dated 1 July 2026) says: "For Anthropic API users, we automatically delete inputs and outputs on our backend within 30 days of receipt or generation." Clear enough.

The developer docs page on API data retention says: "Conversation content (your prompts and Claude's outputs) is not retained by default." Also clear! Also... apparently the opposite?

It isn't, once you read the paragraph around it. The docs sentence sits under a heading about features that necessarily require storage - batches, files, code containers - and it's a promise about what those features keep, not about the backend window. The Messages API itself is, per the docs, "stateless": you resend the whole conversation every time, and nothing is kept as an object you can come back to. So my reading, and it's a reading, not a quote: Anthropic stores no conversation state for you, and its backend copy of the request is deleted within 30 days. Same shape as OpenAI's split between "application state" and "abuse-monitoring logs", just written across two websites.

The exceptions to the 30 days, from the privacy centre:

• Things you chose to store - the Files API keeps files until you delete them or they hit an expiry you configured.
• A Zero Data Retention agreement, below.
• Usage Policy enforcement - if a request is flagged, inputs and outputs can be kept for up to 2 years, and trust-and-safety classification scores for up to 7 years.
• The law - "as required by law", which is the sentence every one of these pages has and none of them can promise you about.

Training on API data needs your "express permission". By default it doesn't happen.

#The feature-by-feature table

The docs publish a table of which API features are eligible for ZDR and what each one keeps. The rows that matter if you're building something that handles meetings:

From Anthropic's API data retention docs, read 27 October 2026. The page shows no date; re-check before relying on it.

The batch row is the one I'd have got wrong if you'd asked me yesterday. Batches are the cheap way to run a pile of summaries overnight, and they're also the one route where your transcripts are deliberately parked on Anthropic's side for most of a month.

#Zero Data Retention, and the models it doesn't cover

ZDR is what you'd expect: Anthropic "does not store customer prompts or responses at rest after the API response is returned." Flagged content and legal holds still override it.

How you get it is also what you'd expect, sadly: "contact the Anthropic sales team." It's enabled per organisation by your account team. No toggle. Of the four providers I've now read, only Groq lets you flip this yourself, and Anthropic sits firmly with OpenAI in the "let's set up a call" camp. If you're a three-person agency, that call is not happening.

Then there's a newer wrinkle, effective 9 June 2026: Anthropic designates some models as Covered Models - currently Claude Fable 5, Fable 5.1, Mythos 5 and Mythos 5.1 - and prompts and outputs for those "are retained for 30 days to support our safety work, on every platform where these models are offered." Its explanation is that some attacks only show up across many requests, so they want the window to look for patterns. For a normal API account this changes nothing, because you were on 30 days anyway. For a ZDR organisation it means those models aren't available under ZDR unless Anthropic expressly authorises it, and the workaround is to switch on 30-day retention for one specific workspace.

Which is the first time I've seen a provider tie retention to which model you picked. It won't be the last, I suspect.

#No EU region to pick

Germany and the UK send this site more readers than anywhere except the US, so this matters. The first-party API has an inference_geo setting with exactly two values: "global", the default, and "us", which costs 1.1x the standard rate on Claude 4.6 and later models. For data at rest, the docs say plainly: "Currently, "us" is the only available workspace geo."

So there is a US-only option and there is an anywhere option. There isn't a Europe option. If you need EU processing for GDPR reasons, the route is Claude through a cloud platform like Bedrock or Google Cloud, where the region is set by the endpoint you call - and then you're reading that platform's boring document instead. (The boring documents, it turns out, are a fractal.)

#No court order this time

With OpenAI, the honest 2025 answer to "do they delete my data?" was "a judge says no," until the preservation order ended in September 2025. I went looking for an equivalent at Anthropic and found nothing: none of the retention or training pages I read mention a court order or preservation order. The docs list "legal holds" as something that overrides any arrangement, and the privacy centre says chats may be kept "as required by law." That's the standard clause, not a live event, as far as Anthropic's own pages show.

Same structural point as last time, though, and it doesn't expire: a retention policy is a promise, and a court can overrule it without asking either of you.

#The four providers, side by side

Anthropic column read 27 October 2026; OpenAI 28 July 2026; Groq 11 August 2026; Deepgram 4 August 2026. All four change - re-check before a compliance decision.

Anthropic's API comes out solid and unsurprising: no training, a 30-day window, a stateless API. Its two soft spots are that the off switch needs a salesperson and that Europe isn't on the map. If I had to rank the defaults alone, Groq still wins. If I had to rank "how much can a small team control without a phone call", Groq wins again, and that's getting slightly embarrassing for everyone else.

#What this means if you record meetings

Say you record a client call and send it to Claude through your own key. In Humla's case, that looks like this:

1. The audio never goes to Anthropic at all. Humla doesn't use Claude for transcription, so that happened somewhere else first - on your Mac, or with OpenAI, Deepgram or Groq, each with its own policy above.
2. The transcript goes to /v1/messages to make the summary. Backend copy deleted within 30 days, no training, up to two years if something in it trips a Usage Policy flag.
3. Chat over your notes, if you use Claude for it, sends the note excerpts Claude asks for to the same endpoint. Anthropic has no embeddings endpoint, so Humla searches your notes by keyword for Claude instead of shipping chunks off to be vectorised.

None of it trains a model. All of it crosses a network boundary you don't control, and step two keeps a copy for up to a month. Whether that's fine depends on what was said: a sprint planning call, sure. A therapist's session or a lawyer's client meeting is a different conversation, and no vendor's retention page discharges a duty of confidentiality on your behalf.

One Humla-specific thing worth knowing: the model picker lists whatever your key can use, straight from Anthropic's /v1/models, and the default is Claude Sonnet 5. If you pick Fable from that list, you're on a Covered Model. On a normal account that's still the same 30 days. On a ZDR account, it's the exception the previous section described.

#The option of not sending it

This is where my bias lives, so weigh it accordingly.

Humla is my open-source Mac notetaker. It records your mic and the call's system audio with no bot in the meeting, and when you choose Anthropic it talks to exactly two endpoints: /v1/messages for summaries and chat, and /v1/models to check your key and fill the model list. You paste the key once; it lives in your macOS Keychain; the bill is yours. So the analysis above is the analysis for Humla-with-Claude, with no middleman in between - which is the whole point of bring-your-own-key, and also means there's no magic either.

Or you skip the question. Whisper on your Mac's GPU, offline speaker labels, and summaries and chat against a local model through Ollama. No request, no retention policy, nothing to read. You can mix it per language, too, so one language stays on the Mac while another goes to a cloud provider.

The tradeoffs, honestly: a local model on a laptop is a real step below Claude on a long, messy meeting, and I'm not going to pretend otherwise. On-device transcription is slower and wants a few gigabytes of model download first. Humla is Apple Silicon only, with no Windows, no mobile, no integrations, and a much smaller community than the big names. If the content isn't sensitive, Claude through your own key is a perfectly defensible, well-documented choice. I'd just like you to make it having read the document. If you'd rather start with one of the other three providers, there's a walkthrough for getting an OpenAI, Deepgram or Groq key.

#The short version

Anthropic's API story is clean: no training without express permission, a stateless Messages API, backend deletion within 30 days, two years for flagged content. The Claude apps are the one to think about, because saying yes to training moves you from 30 days to five years. ZDR exists but needs a sales call, Covered Models carve their own 30-day rule into it, and there's no EU region on the first-party API.

Four providers in, the pattern I keep finding is that the interesting fact is never on the landing page. This time it was split across two websites that read, at first glance, like they disagree. See what runs locally and what private means here - and if you send meetings to Claude, go and read the batch row of that table, because that one surprised me.

All facts above verified against Anthropic's published pages - the privacy centre's retention, training and Covered Models articles, the API data retention and data residency docs, and the August 2025 consumer terms announcement - on 27 October 2026, and all Humla claims against the app's own source on the same date. These pages change; re-check before making a compliance decision on them.