Data Retention & What Goes Back

Detail

Per the Privacy Center: on consumer plans, conversations are saved to your account until you delete them. The governing setting’s printed label is “Help Improve Claude” (Settings > Privacy — the Privacy Center’s walkthrough names the exact path and also calls it “the model training setting”; the storage article’s prose calls it “the model improvement setting”). Earlier drafts of this curriculum and the group’s document called it “Improve Claude for everyone”; that phrase appears on neither current page (both fetched and searched 2026-07-28 — receipts R33, R34).

The setting’s reach, stated declaratively:

• On — new and resumed chats and coding sessions may be used for model training, and can be retained in de-identified form for up to 5 years in the training pipelines.
• Off — Anthropic commits not to use “any new chats and coding sessions” for future model training. That is a use commitment; the page publishes no automatic retention window for this state, so the off-state retention period is unstated — not zero, and not “kept only until you delete them.” One wording divergence between the two chipped pages, noted rather than smoothed: the settings walkthrough says “any new chats and coding sessions” while the storage article’s version reads “your previous or new chats or coding sessions” (both quoted in the receipts). This node teaches the narrower new-chats-only reading — the conservative direction: if the wider promise holds you lose nothing, and nothing here overstates your protection if it doesn’t. Which wording governs is a fair confirm-with-your-contact item.
• Delete — a deleted conversation is purged from Anthropic’s back-end systems within about 30 days, and that applies regardless of the toggle.

Three scope facts that matter for this group: both articles scope themselves to “our consumer products such as Claude Free, Pro, Max and when accounts from those plans use Claude Code” — so consumer Claude Code sessions ride the same toggle, which is easy to miss when the coding surface feels separate from chat. Max is inside this claim, same as Free and Pro. And the toggle’s default for a new account is stated on neither page — the settings article gives change instructions only, and the storage article names no default (searched both, fetched 2026-07-28, for “default”, “by default”, “new account” — receipts R33, R34) — so “what happens if I do nothing?” is itself a confirm-don’t-assume item: Where to confirm the training toggle is off, plan by plan.

Source & currency

• Source: How long do you store my data? — Anthropic Privacy Center (cited by the source doc) + How do I change my model improvement privacy settings? (the settings walkthrough the label and path rest on — added 2026-07-28).
• Status: CONFIRMED 2026-07-28 (receipts R19, R31, R33, R34) — raised from ▪ at CG4, re-verified 2026-07-28 with the settings walkthrough added as a second source. Both figures are verbatim on the storage page: “we may retain your data in a de-identified format for up to 5 years in our model training pipelines” and, on deletion, “Deleted from our back-end storage systems within 30 days.” The page also scopes itself explicitly — “This article is about our consumer products such as Claude Free, Pro, Max and when accounts from those plans use Claude Code” — which is both why nothing here should be read across to a commercial plan and why consumer Claude Code sessions are inside the claim. The toggle’s label was corrected 2026-07-28: the settings article prints “Help Improve Claude”; the “Improve Claude for everyone” label this node previously carried appears on neither page (receipts R33, R34).
• Why the chip could move on a claim that rests partly on an absence. This node carried ▪ because the OFF state has no published automatic retention window, and an unverified absence is not a verified fact. At CG4 the absence was verified rather than assumed, two ways (§4.4): every duration string on the page was enumerated — 30 days, 5 years, 2 years, 7 years, 5 years — and none is tied to the setting being off; and the OFF paragraph was read in full, where the page makes a use commitment (“we will not use your previous or new chats or coding sessions for future model training”) and not a retention one. The absence is now a finding with a receipt behind it, so the node states what the source states.
• One structural detail, recorded because it explains the gap without excusing it. The article carries a heading — “Standard Retention Timeframe” — with no text under it. The section that would name a default window is present and empty, confirmed in the page’s raw HTML across three independent retrievals. Read the honest claim precisely: no automatic retention window for the off state is published here, not your data is not retained when the setting is off. The page does not say the second thing, and the difference is exactly the sort a compliance decision turns on.
• ⚑ Compliance-critical: retention windows move — confirm current specifics with your Anthropic contact before relying on this for a compliance decision. This is a teaching aid, not compliance or legal advice.

Related

• Data Retention & What Goes Back — parent domain
• Commercial plans: no training by default and by contract — the exception is data you hand over yourself — the contractual opposite on Work/Education/Gov plans
• Where to confirm the training toggle is off, plan by plan — where the toggle actually lives, plan by plan
• Risk of using personal Pro/Max accounts for org work — why personal accounts for org work are the live risk here
• Retention (data) — the term, plainly
• Ensuring a rare disease is well-represented in the model — the same training pipeline, seen from the other side: the group also wants its disease represented in future models

Detail

The contractual commitment. Anthropic’s Commercial Terms of Service, §B (Customer Content): “Anthropic may not train models on Customer Content from Services.” “Customer Content” is defined in the same section as Inputs and Outputs together — your submissions and Claude’s responses. That sentence carries no toggle and no conditions.

The policy statement. Anthropic’s Privacy Center, in the article scoped to commercial products: “By default, we will not use your inputs or outputs from our commercial products (e.g. Claude for Work, Anthropic API, Claude Gov, etc.) to train our models.”

The exception — and it is the one that matters operationally. From the same article: “If you explicitly report feedback or bugs to us (e.g. via our thumbs up/down feedback button), or otherwise choose to allow us to use your data, then we may use your chats and coding sessions to train our models.”

This resolves an apparent discrepancy in the documentation. The API documentation phrases the commitment conditionally — “Retained data is never used for model training without your express permission” — which reads, in isolation, like a weaker promise than the Commercial Terms’ flat prohibition. They are consistent: the feedback button is the express permission. Clicking thumbs-up/down is the act of granting it. Nothing else in ordinary use grants it.

What that means in practice. The structural difference between plan families holds: consumer plans govern training by a user-facing setting (Consumer plans: the training toggle governs training use — up to 5 years if on; no automatic window published if off); commercial plans are governed by contract, with an opt-in path rather than an opt-out toggle. For an organization deciding whether to standardize on an org plan rather than members’ personal accounts (Risk of using personal Pro/Max accounts for org work), this is a genuine argument in favor — but the protection is only as good as the discipline around the feedback button. See The feedback button hands over the whole conversation for 5 years — and it is the one thing that permits training: that button hands over the entire conversation, for five years, and may train on it.

Still worth asking your account team (Which contract terms actually govern the working group's accounts): confirm which contract governs each account, and whether your executed terms match the published Commercial Terms.

Source & currency

• Source: Is my data used for model training? — Anthropic Privacy Center, commercial · Anthropic Commercial Terms of Service §B · API and data retention.
• Status: CONFIRMED — re-verified 2026-07-28 (receipts R17, R18, R12, R36, R40) — all three passages verified on the live pages 2026-07-25 and re-verified 2026-07-28, when the API docs page (home of the “express permission” sentence the node reconciles) became the third chip.
• Correction history — worth reading, because this node was wrong twice in opposite directions. Before 2026-07-25 it claimed commercial data is “not used to train models — ever … not a setting someone could flip,” sourced to the consumer retention article, which excludes commercial readers by its own scope line. During the 2026-07-25 patient-safety round it was over-corrected to STALE-RISK on the reasoning that no source stated the commitment — but that check had only covered the retention articles, and the commitment is published in the training article and in the Commercial Terms. The independent review caught it the same day. Asserting a negative (“no source says this”) requires searching where the claim would actually live; a partial search reported the house empty after opening one drawer. The current text carries the commitment and its real exception.
• ⚑ Compliance-critical: plan-tier guarantees are contract terms — confirm your executed terms with your Anthropic contact before relying on this for a compliance decision. This is a teaching aid, not compliance or legal advice.

Related

• Data Retention & What Goes Back — parent domain
• Consumer plans: the training toggle governs training use — up to 5 years if on; no automatic window published if off — the consumer-side contrast, which is a toggle
• The feedback button hands over the whole conversation for 5 years — and it is the one thing that permits training — the exception, in full: what the feedback button actually hands over
• Which contract terms actually govern the working group's accounts — confirming which contract governs each account
• Risk of using personal Pro/Max accounts for org work — the standardization decision this fact supports

On the 7-vs-30-day figure (resolved 2026-07-23): the working group’s own document recorded API log retention as 7 days. Current Anthropic documentation supersedes that — the standard is not retained by default, with a 30-day backend-deletion window where retention applies. Treat the doc’s 7-day figure as outdated. Retention stays compliance-critical: confirm current specifics with your Anthropic contact.

Detail

The API tier has the shortest default window of the product family — and the only fully negotiated one: ZDR, where data isn’t stored beyond real-time safety processing. Two sharp edges to carry with it: custom Skills sit outside ZDR (Agent Skills sit outside ZDR, and outside HIPAA coverage on the API path — Skills carry instructions, never data), and in the HIPAA context, Claude Code is BAA-covered only when ZDR is enabled — and ZDR blocks access to some newer models (Claude Code: not covered on the API path at all; on Enterprise only with ZDR. Cowork: not yet covered). ZDR is an arrangement you verify you actually have, not an assumption.

What “not retained by default” does not cover

Start with the exception the source attaches to the sentence itself, which the earlier version of this node dropped: “Conversation content (your prompts and Claude’s outputs) is not retained by default; the exception is Covered Models, which require 30-day retention.” Covered Models are named: “Claude Fable 5 and Claude Mythos 5 are designated Covered Models … and require 30-day data retention.” So if you are using Fable 5 or Mythos 5, your prompts and outputs are retained for 30 days — mandatorily, not as a fallback. This is the same rule that makes those models unavailable under ZDR, stated from the other side, and it applies to organizations that never considered ZDR at all. The 30-day requirement “applies wherever Covered Models are offered.”

Beyond that, several things sit outside the default, and one of them catches people building real tools:

• Files you upload stay until you delete them. The feature-eligibility table’s Files API row reads: “Files retained until explicitly deleted.” Upload a registry spreadsheet or a scanned document through the Files API and it persists — no 30-day clock, no ZDR coverage (the row is No for both ZDR and HIPAA). If you upload it, you own deleting it. Note the related distinction: HIPAA eligibility “applies to PDFs sent inline through the Messages API, not through the Files API.”
• Stateful features are stateful by design. The documentation is explicit that the “No” rows are not oversights: “Features marked ‘No’ for ZDR are fundamentally stateful: the Batch API stores your jobs, the Files API stores your files, and code execution runs in persistent containers. Data for these features is retained per the feature’s documented policy. Using them is a choice to step outside your ZDR arrangement for that specific data.”
• Flagged content and legal holds are retained regardless of arrangement — see the status note below.
• Other retention models sit alongside this page entirely: Compliance API data “follows its own retention model,” the Activity Feed “retains data for 6 years,” and claude.ai chat, file, and project content “follows your organization’s retention policy” set in Organization settings > Data and privacy — which means part of your retention posture is a control you hold, not one Anthropic sets.
• JSON schemas are cached separately and do not receive PHI protections — Never put PHI in a JSON schema — cached schemas are not protected like message content.

Source & currency

• Source: API and data retention — Claude platform docs + How long do you store my organization’s data? — Anthropic Privacy Center.
• Status: CONFIRMED 2026-07-23 — discrepancy resolved (receipts R20, R21, re-fetched this round): “Conversation content (your prompts and Claude’s outputs) is not retained by default” (R20); “we automatically delete inputs and outputs on our backend within 30 days of receipt or generation” (R21). The source doc’s earlier 7-day figure did not corroborate on any current primary source and is treated as outdated — the current standard is not-retained-by-default with a 30-day backend-deletion window. ZDR is per-organization. Exception (retained regardless of arrangement): flagged content may be retained up to 2 years (safety-classification scores up to 7 years), plus legal holds. Re-verified 2026-07-28 (receipts R36, R38) when the Privacy Center storage article — the page the 30-day deletion figure is quoted from — joined the chip list. Re-verified and expanded 2026-07-25 (receipt R12): the exception surface now also states the Files-API “retained until explicitly deleted” rule, the stateful-features principle in the documentation’s own words, and the retention models that sit outside this page (Compliance API, 6-year Activity Feed, org-configurable claude.ai retention). The Files-API case is the sharpest for this audience — an uploaded registry file has no expiry clock. Corrected the same day, after independent review: the node had quoted “not retained by default” while truncating the clause that follows it in the source — “the exception is Covered Models, which require 30-day retention” — and the Covered-Models rule appeared nowhere else here. That was an understatement of exposure (the only one this round produced) on a node whose title is the reassurance line; the mandatory 30-day retention is now stated first among the exceptions. ⚑ Compliance-critical: confirm current specifics with your Anthropic contact before relying on this for a compliance decision. Teaching aid, not compliance or legal advice.

Related

• Data Retention & What Goes Back — parent domain
• Agent Skills sit outside ZDR, and outside HIPAA coverage on the API path — Skills carry instructions, never data — the ZDR carve-out every Skill builder must know
• Claude Code: not covered on the API path at all; on Enterprise only with ZDR. Cowork: not yet covered — where ZDR meets BAA coverage
• Never put PHI in a JSON schema — cached schemas are not protected like message content — “prompts not stored” does not extend to the cached JSON schema
• ZDR — Zero Data Retention — the term, plainly
• Covered Model — the designation behind the mandatory-retention exception

Detail

Both exclusions, from the same feature-eligibility table. The API documentation’s row for Agent skills marks it No for ZDR and No for HIPAA, with the note: “Skill data retained per standard policy.” The Agent Skills documentation states the ZDR half in its own words: “Agent Skills is not covered by ZDR arrangements. Skill definitions and execution data are retained according to Anthropic’s standard data retention policy.”

Scope note — read this before quoting the HIPAA half. That table’s own lead sentence scopes it: “The following table lists which Claude API features are eligible for ZDR and HIPAA readiness arrangements.” So the No/No is documented for the API path. On the Enterprise path, the authority is a different document, and the Help Center says so plainly: “The Implementation Guide for HIPAA Entities on the Anthropic Trust Center lists every feature’s status and is the authoritative source.” That guide is access-gated, so this curriculum cannot verify the Enterprise-side status of Skills — and will not guess it. If you are on Enterprise with a BAA, check the Implementation Guide for Skills specifically rather than assuming either way.

Why the HIPAA half is the one that was missing here. For a clinician or a registry-handling nonprofit, “is it under our BAA?” is the primary question, and this curriculum previously answered only the ZDR question — leaving a reader with a BAA to assume their Skills were covered by it. On the API path they are documented as not covered. Either way the prudential rule below holds, which is why the design advice does not depend on resolving the Enterprise question.

The design rule this produces — treat it as first-class, not a footnote: a Skill carries instructions, never data. Every use case this curriculum teaches (Skills) is a method — an abstract-review rubric, a house style, an intake template, a minutes format, a lay-summary pattern. None of them needs a real patient record inside the Skill to work. That is fortunate, because the retention posture means anything you did put inside would follow standard retention regardless of your ZDR or HIPAA arrangements. Good Skill design and safe Skill design are the same discipline here.

The group’s own version of this question for the contact: Retention considerations specific to Skills, outside ZDR.

Source & currency

• Source: API and data retention — Claude Docs (the feature-eligibility table: Agent skills — ZDR No, HIPAA No) · Agent Skills overview — Claude Docs §Data retention.
• Status: CONFIRMED — re-verified 2026-07-28 (receipts R12, R15, R35) — both passages verified on the live pages 2026-07-25 and re-verified 2026-07-28 via the Agent Skills overview, whose own Data-retention section states the exclusion (“Agent Skills is not covered by ZDR arrangements”) and which joined the chip list that day. R35 is this node’s 2026-07-28 evidence row; the API-retention page stays the chip for the eligibility-table claim, on its 2026-07-25 verification (receipt R12). Corrected 2026-07-25 on two counts: (1) the node taught only the ZDR carve-out and never stated the HIPAA exclusion, which for a PHI-handling reader is the more consequential of the two; (2) the source: chip pointed at the consumer retention article while the status block cited the API table — the chip now points at the document the claim actually comes from.
• ⚑ Compliance-critical: ZDR and HIPAA coverage boundaries move — confirm current specifics with your Anthropic contact before relying on this for a compliance decision. This is a teaching aid, not compliance or legal advice.

Related

• Data Retention & What Goes Back — parent domain
• API: content not retained by default; inputs/outputs deleted within 30 days; ZDR negotiable — the ZDR baseline this claim carves out of
• Most ways of using Claude are never covered by Anthropic's BAA — and on two cloud platforms, coverage is someone else's call — the coverage picture this exclusion belongs to
• Retention considerations specific to Skills, outside ZDR — the working group’s bring-to-the-call version
• Skills — where Skills are taught (build them data-free)
• ZDR — Zero Data Retention — the arrangement Skills sit outside of
• Skill (Claude) — the term, plainly

Detail

What is retained, and for how long. Identical wording on both the consumer and the commercial retention policies: “Where you have provided feedback to us (e.g. by submitting feedback through our thumbs up/down button or sent bug reports), we retain data associated with that submission for 5 years.” The training article is more specific about the scope: “When you provide us feedback via our thumbs up / down button, we will store the entire related conversation, including any content, custom styles, conversation preferences, or model settings, in our secured back-end for up to 5 years.” Not the rated message — the conversation.

What it permits. This is the sharp edge. Commercial plans are not trained on by default (Commercial plans: no training by default and by contract — the exception is data you hand over yourself) — but: “If you explicitly report feedback or bugs to us (e.g. via our thumbs up/down feedback button), or otherwise choose to allow us to use your data, then we may use your chats and coding sessions to train our models.” The API documentation’s phrase “never used for model training without your express permission” describes exactly this: the feedback button is how express permission is given. A contractual protection you did not know you could waive, waived by a one-click control placed beside every response.

One limit worth knowing, from the same source: “Feedback data does not include raw content from connectors (e.g. Google Drive), including remote and local MCP servers — though data may be included if it’s directly copied into your conversation with Claude.” So a connector-sourced document is not swept in by reference; a passage you pasted is.

Three further properties:

1. Plan-independent. The same 5-year sentence is published for consumer products and for organization data. No plan tier shortens it.
2. Outside the training toggle. The “Help Improve Claude” model-improvement setting (Consumer plans: the training toggle governs training use — up to 5 years if on; no automatic window published if off) does not govern this; turning it off does not turn this off.
3. Deleting the chat does not undo it. Retention attaches to the submission, not to the conversation in your history — and “your data will still be included in model training runs that are already in progress, or in models that have been trained” (Actual timeline to purge data after deletion).

The practical teaching. The feedback buttons are the most frictionless controls in the interface and they carry the longest ordinary-use retention window plus a training permission. The habit to build: if an exchange contains patient information, member-identifying detail, or anything you would not want held for five years or trained on — do not rate it, and do not attach it to a bug report. Reproduce the problem with a sanitized example, then rate that. This is the single highest-value habit in this module, because it is the one place where a careful organization’s contractual protection can be undone by a well-meaning click.

Source & currency

• Source: How long do you store my organization’s data? (commercial) and How long do you store my data? (consumer) — the 5-year sentence is identical on both · Is my data used for model training? — the entire-conversation scope, the training permission, and the connector limit.
• Status: CONFIRMED — re-verified 2026-07-28 (receipts R13, R14, R18, R34, R38, R40) — verified verbatim on all three live pages 2026-07-25 and again 2026-07-28, when all three joined the chip list (the chip previously named one of them). The “identical on both” claim about the 5-year sentence was re-verified byte-for-byte across both storage articles before re-chipping, not assumed. Node created 2026-07-25: the 5-year feedback window was absent from this curriculum entirely. Expanded the same day, after independent review, with the two facts that make it consequential rather than merely notable — the entire conversation is stored, and feedback is the act that permits training on commercial data.
• ⚑ Compliance-critical: confirm current specifics with your Anthropic contact before relying on this for a compliance decision. This is a teaching aid, not compliance or legal advice.

Related

• Data Retention & What Goes Back — parent domain
• Commercial plans: no training by default and by contract — the exception is data you hand over yourself — the training question, which this is independent of
• Consumer plans: the training toggle governs training use — up to 5 years if on; no automatic window published if off — the setting that does not govern this
• Actual timeline to purge data after deletion — the companion lesson: what deletion does and does not reach
• Which contract terms actually govern the working group's accounts — worth raising in the same conversation with your account team