The AI Data Retention Tracker, release one: Anthropic's five retention states, and which cannot run zero-retention

The AI Data Retention Tracker, release one: Anthropic's five retention states, and which cannot run zero-retention

Four of the five model-and-tier cells this tracker's first release could verify from dated primary documents, every one of them an Anthropic commercial surface, retain your prompts by default. The clearest case is also the newest: Claude Fable 5.1, launched September 1, 2026 in GitHub Copilot, is one of only two Claude models in that product that cannot run unless Anthropic retains prompts and outputs, and zero-data-retention (ZDR) access exists only for enterprises approved under a time-bound exemption the changelog dates as running "through the end of the calendar year," meaning December 31, 2026 1.

The simple version of retention is a setting you switch off. Across Anthropic's own commercial surfaces it is currently three different things: a default you must override (Claude Enterprise conversations), a condition of use you accept to run the model at all (Fable 5 and Fable 5.1 in Copilot), and a floor that cannot be lowered (Covered Models under a Business Associate Agreement, held at 30 days). This tracker assembles those states per model and tier, with the date each policy took effect.

Four of five verified cells keep prompts unless someone acts

Method: a cell is a model-and-tier combination with a public, dated policy document from the provider itself, meaning a help center, privacy center, or changelog. A cell counts as retention-by-default if, once the model is in use, no available configuration stops prompts from being stored. Release one, as of September 2026, covers Anthropic's commercial surfaces and maps five cells, one retention state per cell (figures last re-verified September 25, 2026):

• Claude Fable 5.1 in GitHub Copilot (Pro+, Max, Business, and Enterprise plans): retention is required to operate Anthropic's safety classifiers. Copilot Business and Enterprise administrators must switch the model on themselves, and enabling it acknowledges the retention requirement. Retained data is not used to train Anthropic's models 1. Neither that changelog nor GitHub's model-hosting documentation states how long Fable 5.1's retained prompts are kept 12.
• Claude Fable 5 in GitHub Copilot: the same retention requirement, and the launch changelog for Fable 5 states the duration: Anthropic retains prompts and outputs for up to 30 days to operate safety classifiers, then deletes them 3.
• Every other Claude model in GitHub Copilot: operates under zero data retention 1. Claude Opus 5.5, added September 22, 2026, lands in this cell: Anthropic's launch post states "Like previous Opus models, Opus 5.5 is available with zero data retention" 4, and GitHub's model-hosting documentation keeps every Claude model except the two Fables under ZDR 2. The carve-out sits one level down, at features rather than models: the GitHub-Anthropic ZDR agreement covers generally available features, and Anthropic beta or public-preview features, such as tool search via the Messages API, are not covered and may be retained by Anthropic 2.
• Claude Enterprise conversations and projects: data is retained indefinitely by default; an owner can set a custom retention period, with a 30-day minimum counted from last activity 5.
• Anthropic Covered Models on a Business Associate Agreement: 30-day retention is required and ZDR is not available 6.

That is four of the five cells.

ZDR Lag: at least 121 days outside the exemption

The tracker's second metric, ZDR Lag, counts days from a model's launch to the date a typical paying customer can run it with zero retention. Fable 5.1 makes the point sharply. Eligible enterprises get ZDR on day zero: approved customers' requests route to ZDR endpoints under the exemption, and eligibility runs through the GitHub account team rather than a published setting; GitHub Support cannot determine eligibility or enable access 1. The exemption also narrows what an approved enterprise may do: customers who enable a Fable model with ZDR agree to use it only for internal operations, including developing and evaluating products for their own customers, and may not make the model endpoints or outputs available externally 2. Everyone else waits. The exemption covers Fable 5.1 and Fable 5 with zero data retention "through the end of the calendar year," in the changelog's words: December 31, 2026, on a post dated September 1, 2026. After that point, continued use of either model in Copilot requires Enterprise Frontier Safeguards (EFS), whose phase-in is dated only as "later this fall" 7. September 1 to December 31 is 121 days, so an unapproved Copilot Business organization faces ZDR ineligibility spanning at least that 121-day window, with the end point, the EFS rollout, carrying no date. Fable 5's window is longer: it reached Copilot on June 9, 2026 under the same requirement, which makes at least 205 days of ineligibility, June 9 through December 31, for an unapproved organization, a span that includes the three days the model ran before Anthropic suspended it across Copilot on June 12 and its re-enablement on July 1 31. The metric cuts both ways, and September 22 proved it: Claude Opus 5.5 arrived in Copilot with zero retention from day one, no exemption required 4. EFS itself is pitched as "complete privacy (the same as a zero data retention policy)," with data stored in cloud infrastructure controlled entirely by the customer while automated safety monitoring continues 71.

The dated record, assembled from the provider's dated primary documents:

• October 13, 2025: the most recent systematic academic comparison we could locate, an arXiv study of ZDR architecture in Salesforce AgentForce and Microsoft Copilot 8.
• March 16, 2026: Anthropic ships custom retention controls for Enterprise, with an indefinite default and a 30-day floor 5.
• June 9, 2026: Fable 5 reaches general availability in Copilot with the retention requirement; GitHub's changelog states retained prompts and outputs are kept for up to 30 days, then deleted. The model is suspended across Copilot on June 12 and re-enabled July 1 3.
• September 1, 2026: Fable 5.1 reaches general availability in Copilot, retention attached 1.
• September 22, 2026: Claude Opus 5.5 is added to Copilot and runs with zero data retention from day one, no exemption required 4.
• September 24, 2026: GitHub announces a global default policy for generally available Copilot features, effective October 22, 2026: unconfigured eligible features follow the administrator's chosen default, and explicit decisions are preserved 9. The Opus 5.5 changelog states that under default model enablement, new models are automatically enabled unless the global default is off or the model is explicitly disabled 10. Neither changelog says whether the Fable retention policy counts as an eligible feature under the new default, so the enablement mechanics this tracker documents stand as written on September 1.
• End of the 2026 calendar year: the exemption expires and EFS becomes the stated condition of continued Fable use 1.

The cells where zero-retention is not purchasable

Anthropic's BAA terms state it flatly: Covered Models "require 30-day data retention and aren't available with zero data retention (ZDR) enabled" 6. The same document says Claude Code is covered under the BAA only when ZDR is enabled, and draws the consequence itself: those services cannot use Covered Models under the BAA. Six Claude Code surfaces are listed as incompatible with ZDR altogether: desktop remote mode, plus the web, Code Review, Code Security, Computer Use, and Remote Control betas 6. A HIPAA customer therefore faces a grid with no fully covered corner: chat on a Covered Model means 30-day retention with no ZDR, Claude Code means ZDR or no BAA coverage, and the remote and web Claude Code surfaces qualify for neither.

Retention and training are separate rows, and the difference matters. GitHub's changelog states that data retained for Fable 5.1's safety classifiers is not used to train Anthropic's models 1. A prompt can be stored and still never become a training example, so the tracker scores the two axes independently.

The questions worth asking any vendor, in order: what is retained by default, and is ZDR purchasable for this model at all? The first answer lives in the provider's own docs, dated above. The second, for Fable 5 and 5.1 outside the exemption and for every Covered Model under a BAA, is currently no. Opus 5.5, added September 22, is the counterexample that keeps the question honest: there the answer was yes on day one 4.