GPT Image 2.5 is live — OpenAI's newest image model, targeted edits that leave the rest of the frame alone
All resources

· reAPI Team

API data retention: save outputs before links expire

Plan API data retention for generated files, input uploads and backups. Save outputs before links expire, verify copies, and check what deletion removes.

API data retention: save outputs before links expire

The video finished successfully. Your database contains the task ID and a download URL. That is enough to show where the result was, but it is not an archive of the video.

Replicate's documented default makes the distinction urgent: API prediction inputs, output values, output files and logs are removed after an hour. Predictions created on its website follow a different rule and remain until manually deleted.[1] A file you can still see from a web session does not establish how long an API result will remain downloadable.

For an output you need later, the completion workflow should copy the actual bytes into storage you control and verify that copy. Deletion then needs to account for both the provider's version and the copies you created. Follow one file through those steps, and the retention questions become concrete.

Give the download its own completion state

Consider a hypothetical generated video. The generation completes, and an archival worker begins copying the output. Keep “generation completed” separate from “copy verified.” If the download fails, the first state can be true while the second remains false.

Read the completed task's documented output fields and identify each file you intend to retain. For a task with several outputs, save the output index or equivalent identity alongside the task ID. A single success flag cannot show which of several files was copied.

Download through your application's authorized access path, write the destination object and check it before presenting it as durably saved. Confirm the expected object exists, the transfer completed and its size and media type are plausible. Open or decode the owned copy where practical. A checksum can verify consistency with the bytes downloaded; it cannot prove that the generated scene is correct or that the video meets the customer's request.

Keep the original URL, task ID, output identity, destination object key and copy time together. Use your own object identity for later application access. Saving the temporary provider URL in another database column does not create an independent copy.

If the same completion notification arrives again, use the task and output identity to recognize an already verified copy. Retry an incomplete transfer as appropriate, but avoid producing unexplained duplicates merely because the notification was repeated. This is a proposed application workflow, not a claim that every provider implements your archive for you.

The worker can fail after the model succeeds

In the video example, suppose the archival worker is unavailable when the generation finishes. The copy should remain pending. When the worker returns, it can attempt retrieval while the source is still available. If the source has expired, record the preservation failure instead of silently marking the task archived.

Do not regenerate the video and describe that as recovery of the original file. Another generation is another operation and may produce different content or incur another charge. First check for an existing verified copy or another retrieval path documented by the provider. fal, for example, says expired CDN files are permanently deleted and cannot be recovered.[2]

A retention window is therefore an input to the design of the copy job. If the application cannot reliably archive within the applicable window, change that workflow or arrange documented storage before promising later access. A link that happened to work longer in one observation is not a storage guarantee.

reAPI's task documentation also advises copying generated outputs promptly for long-term archival. It describes rehosted CDN outputs and says lifecycle policies are configured per deployment rather than guaranteed by that contract.[3] That supports saving your own copy; it does not establish a fixed expiry date or permanent hosting for every returned URL.

For a recovery check, use an authorized synthetic asset and read the destination copy without relying on the original provider link. Keep the result of that check. An upload reported as successful is less useful than confirmation that the application can retrieve and use the owned object.

A request record and a media file can have different lifetimes

The hypothetical video's task can remain visible while its file is no longer available, depending on the service's rules. Conversely, removing request content need not remove every media object. Name the object before assigning a retention period to it.

fal documents request inputs and outputs in history as retained for 90 days by default. Its CDN files have separate configurable retention. X-Fal-Store-IO: 0 controls request-history payload storage; X-Fal-Object-Lifecycle-Preference controls media lifetime, with per-request settings taking precedence over account defaults.[2] The history setting alone does not switch off all storage.

An uploaded input adds another object. Your source image might remain in your bucket while a provider copy is used to generate the video. Keep the relationship between those copies if they need later removal. The output video's lifetime does not establish the lifetime of its reference image.

Collections and shared views are another layer. fal says enabling Assets, putting a file in a collection or sharing that collection does not extend the underlying file's lifetime.[4] A saved library relationship can outlive the usable media it points to.

The same reasoning applies to document workflows. Reusable PDF annotations can contain extracted content, not just references to a file. If your application stores them, include that copy in its retention decisions alongside the original document.

Deleting the library item is not the same as deleting the file

Suppose the video is now archived in your bucket and you want to remove the provider-side version. Choose the operation that matches that goal. Hiding it from a library, removing a collection membership and deleting the CDN file are different actions.

fal's Assets documentation says deleting an asset removes it from the library and search but does not itself delete the underlying CDN file. Removing it from a collection is narrower still: the asset can remain in the library and other collections.[4] Checking only the collection after removal would verify the view, not the file's deletion.

For a completed generation, fal documents a separate request-payload deletion operation that also targets the CDN output files. It does not remove CDN input files, which may be reused by other requests.[2] If the reference image also needs removal, handle it through the appropriate file operation after checking its intended use elsewhere.

Inspect the result at file level. fal's deletion response contains cdn_delete_results, with the targeted link and an exception value for each file; null indicates success for that entry.[5] An overall response is not a reason to ignore a reported exception for one of several outputs.

Record the operation time, targets and unresolved failures. A subsequent failed download is evidence about that retrieval path, but it does not prove every internal backup or downstream copy has vanished. Keep the statement of deletion within the documented operation's scope and the results you actually checked.

Replicate's website Delete action has its own documented scope: it removes the prediction and associated output data and files.[1] Use the rule for the creation path involved rather than assuming every dashboard's delete button behaves like fal's library action.

Follow the copy into exports and backups

Removing the provider version does not remove the video now in your bucket. If the customer downloaded an export, that is another copy. Revoking a sharing link stops access through that link, but does not retrieve files recipients already downloaded; fal's sharing guidance explicitly distinguishes those outcomes.[4]

For the copies you control, retain the task and output identity, owner, storage location and removal rule with the object. This makes a later deletion request actionable without depending on a temporary URL or an unlabeled export filename. Include generated previews, stored versions and migration copies where your workflow creates them.

Backups need an explicit decision too. Keep them for the recovery purpose you intend, and account for how restored data is reconciled with later deletions. Otherwise, restoring an older backup can reintroduce content the application had already removed. During a storage migration, verify the destination before switching references, then carry the ownership and retention decision with the bytes.

Those originals, previews and backups also occupy your destination storage. Include the number of copies and how long each remains when estimating that storage cost; deleting the provider's version does not eliminate the cost of your own archive.

There is no universal seven-year retention requirement in the generated-output contracts discussed here. Different record types can have different requirements; a period applied to an invoice should not automatically become the lifetime of every generated video. Define the applicable policy for the object and purpose rather than selecting a familiar number.

Where is the data stored? In this example, there can be provider request history, CDN media, your owned object, exports and backups. Geographic residency is a separate contract question. A media URL does not identify where every copy is processed or stored.

Keep storage, sharing and training questions separate

A statement that data is not used for training does not establish that it is never stored. OpenAI's API data guide distinguishes abuse-monitoring logs from application state. Its Zero Data Retention controls are available to eligible approved customers and remain subject to endpoint and feature limitations.[6] Confirm the scope applying to the actual organization, project and operation.

fal likewise says its request-history and CDN settings do not configure a deployment owner's application logging or application-managed persistent storage. It asks users requiring ZDR to confirm endpoint coverage.[2] One header cannot describe every copy in a workflow that also uses an application, tools and your own storage.

For the hypothetical video, the final record should now answer two practical questions: which verified copy will serve future downloads, and which other copies were removed or remain under a stated rule? Keep a failed transfer or per-file deletion exception visible until resolved. That lets the application distinguish a completed generation, a usable archive and a completed removal without pretending they are the same event.

References

  1. Replicate, “Data retention,” official predictions documentation. Read November 16, 2026; API defaults and website deletion have different scopes.
  2. fal, “Data retention and storage,” official Model APIs documentation. Read November 16, 2026; request-history, CDN and deployment-owner responsibilities are distinct.
  3. reAPI, Tasks documentation, output shape and archival guidance. Reviewed November 16, 2026.
  4. fal, “Retention and deletion,” official Assets documentation. Read November 16, 2026; library actions, file lifetime and downloaded copies.
  5. fal, “Delete request payloads,” official Platform API reference. Read November 16, 2026; per-file deletion results. No deletion was performed for this guide.
  6. OpenAI, Your data. Read November 16, 2026; training policy, retention controls and endpoint limitations.
All resources