Replicate deletes API outputs after an hour: what Sume keeps

How long does Replicate keep API prediction output?

One hour by default. Replicate's data retention page says that for predictions created through the API, all input parameters, output values, output files and logs are automatically removed after an hour. Predictions made in the web interface are kept indefinitely, and you can delete any prediction manually from the dashboard.

Sume works the other way for generated media: completed jobs return public media.sume.com artifact URLs, and Sume's run documentation says those URLs do not expire. Both are design choices, and each one decides where your pipeline has to be careful.

What does the one-hour window change in your code?

Replicate's prediction object includes a data_removed field, and its reference says that after removal the output key persists but returns null. Its webhooks page frames webhooks as the way to persist data before automatic deletion, since a webhook delivers the metadata for a completed prediction so you can save output files to your own storage.

The practical rule on Replicate: treat the output URL as a transfer handle, not a storage location. Download or copy the file as soon as the prediction succeeds, and poll with the clock in mind, because a slow consumer that returns after an hour finds nothing.

What does Sume return, and what is the catch?

A completed Sume job exposes artifacts with url, type and content_type, as in Jobs and results. The run receipt docs state that media URLs are durable media.sume.com HTTPS URLs that do not expire and are public to anyone holding the URL. See Run receipts.

The catch is the other half of that sentence. A durable, public URL is not an access-controlled one. If your product needs per-customer access, proxy or copy the file into your own storage and serve it from there. Sume does not document a private or signed-URL mode for these artifact URLs in the pages above, so do not assume one.

Comparison at a glance

Retention facts are from each vendor's own pages.

Which should you plan around?

If you cache outputs in your own bucket anyway, the difference is small: save on completion and move on. The difference shows up in three places.

• Retries and slow consumers: on Replicate a queue backlog can outlive the output; on Sume a late reader still gets the file.
• Audit trails: a Sume job id keeps resolving to its result and its usage rows; a Replicate record loses its content after the window.
• Privacy: short retention is a feature if you do not want vendor-side copies. Sume says nothing on that page about deleting a URL on request, so ask support before promising a customer erasure.