Predictions now have a starting status
Oct 16, 2026
·1 minute read
Predictions now have a starting status between pending and started. A prediction is starting when it lands on a cold runtime and the model code and weights are still loading. On a warm runtime, it goes straight from pending to started.
A prediction on a cold privacy filter model, moving from starting to started to succeeded.
What changed
Previously, started covered both setup and inference. On a cold runtime, a prediction could sit in started while the model was still loading, and there was no way to tell whether it was setting up or already running. Now started always means the prediction is actively running.
The move to starting is also sent to your webhook if you've set one. If your code or webhook handler matches on specific statuses, add starting to the ones you treat as in progress. Not every prediction will pass through it.
The latest SDKs include starting in their status types. Older versions keep working, since polling only waits for a terminal status.
See the full prediction lifecycle for every status.