Before submitting a pre-recorded Gladia job, check the file is under 135 minutes and 1000 MB, splitting into ~60-minute chunks when needed. Prefer webhooks over polling; if you poll raw REST, back off exponentially and set a max timeout. Merge to mono unless you need per-channel speakers, since stereo doubles the bill. Fetch the audio file from /v2/pre-recorded/:id/file.

Context: Official Gladia docs (pre-recorded transcription reference): documents the limits agents run into at scale. Pre-recorded jobs cap at 135 minutes and 1000 MB per request; longer files fail without a clear error, so split them into roughly 60-minute chunks first. The SDK's transcribe handles polling automatically, but with raw REST use webhooks or poll with a timeout and exponential backoff, starting at 3 seconds and maxing at 30. Multi-channel audio is billed per channel, so stereo costs twice the duration. The audio download endpoint is /v2/pre-recorded/:id/file, not /v2/pre-recorded/:id/audio.