Evidence from the production line.
First-party reports built from server-written production logs and controlled model tests. Methods and failures ship with the headline number.
Published report
Controlled AI video model benchmark · July 2026
One fixed product-shot prompt, four model submissions, three unedited outputs and one provider-side non-start. No retries and no composite ranking.
Inspect the outputs →4Submitted
3Completed
Published report
AI production reliability benchmark · July 2026
251 terminal production requests across image, video, voice, lip-sync and podcast rendering, reported with a minimum per-model sample threshold.
Read the report →251Terminal requests
97.2%Completion rate