2026-10-29

VibeVideoBench: agents must deliver an MP4 with sound within one hour across 239 tasks. The best system meets all core requirements on 76.3%; counting missing videos as zero changes rankings. Read it for rendering-time planning, not just visual quality. First submitted 2026-10-26. → VibeVideoBench: evaluating finished video delivery (source)

Reasoning effort: on easy tasks solved at every setting, top effort costs up to 42x as much as the lowest. Starting low and escalating after verified failure matches or improves performance in the tested settings. Read it for the dependence on a reliable failure check; the abstract leaves that mechanism underspecified. First submitted 2026-10-25. → Reasoning effort: cost and task-level reliability (source)

• CapCut's 2026-10-28 guide connects Seedance 2.5 drafts with timeline editing for captions, music and pacing. This is a documented workflow, not a verified new model launch; output quality and account-specific export limits remain untested. → CapCut (source)

• CapCut is a new entity for review, scoped to the vendor guide. (source)
• Gemini 4 Carbon and both paper notes add specifications, methods and evidence limits. (model) (video) (effort)

• Gemini 4 Carbon records planned introductory $2/M input · $10/M output, rising to $4/M input · $20/M output after an unspecified introductory period. These are announced rates, not proof of public API availability. (source)