Skip to main content

Smoother video narration, sturdier voice and a roomier chat

· 2 min read

Video narration now fits each scene and can use one project-wide voice, voice chats recover from failures, and the chat and public page refresh are clearer.

Video scenes now stretch to fit their full narration, and you can cast one narrator voice for a whole project. Voice conversations recover on their own when something goes wrong, and the coding-agent chat has more room and fewer layout jumps. Public page refreshes also report failures clearly.

Video​

  • Scenes now run as long as their narration needs, so voiceovers are no longer cut off. Pictures hold their final frame, and the voice stays consistent from clip to clip.
  • Cast a single narrator voice for an entire video project instead of setting it scene by scene.
  • Style reference stills are no longer used in video generation, which stops characters blending together and drifting between scenes.
  • Deleted styles no longer reappear when an old browser tab is still open.
  • The storyboard is tidier: unused style stills are gone and the "Add character" controls now sit at the top.
  • Long video renders are more stable and no longer compete with other work for resources.

Voice​

  • Voice conversations switch to a backup model when the main one fails, hits a rate limit or has an outage. Stray invalid characters that could break a session over and over are now cleaned up.
  • Asking to switch between email and notifications in the inbox now changes the view as you expect.

Coding agent chat​

  • The chat area is wider, session chips are sized better, and the coding-agent queue now sits above the message box so it no longer shifts your typing controls.
  • Finished coding-agent sessions can be hidden from the chat strip. Your choice is remembered per browser profile.

Pages and billing​

  • Public page refreshes now show a clear failure message and reason, even when the page already has content. A failed check is no longer reported as a success.
  • Image costs for gpt-image-2 auto-quality requests are now calculated correctly instead of using an incorrect flat rate.

Plus 2 smaller improvements.