Skip to main content

New AI Models, Faster Onboarding, and a Smarter Video Editor

· 2 min read

Setup is now a guided checklist, a new Spend Guard keeps run costs under control, computers that go offline mid-run now fail over automatically, and video, image, and voice tools got a round of model and reliability fixes.

This batch brings a much smoother computer setup experience, a new Spend Guard to keep run costs in check, more resilient handling when a computer goes offline mid-run, and a round of fixes across video, image, and voice tools.

Faster, guided setup​

  • A new setup checklist walks you through hardware checks, tool installs, and sign-ins, with automatic retries for stalled installs, hidden items for unsupported hardware, and live progress tracking.
  • Fixed stale setup reports that could cause failed setups or duplicate checklist cards, and cleaned up chat bubbles during computer hand-offs.

Spend Guard​

  • New Spend Guard lets you set a spend limit per run, with automatic pausing, in-app and email alerts, and a raised default limit of $4 per run to better support longer tasks.
  • Fixed a race condition where responding to a spend-limit pause too quickly could fail to register.

Deployment risk controls​

  • Deployment cards now support the full five-level Risk Tolerance scale, up to "Very High," plus custom routing rules for more granular control over what gets auto-deployed.

Staying resilient when a computer goes offline​

  • Runs affected by a lost or disconnected computer are now automatically detected, paused, and failed over to another machine, with clear notifications to the run owner instead of runs silently stalling.

AI models, video, and voice​

  • Video and image tools now automatically hide retired models and clearly show your platform default, preventing generation failures from sunset models like Sora 2.
  • Added masked inpaint editing for OpenAI GPT Image models, and fixed a bug where confirming a scene regeneration in the video editor could loop indefinitely.
  • Fixed voice replies ignoring your selected model and cleaned up tool-call rendering so saved chat cards display correctly.

Also this week​

  • Moved calendar settings into the main toolbar for quicker access, and redesigned the chat composer's voice and send button.
  • Fixed sign-in window issues on Windows and improved terminal reliability for OpenCode sessions.

Plus 8 smaller improvements.