Release roundup—March 2nd, 2026

Descript has a new look. We used the refresh to audit and improve contrast across the interface—the previous palette wasn't fully WCAG 2.0 compliant in a number of places, and the new one is. That means better legibility for everyone, including users with color vision deficiencies. If you're seeing anything that still looks off, flag it—there are a few known issues still being wrapped up.

Also, probably not important for you but maybe: things are a bit redder now and our brand colors have been updated across the board.

The wait is over, Descript has added the Media Library. You can now upload and organize media at the drive level, making it easier to reuse files across multiple projects, saving you media minutes, as well as time waiting for transcription, and file uploads.

Click Media Library in the left sidebar of your home page. Any media you upload here is shared with everyone on your drive. When you add media from the library to a project, it uses media minutes for transcription. This has been a longstanding feature request, and we’re excited to finally bring it to life.

Import from other projects

You can now preview audio and video before Underlord applies them to your project. When generating media like avatars or video clips, Underlord will show you preview options so you can confirm the results match what you're looking for before committing to the generation.

This helps you avoid wasting time and credits on results that don't meet your needs, especially for time-consuming or resource-intensive generations.

We've improved how layouts work in your projects. When you apply a layout, the editor now better understands which layers should be replaced and which should stay.

These changes make layouts behave more consistently and give you better control over your scenes.

Two new models are now available in Underlord's model picker.

sits between Sonnet 4.5 and Opus in terms of quality and cost. It's Underlord's most precise instruction-follower, making it a strong choice for complex, multi-step editing workflows, creative work built from existing clips, and script-based generation. It had the fewest errors across all test cases. Note that it's the slowest model in the lineup, and for audio-heavy tasks like Studio Sound, Opus remains the better option.

sits between Claude Sonnet 4.5 and Sonnet 4.6 in quality and cost. It's a cost-effective option for longer projects and everyday editing tasks, and leads the lineup on volume leveling. For tasks like removing filler words, pauses, or generating Studio Sound, Claude models will produce cleaner results.

To switch models, open the model picker in Underlord’s settings by clicking the cube icon.

We've added Nano Banana 2 to our image generation model options. This model is built on improved architecture and offers better performance across all aspects of image generation.

You can select this in the Settings menu, or under the cube icon within the generate an image section of the AI tools sidebar.

Rooms now alerts you when your CPU is overloaded during a call. When this happens, you'll see a warning with suggestions to improve performance, like turning off background noise reduction or disabling your video.

Background noise reduction is now easier to access with a toggle in the microphone menu. If your CPU gets overloaded, you'll see a warning, and the feature will automatically disable itself to prevent audio issues instead of crashing.