The AI editing agent
Press ⌥⌘A (or the Agent button in the rail) and describe what you want. The agent isn't a separate "AI mode" — it operates your real timeline with the same operations you use by hand, so its work is inspectable and every change is one ⌘Z from gone.
What it can do
- Cut and arrange — add, split, move, retime, and ripple-edit clips; build side-by-sides, picture-in-picture, and grids; sync clips by audio.
- See your footage — it inspects frames and transcripts, and searches media semantically: "find the shot where the bike jumps" searches what's in the video, both visual and spoken.
- Transcript edits — remove filler words and silences, down to the word.
- Captions and titles — transcribe and lay down styled caption tracks, add and update text elements.
- Color and effects — apply grades, LUTs, and effects, read the scopes, denoise audio.
- Multicam — build synced groups and switch angles.
- Generate media — create video, images, voiceover, music, and sound effects with the same models as Studio, imported straight into the project.
- Project housekeeping — organize the media panel, change project settings, run exports.
What it costs
The agent runs on Claude and is metered by tokens, converted to credits. It defaults to Sonnet — the reliability sweet spot for editing — and the model picker in the agent panel lets you drop to Haiku for cheap quick tasks or step up to Opus for the hardest edits:
| Model | Input (per M tokens) | Output (per M tokens) |
|---|---|---|
| Haiku | 150 credits | 750 credits |
| Sonnet (default) | 450 credits | 2,250 credits |
| Opus | 750 credits | 3,750 credits |
Minimum 1 credit per turn. A typical editing exchange costs a handful of credits; generation the agent triggers is billed like any other generation.
Working style that works
Give it outcomes, not clicks: "make this feel faster and add captions for social" beats a list of micro-instructions. Check the result, then refine — it keeps the context of the conversation and the cut.