Can ChatGPT edit videos?

Partly. ChatGPT accepts video uploads and can tell you what's in them, but OpenAI warns it may not analyze the whole video or its audio, and it has no timeline to cut on. To edit your own footage reliably, use Codex with ffmpeg and Whisper on your computer, or the invideo plugin if you'd rather work in a hosted editor.

What ChatGPT does with your video

Attach a video and ChatGPT takes it, Free plan included, and can tell you what it finds. OpenAI's own help page is careful about the rest: ChatGPT may use tools on the file, its analysis can be incomplete or inaccurate, and it may not cover the entire video or read the audio correctly.

Three facts decide what you can do with that:

• Upload limits, not a video limit. Video counts toward your plan's file-upload limits. OpenAI gives 512 MB per file as the ceiling for documents and audio, but publishes no separate size figure for video.
• Transcripts come from audio uploads, which need a paid plan. ChatGPT won't take an MP4 that's identified as video as an audio file, so export the soundtrack first with the command below.
• No edited file by design. Data analysis runs Python in a sandbox with no internet access, and OpenAI documents it for tables, charts and calculations. It doesn't document video editing, so treat a trimmed file you coax out of it as luck, not a feature.

ffmpeg -i talk.mp4 -vn -c:a aac talk.m4a

GPTs and plugins

The “video editor” GPTs in search results are ChatGPT with instructions on top. A GPT can use web search, image generation, canvas and Code Interpreter, plus either connected apps or custom actions that call an outside API. It can't do more with your footage than those tools, or the service behind its action, can do.

They're also on the way out. OpenAI has scheduled custom GPTs to retire on December 11, 2026, and is moving creators to plugins.

Plugins are the stronger option. invideo is an official ChatGPT plugin wired to the invideo Editor: the chat opens a project, brings media in, edits the timeline and exports, and you can take over in the editor at any frame. invideo says manual editing there is free, while AI generation such as b-roll and voiceover uses your invideo credits. Runway's plugin is aimed at generating new clips and ads from prompts and images, not at trimming yours.

The reliable route: Codex, ffmpeg and Whisper

For your own footage, the dependable setup runs on your computer. Codex, OpenAI's coding agent, comes with ChatGPT plans and runs commands in your terminal. ffmpeg does every cut, crop and encode on the full file, with no upload cap. Whisper, OpenAI's open-source speech model, gives word-level timestamps, which turns cutting video into editing a transcript.

Codex writes and runs the commands; you approve the plan. Every step writes a new file and leaves your original alone, and the commands stay in the transcript, so the next video is the same edit with a new filename.

Set up once

# Codex CLI
npm install -g @openai/codex

# Whisper (needs Python and ffmpeg on your PATH)
pip install -U openai-whisper

# ffmpeg with the subtitles filter, on macOS
brew install ffmpeg-full
export PATH="$(brew --prefix ffmpeg-full)/bin:$PATH"

On a Mac, plain `brew install ffmpeg` can't burn captions. Homebrew's standard ffmpeg is built without libass, the library behind ffmpeg's subtitles filter, so the caption step fails. Install ffmpeg-full instead. It's keg-only, so put it first on your PATH as above, then check that ffmpeg -filters | grep subtitles prints a line.

1. Start in a copy

   Put the recording in an empty folder, run codex there and keep the original untouched.

2. Transcribe with word timestamps

   Whisper reads the MP4 directly. Word timestamps are marked experimental, so leave a little padding around every cut. Whisper can also leave filler words out, and OpenAI's speech-to-text guide lists keeping them as a job for the prompt, hence an --initial_prompt full of ums.

3. Find the dead air and fillers

   ffmpeg's silencedetect filter prints the start and end of every pause. The transcript shows the ums and restarts. Ask Codex for the full cut list before it cuts anything.

4. Cut

   Codex turns the keep-list into one select and one aselect filter, so picture and sound are cut at the same timestamps.

5. Reframe to 9:16

   A center crop keeps the middle third of a 16:9 frame. That works for a talking head who stays put; it won't follow someone who walks.

6. Add b-roll

   Overlay a second clip for a few seconds while your audio keeps running underneath.

7. Caption last

   Transcribe the finished cut again so the timings match, ask for three words per line and burn the SRT in. Captioning after b-roll keeps the words on top of everything.

The commands Codex will run

# 2. Transcript with word timestamps
whisper talk.mp4 --model turbo --word_timestamps True \
  --initial_prompt "Um, so, uh, I was like, you know, thinking." \
  --output_format json --output_dir work

# 3. Every pause longer than half a second
ffmpeg -i talk.mp4 -af silencedetect=noise=-30dB:d=0.5 -f null -

# 4. Keep 0-4.1 s and 5.9-12.1 s, drop the rest
ffmpeg -i talk.mp4 \
  -vf "select='between(t,0,4.1)+between(t,5.9,12.1)',setpts=N/FRAME_RATE/TB" \
  -af "aselect='between(t,0,4.1)+between(t,5.9,12.1)',asetpts=N/SR/TB" \
  cut.mp4

# 5. Center crop to 1080x1920
ffmpeg -i cut.mp4 -vf "crop=ih*9/16:ih,scale=1080:1920" -c:a copy vertical.mp4

# 6. B-roll from 5 s to 9 s, your audio underneath
ffmpeg -i vertical.mp4 -i broll.mp4 -filter_complex \
  "[1:v]scale=1080:1920:force_original_aspect_ratio=increase,crop=1080:1920,setpts=PTS-STARTPTS+5/TB[b];[0:v][b]overlay=enable='between(t,5,9)':eof_action=pass[v]" \
  -map "[v]" -map 0:a -c:a copy with-broll.mp4

# 7. Captions, three words per line, burned in
whisper with-broll.mp4 --model turbo --word_timestamps True \
  --max_words_per_line 3 --output_format srt --output_dir work
ffmpeg -i with-broll.mp4 \
  -vf "subtitles=work/with-broll.srt:force_style='Fontsize=16,Outline=2,MarginV=60'" \
  -c:a copy final.mp4

You don't need to type any of that. Hand Codex the whole job in one message and let it adapt the numbers to your file:

Working in Claude Code instead? See Claude video editing. Need a video from nothing rather than an edit? See Can ChatGPT make videos?

What this won't do

• Pick your best take. Codex reads the transcript, not your face. Review its cut list; it will happily remove a pause you meant.
• Track a moving subject. The 9:16 crop is fixed. For a speaker who moves, ask for a crop that pans between set points, or reframe that shot in an editor.
• Land every cut on the syllable. Whisper's word timestamps are experimental, so a cut can clip the start of a word. The 0.1 s padding is there for that; widen it if you hear it.
• Judge the grade or the mix. ffmpeg can adjust color and level audio, but deciding whether it looks and sounds right still takes someone watching.