We gave two AI editing agents the same 80-second clip and the same prompt. In the clip, our founder Allen speaks ten edit orders out loud, from "an editorial background" to "when I snap, go back to normal," and each agent had to find the orders in the audio and carry them out in sync. The two finished edits were labelled Gemini 4 Argon and Claude Opus 5.5.
The labels do not tell the whole story. Codex, OpenAI's coding agent, made the Argon-labelled entry as an Argon-style interpretation. Opus 5.5 designed the other edit, ran out of session before it could export, and Codex finished that saved design. So this was a comparison of two finished edits, not a clean model benchmark, and we credited it that way when we published it.
Below: who made what, what each edit did with the same orders, the timing mistake worth learning from, where to run each model, and a full guide with the prompt to run an AI edit of your own footage, alone or as a fair head-to-head.
Short answer: Our test does not prove whether Gemini 4 Argon or Claude Opus 5.5 is the better video editor, because Codex made the Argon-labelled edit and finished the Opus design. To compare models fairly, give each agent the same raw cut and the same prompt in a fresh session, and write down which model really made each edit.
- Opus 5.5 designed every effect in its edit. Codex restored one missing caption function from Opus's own code, moved one label and rendered it, with no new effect designs.
- Google says Gemini 4 Argon leads on understanding long video (91.7% on LVBench). Understanding a video and editing one are different jobs.
- The most useful lesson came from the snap: the Opus-designed reset landed about half a second after the real finger snap.
- The full prompt is below, with personal paths and settings taken out, plus the steps to record, cut, edit and review.
How the test was set up
The clip was recorded on 14 October 2026 in horizontal 4K at 30 frames per second. The raw cut runs 80.2 seconds across 13 cuts, with no effects, grade or music. The orders were improvised. In order, the clip asks for:
- An editorial background.
- A Vox-style look.
- Paper-cut effects.
- A new caption style, then another, then another.
- Back to normal, then motion graphics where he points.
- A glowing outline around him, like a hologram.
- A more modern style for the whole video.
- A look like it was recorded in the 1950s.
- A split screen comparing the modern and the old look, with motion graphics.
- "When I snap, go back to normal."
The plan was to paste one prompt, word for word, into Google's Antigravity with Gemini 4 Argon and into Claude Code with Opus 5.5, changing only the output folder names. It did not run that way.
Who actually made each edit
- The Argon-labelled edit was produced in Codex as an Argon-inspired entry. Its own working notes say it must not be presented as an actual Gemini benchmark. We rejected its first pass, and Codex rebuilt it in a bolder, high-impact direction.
- The Opus-labelled edit was designed in an Opus 5.5 session in Claude Code. The effects, camera moves, timings, fonts, logos, captions and sound design all came from that session. It hit the session limit before the edit was finished. Codex resumed the saved work, reused Opus's transcript, cleaned voice, cut-out masks and face and hand tracking, restored one missing caption function from Opus's own earlier code, moved a caption label that covered the words, and rendered it.
Both edits went into the same white layout with rounded windows and got the same light skin retouch. Our public credit for the comparison reads:
"Edit credit: the Argon-labeled version was made in Codex. The Opus design was finished in Codex after its session ended. Judge the edits here; this is not a clean model benchmark."
What each edit did with the same orders
| Order | Opus 5.5 design | Codex "Argon" entry |
|---|---|---|
| Editorial background | Newspaper masthead, a typewriter quote, paper background, the speaker cut out in front | An oversized magazine cover with a heavy black serif masthead |
| Vox style | Yellow circle, bold editorial copy, a diagram arrow, paper texture | A hard yellow takeover with a huge EDIT behind him and diagonal ink shapes |
| Paper cut | Torn-paper layers, stop-motion letter tiles, a white border and drop shadow around him | Torn edges, stacked paper shadows and offset layers of him |
| Three caption styles | Outlined single words, then stacked paper word slips, then outline-to-solid karaoke words | Three different animated caption styles, held long enough to read |
| Motion graphics | Graphics tracked to his fingertip, with orbiting shapes and lines | A rotating 3D cube and a spark at the two spots he points to |
| Hologram | A drawn-on glow, a cyan and magenta rim, scan lines and rising particles | A cyan scanning rebuild of him, a depth grid and scan-line interference |
| Modern | A rounded camera frame with a soft background blur | A technology poster with giant NOW lettering |
| 1950s | Black and white film, grain, gate weave and scratches | A black and white TV set, halftone type and thin, old-radio sound |
| Snap back | Flash, a short camera shake and a reset | One clean cut back to normal on the snap |
The two editors took the same words in very different directions. "Make it look more modern" became a subtle camera frame in one and a full poster in the other. If you care what "modern" means, say it in the order.
The snap: the mistake worth learning from
When we later built a clean package for a proper re-run, we checked the snap frame by frame. The snap's sound spike and the visible finger release land within a frame of each other. The Opus-designed effect map put the reset about half a second later, roughly 15 frames at 30 frames per second.
The notes do not say why it drifted. A transcript gives the agent word times, and a gesture needs its own check against the sound and the frames. That is why the prompt below says "Check the frames, not only the words." In your own orders, name the trigger outright: "reset on the sound of the snap."
Where to run each model in October 2026
| Model | Where you use it | What you need |
|---|---|---|
| Claude Opus 5.5 | Claude Code, in a terminal or the Code tab of the Claude desktop app | Claude Pro ($20 a month, or $17 billed yearly) or higher. API: $4 per million input tokens, $20 per million output. |
| GPT-6 Astra | Codex and ChatGPT | ChatGPT Plus ($20 a month), Pro, Business or Enterprise. |
| Gemini 4 Argon | Google Antigravity and the Gemini app | Google AI Pro or Ultra (Ultra starts at $100 a month). Antigravity gives Argon its own weekly quota. API: $2 and $10 per million tokens at the intro price, then $4 and $20. |
Google opened Argon in the Gemini app to AI Pro and Ultra subscribers on 13 October 2026, and in Antigravity to AI Pro and Ultra in the days after. Our clip was recorded on 14 October, in the middle of that rollout.
What the model makers say
None of these numbers come from our test. They are each company's own claims.
- Google on Gemini 4 Argon: a frontier model for real-world software engineering, with an output limit raised to 1 million tokens from 64,000. Google reports 91.7% on LVBench, a long-video understanding test, and says it can "identify details from long videos." It also reports 77.9% on DeepSWE v1.1.
- Anthropic on Claude Opus 5.5: released 22 September 2026 and "particularly good at long and sprawling jobs." One early tester says it stayed on an unattended engineering task for over 18 hours. Our session still ran out mid-edit, so plan limits matter as much as the model.
- OpenAI on GPT-6 Astra: announced 4 September 2026. In Codex it has an experimental way to keep notes across context windows on long jobs, and a full video edit is a long job.
Understanding a long video is a different skill from building one. An editing agent has to plan effects, write the code that draws them, cut you out of the background and render without breaking sync, so we judge these models on finished edits.
Run an AI edit of your own footage
You can use the same prompt to have one agent edit your own clip, or run two agents against each other. Before you start, you need:
- An agent that can run programs on your computer: Claude Code, Codex or Antigravity. A chat window that can only describe an edit will not work, because the prompt asks the agent to transcribe, render and check files.
- FFmpeg, the free tool that cuts and encodes video.
- Local Whisper for transcription: OpenAI's open-source Whisper (MIT licence) installs with
pip install -U openai-whisperand needs FFmpeg. Its default turbo model wants about 6 GB of video memory; the large model about 10 GB. - A graphics card helps. Cutting you out of the background means a mask for every frame, and our 80-second clip at 30 frames per second needed 2,406 of them.
- Disk space. Our 80-second 4K raw cut alone was about 412 MB, before any working files.
You do not have to install the tools yourself. Ask the agent to install whatever the prompt needs before it starts.
Stage 1: Record the clip
- Film horizontally on a steady camera, in even light. Ours was 4K at 30 frames per second; 1080p works too and is lighter to process.
- Say each order as its own short sentence, then hold still and stay quiet for a moment. Our raw cut kept natural holds of 1.5 to 2 seconds after each order, so each effect has room to play.
- Make gestures big and clear. Point to where graphics should appear. If a snap is a cue, make it loud.
- End each effect out loud with "back to normal" or "remove it." The prompt tells the agent an effect stays on until you say so.
- Open with an intro line and close with a call to action. The prompt treats your first and last lines as those and asks the agent to make them look good.
- If you stumble, say the whole line again. False starts are cut out in the next stage.
Stage 2: Make the raw cut
The editor should get one clean file: your best take of each line, joined with plain cuts. This keeps the edit fast, and in a head-to-head it guarantees both editors get identical footage. Paste this into Claude Code with your paths:
Make a raw cut of [path to my recording].
Transcribe it with local Whisper, with word timestamps.
For each line, keep my best full take and drop false starts, repeats and filler.
Cut only in silence and keep the pause after each edit order.
Add nothing else: no grade, no effects, no music.
Save it as [output path] and list every cut with its source times.You should get a transcript, a list of cuts and the file. Watch it once before you move on. Every order should be there, whole, with its pause after it.
Stage 3: Run the edit
- Make three folders: one for the raw cut, one for the finished edit, one for scratch files.
- Open a fresh session in your agent and pick the model by name. In Claude Code, type
/model. Take a screenshot of the picker if you are running a test. - Paste the prompt below and fill in INPUT, OUTPUT and SCRATCH. If only one agent is editing, use the solo changes listed after the prompt.
- Let it work. It transcribes the clip, finds each order and its time, plans the effects, cuts you out of the background, builds the graphics and sound, renders, checks the frames and writes EDIT-NOTES.md. Claude Code sessions usually start in auto mode, where a safety classifier reviews each command, so a long edit can run on its own. Press Shift+Tab for manual mode if you want to approve each step.
- If the session runs out, open a new one and say: "Resume this edit from the files in SCRATCH. Reuse the transcript, masks and tracking. Do not regenerate them." That is how the Opus design was finished without redoing its prep.
Edit my video. This is a test: two AI agents each get this exact prompt and the same clip, and people vote on which edit is better. Do your best work.
INPUT: [the path to your raw video]
OUTPUT: [where to save the finished video]
SCRATCH: [a folder for your working files]
START FROM ZERO
- This is a fresh edit. Open only the input video, your own output folder and your own scratch folder.
- Do not open any other edit, plan, script, prompt, notes or old chat log. That includes the other AI's folders.
THE JOB
- My spoken lines are edit orders. Transcribe the clip with local Whisper (word timestamps), find every order, and do it the moment I say it. Sync it to what I do in the shot (a point, a snap). Check the frames, not only the words.
- An effect stays until I say "remove it" or "back to normal".
- My first lines are the intro and my last line is the call to action. Make them look great too.
MAKE IT CRAZY
- Big, bold, surprising. Something new should happen every second or two.
- Text and graphics: clean black and white, bold sans. Colour only where an order asks for it.
- No italic words. No drawn stand-ins for logos: use the real ones.
COLOUR: VERY LIGHT, MAKE ME LOOK GOOD
- Grade very lightly. Keep it natural, with healthy, warm skin.
- [Optional: lighting to keep, for example "Keep the coloured light on one side of my face. It is part of my look."]
- Do not push contrast or saturation. Do not crush the shadows.
- No skin smoothing, beauty filter or vignette.
TEXT AND SOUND
- Captions: short phrases, bold sans, never on my face or hands.
- The final video stacks your edit above the other AI's edit on a phone screen, so each edit shows at about half size. Make all text big enough to read at that size.
- Clean my voice. Add a music bed and short sound effects, all well under my voice.
HARD RULES
- Same length and timing as the input. No cuts, no speed change. Frame 1 of yours = frame 1 of mine.
- Same resolution and frame rate as the input. H.264, audio at -14 LUFS, peaks under -1 dBTP. One final encode.
- Free local tools only. No paid APIs.
- Run heavy jobs (renders, encodes, transcription, cut-outs) one at a time.
BEFORE YOU FINISH
- Check every frame of your output. Look for: an effect early or late, broken cut-out edges, text on my face or hands, typos, blank frames, sound out of sync. Fix them, render again, check again, until you find nothing.
- Write EDIT-NOTES.md next to your output: each order, its time, what you did, and the tools you used.
Work end to end without asking me questions.For a solo edit, delete the first paragraph's test sentence and the "That includes the other AI's folders" line, and change the stacking line to: "People will watch this full screen on a phone. Make all text big enough to read there." For a calmer business video, swap the MAKE IT CRAZY block for your brand colours and "one effect at a time, nothing that distracts from what I am saying." A single-editor version of this prompt, with a copy button, is on our free prompt page.
Stage 4: Review the edit and send fixes
- Read EDIT-NOTES.md first. Compare its list of orders and times with your own. A missing order is the easiest thing to catch here.
- Watch the edit at the size people will see it, on a phone if it is for social.
- Step through every gesture cue frame by frame: each point and the snap.
- Check the loudness with
ffmpeg -i edit.mp4 -af loudnorm=print_format=summary -f null -. The input integrated figure should be close to -14 LUFS. - Send fixes by order and time, one per line: "The reset after the snap is late. Move it to the frame where the snap sound peaks, check the frames around it, and render again."
Stage 5 (optional): Run a fair head-to-head
- Use the raw cut from Stage 2 for both editors, so the footage is identical.
- Give both the same prompt. Change only the OUTPUT and SCRATCH lines, and keep the folders separate so neither can see the other's work.
- Start each in a fresh session, pick the model by name and screenshot the picker.
- Run them one at a time if they share a computer, so neither slows the other down.
- Write down who made what. If a session runs out or another tool steps in, label the edit that way.
- Do not fix either edit. Stack them at the same size, and apply any layout or retouch to both.
Troubleshooting
| Problem | What to do |
|---|---|
| An effect lands early or late | Name the trigger ("on the sound of the snap", "when I point") and ask it to check the frames around the cue |
| Ragged cut-out edges around you | Give the time range and ask it to redo the masks there and check the edges frame by frame |
| Captions cover your face or hands | Give the time; the prompt already bans it, so ask it to recheck that rule across the whole edit |
| The output is longer or shorter than the input | Ask it to compare frame counts with the input and fix the timeline; frame 1 must match frame 1 |
| The session runs out mid-edit | Resume in a new session from the scratch folder, reusing saved files |
| Your computer crawls | Make sure only one heavy job runs at a time: render, encode, transcription or cut-out |
Honest notes before you publish a comparison
- Usage limits can stop a long edit halfway. Ask the agent to save its working files (transcript, masks, tracking) in the scratch folder as it goes.
- Vague orders split editors. "Make it look more modern" produced two unrelated looks.
- Keep personal style rules out of a public test, or give both editors exactly the same ones.
- Label the result honestly. A comparison that later turns out to be mislabelled costs more trust than the views are worth.
If you only need one good edit, skip the head-to-head and use whichever agent you already pay for. If you want a benchmark score, this is not one. And if you want ad creative and reels made for your brand with tools like these, that is what our creative production team does.
Frequently asked questions
Is Gemini or Claude better at video editing?
Our test cannot answer that. Codex made the edit labelled Gemini 4 Argon, and Codex finished the Opus 5.5 design after its session ended. Google says Argon leads on long-video understanding (91.7% on LVBench), but understanding a video and editing one are different jobs. Run a fair test with the same raw cut and prompt in fresh sessions to decide for your own work.
Where can I use Gemini 4 Argon?
Google opened it in the Gemini app for Google AI Pro and Ultra subscribers on 13 October 2026, and it rolled out to Google Antigravity for AI Pro and Ultra in the days after, with its own weekly quota. Developers can use it through the API at an intro price of $2 per million input tokens and $10 per million output tokens.
Where can I use Claude Opus 5.5 for video work?
In Claude Code, in a terminal or the Code tab of the Claude desktop app. It comes with Claude Pro ($20 a month, or $17 billed yearly) and higher plans. Long edits use a lot of your plan's allowance, so ask the agent to save its working files as it goes.
Can an AI agent edit a video from spoken instructions?
Yes, if it can run tools on your computer. It transcribes the clip with word timestamps, treats each spoken order as an edit at that moment, then builds the effects, renders and checks the frames. Give it a raw cut with a short pause after each order, and name gesture cues such as a snap outright.
How do you run a fair AI video editing test?
Cut the raw clip first with no effects, give both agents the same file and the same prompt, start each in a fresh session with the model picked by name, run them one at a time, do not fix either edit, and label who really made each one.
Why did the snap effect land late?
In the Opus-designed edit the reset was mapped about half a second, roughly 15 frames, after the real snap, which sits where the snap's sound spikes. The notes do not say why. Tell the agent to reset on the sound of the snap and to check the frames, not only the transcript.