Qualixar film · 12:45
WATCH EVERY FILM IN FULL (all 11 contest entries, untouched, with checksums):
[https://drive.google.com/drive/folders/1yBmCb2h63Fp5yr_37LY46f8BSRa_3GiO?usp=sharing]
GPT-6.1, Opus 5.5 and Fable 5.2 take on the same motion-graphics briefs. Which result would you post?
Four rounds; Fable entered three. Excerpts are edited for review, and the checks use the original files.
FREE GUIDE (every prompt, setting and brief from this video, 69-page PDF):
https://qualixar.com/learn/guides/brief2mp4-guide
FREE SKILL from the end of the video, brief2mp4 (MIT):
https://github.com/qualixar/brief2mp4
CHAPTERS
0:00 Same briefs, three AI models
0:31 Round 1: glass logo
1:23 Round 2: data into a 3D chart
3:40 Round 3: the laptop launch film
6:05 Round 4: four worlds in one camera move
9:45 My picks and the scoreboard
10:20 Make your own with brief2mp4
INSTALL BRIEF2MP4
Claude Code plugin (type in Claude Code):
/plugin marketplace add qualixar/brief2mp4
/plugin install brief2mp4@brief2mp4
Claude Code and Codex together, one line in a terminal:
git clone https://github.com/qualixar/brief2mp4.git && bash brief2mp4/install.sh
No terminal? Ask Claude Code:
Install the brief2mp4 skill from github.com/qualixar/brief2mp4 for Claude Code and Codex.
Then start a new session and type:
Use the brief2mp4 skill to make the video in this brief: (paste your brief)
When it says it's done, send one follow-up:
Check it against the brief.
You need Claude Code or Codex, Node.js 22 or newer and ffmpeg (Blender only for Blender briefs). The skill is free; the models run on your own Claude or ChatGPT plan.
HOW THE TEST WAS RUN
Each AI got the same brief and toolbox in its own sealed folder: no skills, no memory, no project notes. None of them had brief2mp4. Effort was medium for all three. Each entry used one follow-up: "Check your output against the brief and fix every mismatch." In Round 1, Opus's toolbox still mentioned two follow-ups; an extra follow-up was stopped within seconds. Time caps count agent time only. Fable sat out Round 1. Full settings, toolbox and briefs are in the free guide.
RESULTS (my picks; one run per model, results can vary run to run)
Logo: GPT-6.1
Data: Fable 5.2
Launch: Opus 5.5 and Fable 5.2, shared
Four Worlds: Fable 5.2
AUTOMATED CHECKS
Logo: GPT 12/12 · Opus 12/12
Data: Fable 12/13 · Opus 13/13 · GPT 13/13 (Fable's miss: an OCR misread of the value order, checked by eye)
Launch: Opus 14/14 · GPT 14/14 · Fable 13/14, 14/14 by eye (a caption misread)
Four Worlds: Opus 14/14 · GPT 14/14 · Fable 12/14 (a thin-font title misread, checked by eye, and one real miss: a 65 ms soundtrack drift)
RECEIPTS
Counted agent minutes (Fable / Opus / GPT):
Data: 26.5 / 36.7 / 11.5 (GPT also waited 26.6 minutes in the render queue)
Launch: 9.9 / 29.4 / 7.7
Four Worlds: 30.0 / 21.3 / 78.6 (GPT also waited 33.6 minutes in the render queue)
Output tokens (Fable / Opus): Data 42.3k / 34.0k · Launch 48.8k / 39.2k · Four Worlds 130.6k / 55.1k
API-equivalent cost meter (Fable / Opus): Data $1.83 / $1.67 · Launch $2.40 / $2.19 · Four Worlds $6.40 / $2.87
GPT total tokens from the Codex log, input included (not comparable with Claude output tokens): Data 157.9k · Launch 147.3k · Four Worlds 700.0k
Launch: GPT's picture holds still for 1.03 s, and 1.9 s on its end card (a quality check, not a brief rule).
Four Worlds: Fable's soundtrack drifts 65 ms from its saved music project in the last section; 7 of 8 sections match.
$ = API-equivalent cost meter. Every run was paid on subscriptions.
HOW THE FILMS WERE MADE
The AIs wrote web pages and rendered them with HyperFrames, or scripted Blender in Python. Fable's four worlds are Three.js inside HyperFrames. Every data chart came out of Blender. In Round 4 all three wrote their own music in SoundCraft.
DISCLOSURES
The narration is an AI replica of my own voice. The reaction clips in the corner are generated from my avatar. In Round 3 the films' original voiceover is swapped for a stock voice; the timing was measured on the original films. The short app clips are recordings made afterwards to show the apps; results come from the sealed runs.
I run these tests as part of my AI Reliability Engineering work at Qualixar: same brief, sealed runs, checks you can read.
Next video: the full brief2mp4 tutorial and all the GitHub repos.
#ClaudeCode #Codex #AIVideo
Connecting official YouTube player…
Press play in the official YouTube player. Playback is never started automatically.
Watch on YouTube ↗Evidence status