Video from code
What it means to have Opus 5.5 write a video as code, when it beats generated footage, what it costs and which tool to pick.
In a code video, Claude Opus 5.5 does not generate pixels. It writes a program (HTML and CSS, or React) that draws every frame as a function of time, and a renderer turns that program into an MP4 on your computer. Text, UI, numbers and logos come out exact, a change is one line, and the same project renders the same video every time. No video credits are spent. You pay for your Claude plan.
When to pick code over generated footage
Pick code when the video is about words, screens and numbers. Pick a video model when it is about people, places and physical things. Most good launch films use both.
| Pick code for | Pick a video model for |
|---|---|
| Product launches and feature teasers | People, faces and animals |
| UI demos with your real screens | Filmed looking places and weather |
| Explainers with diagrams and charts | Photoreal products, materials and light |
| Kinetic type and title cards | Mood pieces and music videos |
| Many language or format versions of one video | One off cinematic moments |
A Chinese creator summed it up: code video "won't misspell text, a detail change is one line, aligns to narration by the frame", but faces, animals and realistic scenes are much stronger in video models (yanhua1010). To combine them, see Mixing footage.
What Opus 5.5 does, and what you do
Opus 5.5 was released on 22 Sep 2026. In a code video it:
- Writes the composition: one file per scene, timed to a beat map.
- Captures your real website or app screens, so the UI is not redrawn from memory.
- Runs the renderer's checks (layout, text overflow, contrast) and takes still frames to look at.
- Can write a soundtrack in code, call a text to speech engine and transcribe audio for captions.
It reads frames, not moving video, and it cannot listen. One creator puts it bluntly: "it can't even watch video" (yanhua1010). So you are the director. You approve each stage and you judge motion, pacing and sound yourself.
The gates
Every serious workflow in our research uses the same order, with a human look between each step. Fixing a still is far cheaper than fixing a render, because a render has to be redone and watched again.
| Gate | What you approve | Skip it when |
|---|---|---|
| 1. Plan | Message, length, format, beat map, the on screen lines | Never |
| 2. One still per scene | Layout, type, colors, real assets, readability | Never |
| 3. Animatic | Pacing at half resolution with placeholder audio | The video is a single scene or a short loop |
| 4. Full motion | Entrances, transitions, one focal action at a time | Never |
| 5. Sound | Music, voice, effects on the beat, loudness | The video ships silent |
| 6. Render | The final file, checked on a phone | Never |
The Iterate page shows how to run each gate and when to stop.
Typical cost and time
Prices checked on 15 Oct 2026. Every run below is a single creator's report on their own setup, so treat them as ranges, not promises.
| Run | Time | Cost or usage |
|---|---|---|
| One line prompt, 26 second launch video | About 1 minute | About 2 USD (deedydas) |
| 7 scenes, 26 seconds, 8 model calls | Not stated | About 1.18 USD in tokens (neil_xbt) |
| 15 videos built in parallel on three engines | 26 to 34 minutes per build | About 4 USD per video at API prices (luniacllama) |
| 3 minute history film in Remotion | About 1 hour | 7% of a weekly plan limit (kimmonismus) |
| Long explainer at max effort with an ElevenLabs voice | 2.5 hours to build, 35 minutes to render | Not stated (mablesjoseph) |
| Ultracode motion video | About 8 hours | 220 USD in API credits (_junaidkhalid1) |
The cheapest way to start is Claude Pro at 20 USD a month (17 USD billed yearly) with HyperFrames, which is free. Expect 2 to 5 rounds of notes on your first film. Long autonomous runs cost 7 to 30 times more on the same prompt with no shown quality gain, so start small (Setup explains effort levels).
Tool choice at a glance
Pick HyperFrames unless your team already writes React. It is free, it has checks built for agents, and it now handles captions and simple cuts on footage too.
| Tool | You work in | Who can use it | Cost | Pick it when |
|---|---|---|---|---|
| HyperFrames (HeyGen) | Claude Code; Opus writes HTML, CSS and GSAP | Anyone with Claude Code | Free, Apache 2.0 | Default for launch films, UI demos, kinetic type and captions on footage |
| Remotion | Claude Code; Opus writes React | Anyone; companies of 4 or more people pay | Free for individuals and companies of up to 3 people | Your team writes React, or you need many variants rendered on AWS Lambda |
| Claude Motion | The claude.ai chat, no code | Team and Enterprise plans only (beta since 8 Oct 2026) | Counts against plan usage | Quick internal explainers and animated charts for decks |
| Higgsfield Katana | A Claude connector to Higgsfield | Higgsfield subscribers | Higgsfield credits, cost shown before each generation | Rebuilding a reference edit with up to 5 of your photos |
Notes:
- Claude Motion makes short animated explainers from text, charts, shapes and images and exports MP4. It has no video model, so "it doesn't produce realistic footage or people" (Claude help).
- Higgsfield says Katana is "powered by Claude Motion" (Higgsfield). Its launch came with dozens of near identical creator posts within hours, so judge it by its own page, not the feed.
- An advanced path skips frameworks: one HTML page with a
renderAt(t)function, screenshots by a headless browser, and FFmpeg. Inbox Zero's founder made about 15 product walkthroughs this way (elie2222).
The pages in this section
Setup
Which Claude plan, effort levels, installing HyperFrames or Remotion, preview, render and hardware.
Brief
The brief structure that works, references, the banned list and five example briefs.
Iterate
The gate loop, notes that work, contact sheets, a taste rules file and when to stop.
Audio and sync
Music, beat maps, voiceover, text to speech and caption timing.
Mixing footage
Seedance, Kling and Veo shots, three.js scenes and Blender renders inside a composition.
Fonts
Picking and loading fonts, including Vietnamese and other non Latin scripts.
Output
Aspect ratios, variants and languages, export settings and cloud rendering.
Troubleshooting
Failure modes and fixes, what code video cannot do, and hype to ignore.
For a complete worked example, follow the launch video recipe. For a one page summary, see the Video from code tool guide.
Sources
- Claude pricing and Get started with Claude Motion
- HyperFrames on GitHub and Remotion license
- Higgsfield Katana (vendor page)
- Posts on X linked inline; more in Sources