Claude vs Remotion vs HyperFrames for AI video

The short answer

• A solo creator or a team of up to three making videos for yourselves: any of them works. Pick Remotion if you know React, HyperFrames or plain HTML if you do not.
• A company of four or more, or a product that renders on behalf of users: HyperFrames or plain HTML. Remotion's licence charges for programmatic renders and its next major version bans bring-your-own-code rendering services.
• You want the most mature tooling and do not mind paying: Remotion.
• You want the result, not the pipeline: a hosted tool such as Video Studio, or Anthropic's Claude Motion if you are on a Teams or Enterprise plan.

Side by side

Remotion: the mature one, with a licence to read

Remotion has been around longest and it shows: a big component ecosystem, captions, a Player you can embed, and Lambda rendering that is well documented. Its Agent Skills launch in January 2026 is what made "Claude makes videos" a genre, and its skill is still among the most installed.

The licence is the catch. It is free for individuals, non-profits and for-profit companies with up to three employees. Above that, a company licence is required: $25 per seat per month for people who make videos in it, and for anything that renders programmatically, $0.01 per render with a $100 monthly minimum. Every successful render counts, drafts and stills included. The licence explicitly allows a service that has AI write Remotion code and renders it, but the upcoming 5.0 terms ban letting users bring or upload their own Remotion code, and start counting contractors toward headcount.

For Claude specifically, React adds weight: more files, a bundler, and more ways for a build to fail before a frame is drawn. Claude handles it well, but in our experience the same video costs more turns in React than in a single HTML page.

HyperFrames: open, HTML first, still young

HyperFrames is HeyGen's open source renderer, released in April 2026 under Apache-2.0. Claude writes ordinary HTML: elements carry data-start and data-duration attributes, and animation runs on a paused, seekable timeline (GSAP, CSS, Lottie, three.js or a custom adapter). Its engine drives Chrome headless shell with beginFrame on Linux, which gives deterministic frames from the compositor itself, and falls back to screenshots elsewhere.

It ships the pieces you would otherwise build: Lambda and Cloud Run renderers, an embeddable player, a linter that catches non-deterministic code, caption blocks and shader transitions. There are no per-render fees and no headcount rule.

Two cautions. It is pre-1.0, so pin the version. And its own agent loop is slow: in one public end-to-end test it took 23.5 minutes on a 15 second brief, where plain Claude Code with no skill finished in 4.4. Using its engine as a renderer and keeping your own authoring loop avoids that.

Plain HTML and Playwright: what we use

The third option has no framework. Claude writes one page that exposes a small contract (size, frame rate, duration, a ready promise and seek(t)), and a few hundred lines of Node drive Chrome through Playwright: seek, screenshot, pipe to ffmpeg. Our step by step guide shows both halves.

The appeal is that nothing stands between Claude and the canvas. A page is one file, previews in any browser, and needs no build. Licence risk is nil. The cost is that you build the scale tooling yourself: segmenting renders across workers, motion blur, a purity check that renders the same frame twice and compares, loudness and safe-area checks. We have run this setup for 72 films and 90 Shorts, and it is the engine under Video Studio today.

A practical middle path is to own the contract and treat the renderer as swappable: write pages against your own seek(t) contract, add a small adapter, and you can capture them with your script or with HyperFrames' engine, whichever is faster on your hardware.

Where Claude Motion fits

Anthropic shipped Claude Motion in beta in October 2026 for Teams and Enterprise plans: describe an idea, a report or a chart, and get an animated explainer as an MP4 that stays editable as code. It is the platform owner entering this space, and for a business user who wants an explainer of a slide deck it removes the renderer question entirely.

It is not aimed at creators running a channel: there is no style library tuned for Shorts, no captions synced to a voiceover and no channel tooling, and it is not on individual plans yet.

Speed and cost in practice

Render speed depends far more on what you draw than on the framework. On a Mac we capture 2D pages at about 22 fps for drafts across three workers and about 4 fps with six-sample motion blur. On a CPU-only Linux server, 2D is fine but WebGL through software rendering crawled at 0.7 fps for us, so 3D styles want a GPU whichever renderer you choose. Remotion's own figures put Lambda compute at roughly $0.02 per minute of 1080p video, on top of any licence fee.

The bigger cost is Claude, not the renderer. In our measurements an agentic Claude Code session costs about $7 to $13 at API list prices for a 60 second video, and most of that is the render and review loop. A framework that needs fewer fix-ups saves more than one that renders a little faster.

How to choose

• You know React and are three people or fewer: Remotion. Enjoy the tooling.
• You will render inside a product, or you are four or more people: HyperFrames or plain HTML.
• You want Claude's code to stay as simple as possible: plain HTML, one file per video.
• You need cloud rendering next week without building it: Remotion Lambda (with the licence) or HyperFrames' Lambda package.
• You need 3D: any of them, on a machine with a GPU.
• You want finished videos and no pipeline: a hosted tool.