All News
anthropicclaudevideo-generationcode-generationai-agents

Claude Opus 5.5 makes explainer videos by writing code, not by generating pixels

Two demos now show Claude Opus 5.5 making explainer videos by writing HTML, Playwright and ffmpeg code, not pixels. Practitioners call the explanation thin.

Vlad MakarovVlad Makarovreviewed and published
3 min read
Claude Opus 5.5 makes explainer videos by writing code, not by generating pixels

Two demonstrations this week converged on one claim: Claude Opus 5.5 can produce a short film with no video model in the loop. On September 23, a poster using the handle Deedy published an animated launch video for an inference startup that he says he prompted and rendered in about a minute for roughly $2. A day later, a Reddit thread showed a game-style fight scene animated entirely from code.

The film is a web page, screenshotted

Deedy answered the obvious question in a follow-up post: no video-generation model was used, only JavaScript, Playwright and ffmpeg. Claude wrote a browser animation, Playwright stepped through it frame by frame, and ffmpeg stitched the frames into an MP4.

The clearest account comes from launchvideo.io, a tool by diggerhq whose code is public: a URL or prompt in, an MP4 out. The product is one agent file, three tools and a form on OpenComputer's serverless platform, running Opus 5.5 at roughly 90,000 input and 15,000 output tokens per film.

Rendering, in its own words:

  • Clock: requestAnimationFrame, timers, Date and CSS replaced by a virtual clock, so every frame is a deterministic seek
  • Output: 1920x1080 at 30 fps, JPEG frames piped into libx264 at crf 18

Practitioners call the explanation thin

Hacker News titled its thread "Opus 5.5 is good at explainer videos", and the skepticism was specific. One commenter said the result is "just taking an existing presentation slides skill and generating a video out of it", with no animation that explains an actual concept. Another asked for a site for his own project and got a film about an AI startup centralizing data. A third noted YouTube is already full of AI explainer slop.

The loudest showcase was not a work explainer at all. On Reddit, a poster asked for a Super Nintendo-style fight in which the Bing alter ego Sydney faces Sam Altman and then Claude itself, music included, with no assets supplied; the thread drew roughly 449 points and 88 comments. A top reply marvelled at hand-written sound synthesis and frame-by-frame animation, then asked how it did not break the bank on tokens.

What would settle it

The cost figures circulating are estimates, not invoices. Deedy's $2 was never labeled API spend or a subscription path, and the reply asking which went unanswered; commenters in the thread cited their own runs at about $3.21 and $4 of OpenRouter usage. Published token accounting at list prices, and independent reproductions by people with no stake in the demos, would turn the estimate into a number. The harder question is quality: whether these films explain anything to someone other than their author. Opus 5.5 shipped two days ago as a cheaper flagship; none of that launch material mentioned video.

Related Articles

Scroll down

to load the next article