← All posts

Opus 5.5 paints, animates and scores music with code alone

2026-09-23 · 4 min read

A day after Anthropic released Claude Opus 5.5, some of the most-shared posts about it on X have nothing to do with benchmarks. They're art the model made by writing code, with no image generator anywhere in the process.

The most-viewed one I found came from Jake Eaton (@jkeatn), who works at Anthropic. He posted four paintings in different styles and explained that "every image here is a python program generated pixel by pixel." There's no image model behind them and no reference pictures. By his account, the agents work only from what they know about each painter, and the system runs to about 7,500 lines of code that imitate different brush styles. The post passed half a million views in about a day.

Two more from outside Anthropic

@Voxyz_ai shared a 29-second animated short called "small print" and wrote, "i didn't write a single line of code." According to the post, Opus 5.5 wrote the story, drew every frame in JavaScript and synthesized the music in Python. The whole thing lives in one index.html file. After rendering it, the model went back through the video second by second and polished it again.

@developedbyed, a developer with about 75,000 followers, ran the same pixel-art test on Opus 5.5 and OpenAI's GPT-6 Astra. His take on Astra's version: it "ends up looking so damn generic." He told viewers to look at the fountain's water animation and the bats' eyes in the cave, and said more tests are coming.

Plenty of other clips, songs included, were made the same way.

Why code instead of pictures matters

An image generator hands you a flat picture. If a color is wrong, you regenerate and hope. In these demos the model writes a program, and the program draws the picture.

That means you can edit the result. A web developer can open the file, change a hex code to match your brand or slow an animation down, and nothing has to be regenerated. The JavaScript versions are small files that run in a browser, and since the output is text, it can sit in version control next to the rest of your site.

Anthropic seems aware of this. Its launch page describes a tester who had several Claude models build a game from a single prompt, and says Opus 5.5 scored higher than any other model on "the strength of its graphics and polish."

My read: real, but filtered

The work is impressive. It's also a highlight reel. People post the attempt that worked, not the ones that didn't, and nobody sharing a clip has to say how many tries it took.

Eaton's own post is the best reality check. He says he's been asking the models to paint "for the past few months," and the painting system is thousands of lines of code. That's a long project that one prompt did not produce. Even Vox's no-code animation had a person deciding what to ask for and whether the result was worth posting.

Dev Ed's comparison is a single side-by-side test, so I'd count it as one data point. It does line up with a complaint I hear a lot: left to its defaults, a general model reaches for the safest, most average version of a thing, and average rarely sells.

There's a legal wrinkle too. Asking a model to imitate an old master is one thing. Asking it to copy a living illustrator or a competitor's look is another, and US copyright protection for work a machine made on its own is limited and still being argued over.

What it means for a business

Most local businesses will never need a code-painted oil portrait. Plenty could use smaller things: a light animated header, a short explainer for a service page, a chart in brand colors, a three-second sound at the end of a video. If a model can build those as code well enough that a developer only has to tweak them, small projects get a lot cheaper.

The model still doesn't know your brand. As we wrote about anti-slop files and taste, someone who knows the business has to reject the generic version, and nobody has automated that part yet.

If you're wondering where AI could take real work off your team, visual or otherwise, our free process audit is a good place to start.

08 / Start here

Find your worst bottleneck. Free.

A 20 minute call. We map where your week goes and pick out the first process worth automating. You keep the map either way, and there is no deck to sit through at the end.

Email

pgorski@newfacedesign.com

Phone

+1 (773) 627-2176

Based in

Chicago area

Working with clients everywhere