A map of the Mongol conquests, 1206 to 1294, from a one-shot prompt

Given only the rules "make it look great, no repeated designs, go wild," it was asked to make 100 HTML files. 20 agents took 5 each, and every page was opened on desktop and phone screens for review.

None of the 100 files is broken, and several are games you can play through to the end.

The screen of the developer tool Claude Code was redesigned to look like paper written on in pen. Opening a new task brings up a single napkin.

It turns a familiar tool into a screen with a completely different feel.

The screens of Letters Abroad, an app for learning languages by exchanging letters with an AI pen pal, were rebuilt with Opus 5.5.

In a few hours it lifted a design that another model could not fix in a week.

The clunky screens of Goodbudget, a budgeting app earning $10,000 a month, were changed in one go to the design language of the money transfer service Wise. The same job took Opus 5 1 hour 39 minutes and Opus 5.5 28 minutes.

Redesigning the screens of an app that makes real money becomes a job of under 30 minutes.

Given just one reference site picked from PrettySites, a collection of well-made websites, it was asked to build an entire landing page with motion.

One reference yields a near-finished page in a few minutes.

A Gyeongbokgung simulator was rebuilt with Opus 5.5 so it can be compared side by side with the GPT-6 Astra and Fable 5.1 versions. Opus 5.5 led on detail of information, and Fable 5.1 led on rendering.

A head-to-head test of three models on a Korean subject.

A web service for managing shipping risk was built from 2 prompts in 18 minutes 16 seconds. The files the model wrote were deployed as they were, without a single line changed.

A complete service comes out exactly as the model made it, with no human edits.

The same prompt, "Draw a PS5 controller in SVG (vector graphics drawn in code)," went to Opus 5.5 and GPT-6 Sol, and the results were placed side by side.

The drawing came out a step above in detail, shading and accuracy.

After trying several models, this video collects the output of Opus 5.5, which had the best visual design.

Beyond writing code, it also leads in the sense for making screens that look good.

With Opus 5.5, T3 Code, an app for handing coding to agents (AI that works through multiple steps on its own), was put inside the game Minecraft. The agent's work status appears at the top of the game screen, and opening a window lets you see the task list and progress and send messages right there.

Even while playing, you can give the agent work and check its progress right away.

An app built with Opus 5.5 draws whatever you type in the input box at the bottom of the screen onto paper on a desk. The drawings come alive, and you can play by making them walk and run with the keyboard, like a stick figure.

The model made a toy where, with no drawing skill, one line of text gives you a character you can control yourself.

The paper-reading site alphaXiv used Opus 5.5 to add a blog-style explainer to every paper on arXiv (a public repository where anyone can post papers). Each sentence in an explainer is marked with the matching part of the paper as its source, and figures and key takeaways are included.

Even a difficult research paper can be read as plain writing with one button, while checking the original text behind each point.

With Claude Code, it built a music player in the style of Apple Music as a native macOS app (one installed directly on the device rather than run on the web) in about 35 minutes, and an iOS vocabulary memorization app in about 63 minutes. It runs them to show importing and playing songs, studying and practice, and also demonstrates a 3D house built from a single floor plan, a first-person shooter, an ink-wash painting animation and a gear machine simulation.

Beyond web pages, apps you install on a computer or phone can also be made to actually run with about an hour of work.