Industrial software that runs a warehouse like a strategy game

Sonnet 5.5 and GPT-6 Astra got the same task, to build what you think you look like, and built side by side on a Minecraft server. Sonnet 5.5 placed blocks directly, with no commands, and raised a smiling sun with orange rays.

Within limits of 45 minutes, 10 build calls and 6,000 blocks, the model decided for itself how it looks and built it.

With a single prompt, it made a launch promo video for Opus 5.5, complete with game-style background music. It used only 12% of the 5-hour usage limit.

It makes a whole video in its default state, with no extra tools or setup.

With Opus 5.5 set to Max effort, its deepest thinking level, a single request asked for a 15-second motion graphic like a showreel (a short compilation video that shows off your skills) to attach to a résumé. The result was a video of bold CLAUDE lettering and the phrase motion designer moving across a red screen.

A self-introduction video that motion designers put real work into can come from a one-line request.

From a single request to explain how camera focus works, it worked for 1 hour 26 minutes and built a hands-on 3D lens lab web page. Turning the focus ring moves the glass elements inside the lens and shifts the plane in sharp focus forward and back through the scene. The API (the way programs call the model) cost $25.66.

In one pass, it turned optics that are hard to follow in text into a teaching tool you learn by turning it with your hands.

A new dashboard for Home Assistant, the smart home program, was built with Claude Code and Opus 5.5. It is a single page of HTML, JS and CSS, with no framework (a prebuilt development scaffold) and no npm. It includes screens for garden watering, home power, the electric car and lights by room.

Opus 5.5 handled everything from writing the code to preview testing, SSH (remote access) deployment and automation setup, and finished a screen that controls the whole home.

Opus 5.5 at medium (middle reasoning effort) built the screens of SharkFin, a personal finance dashboard like Monarch. It connects to bank accounts and shows spending, net worth and a transaction list on computer and phone screens. It also has a page for cash flow by property.

On just 35% of the weekly limit of the $20-a-month Claude Pro plan, you can finish a personal finance app that keeps your data in your own hands.

A train journey with scenery passing outside the window, drawn only with JavaScript code and no image files. It took a single prompt and about 45 minutes.

Even a printmaking (risograph) art style can be expressed in code.

Opus 5.5 made a UI motion video using only code. In it, a single shape keeps turning into screen elements such as a button and a music player. The video editor After Effects was not used at all, and the author also shared a prompt template for making the same video again.

Designers used to polish UI motion scene by scene in dedicated software. Now it can be made with a single prompt and an HTML file.

A short animated story made with code. The official Anthropic account posted it first when introducing works made with Opus 5.5.

Code alone makes an animation that tells a story.

A game made over a weekend with Opus 5.5. A stickman walks around on a website screen and shoots weapons to scatter the text and break the screen. It also supports multiplayer with several people.

In a single weekend, the idea of turning a real web page into a playground became a game that can be enjoyed right away.

After a third round of refinement with Opus 5.5 at medium (middle reasoning setting), Blender (a 3D program), an image-generation AI and three.js (a web 3D tool) were combined into a game that runs directly in the browser. The screen shows a character standing on a skyscraper rooftop and a small map at the bottom right.

Making a game used to mean moving between many tools. One model stitched them together into a result you can open without installing anything.

Asked to design a microduck that could be built from real LEGO, Opus 5.5 designed a life-size model using 1,113 real LEGO parts. It checked 3,204 connections, found 0 places where parts collide, confirmed that every step can be assembled, and confirmed that the center of gravity sits inside the feet. It then made a 141-page LEGO-style manual with 237 steps, priced the parts in the browser and prepared an order on BrickLink (a LEGO parts marketplace).

The whole process of actually building a LEGO piece can be handed to the model, from design and structural checks to assembly instructions and parts ordering.

It moved HAProxy, the open-source tool that distributes requests in front of web services, from C to Rust in full. It passed nearly all of HAProxy's own tests in 9.5 hours, while Fable 5.1 took 12 hours on the same job.

It finished a migration that would take a human team weeks faster than a more expensive model, at 51% lower cost.

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.

This scene was built with Three.js (a JavaScript tool for drawing 3D in a web browser). You steer a boat down a river through Japanese scenery. The weather changes and the light shifts between day and night. It has reflections on the water, a pagoda, cherry trees and 3D people, and you can try it right away on a public site.

A 3D scene with water reflections, physics, weather and buildings can be made to run right in the browser with no install.

A portfolio website showing @batchmerch's work was built with just three prompts to Opus 5.5. T-shirts hang in a row on a clothes rack rail, and a band announcing a newsletter subscription scrolls across the bottom.

A designer can finish a showcase site that keeps the brand's mood in a short conversation, without writing code.

This 3D island town has moving birds, crabs, fish and whales, cloth and signs that sway in the wind, and pier lights that come on at night. Most requests were along the lines of 'add X and Y' or 'this looks off, make it better'. Using several subagents (helper AIs that split up the work), it used 59% of the $200 plan's weekly usage limit over about 8 hours of real time.

With no technical instructions, just repeated short requests, one person can build a 3D world that never drops below 60 FPS (frames per second) at 1440p resolution.

Alongside GPT-6 Astra, it was asked to make five games, each from a single prompt: a 2D platformer, a racing game, spaceship combat, a first-person shooter and a stone golem boss fight. No human wrote a single line of code. The 3D assets (the character and object models in the games) were not downloaded. All of them were made from scratch in Blender (a 3D creation program).

It shows that five games in different genres can be filled out with one request each, from model building to graphics, sound and game rules.

Claude Agent SDK, a development tool for building agents, had its code formally verified in Lean. This method checks mathematically whether code behaves as designed.

The proof process found concurrency bugs that are hard for people to spot, and a few short requests produced 16 code fixes.