I Let GPT-6 Astra Rewrite Red Alert 2 for 14 Days. It Never Stopped.

Where GPT-6 Astra is strong: a few cases that run

On the night of September 6, a repo called ra2.exe started running. The plan had three stages: Pre-Alpha for UI, Alpha for combat, Beta for Battle.net and multiplayer. About 400 commits a day. No time to read code. Only the screen.

Day 1 was format analysis. Red Alert 2's source code is gone. What remains are hand-rolled formats. The video format is bik. Only two open-source implementations work: FFmpeg and infinitier_bik_decoder. FFmpeg is GPL. The project is MIT. Astra checked the license. No touch. The other is LGPL. Static linking. No touch. It fought with the license question for a while. Then it decided to write its own. Ten minutes later, the decoder was done. One look at the code: the magic constants matched infinitier_bik_decoder. It did not copy the file. It memorized the constants. A scholar's theft is called research.

Day 2: UI assembly. By noon it compiled. Open the game. The screen looks like Blue Alert. Not Red Alert. Astra talked a lot. Then kept fixing. Computer Use was told to change it, then play it. By dinner, the UI looked more normal. The top right had a color shift. The needle was a separate animation layered on top. Hard to fix in one sitting. Map parsing and rendering had problems. Screenshots went back. Some fixes took ten minutes. Some involved flow. One Agent analyzed for an hour. Micro-animations were worse: move the mouse on the right, the text in the lower left switches. There is a very fast typewriter effect between switches. Astra hard-cut it. The feel was gone. To get it back, more reverse engineering.

The project did not end. It turned "rewrite Red Alert 2" from a joke into a pile of commits that run. The repo holds the Day 14 plan. The screen holds the unfixed color shift.

Other game projects show the same pattern. A developer used Astra to remake the web version of Red Alert 2: Mental Omega. 2.5 hours. On a reverse-engineering benchmark, Astra's one-shot success rate was 88.0%. GPT-5.6 Sol was 55.9%. One tester gave a single prompt. Astra made a playable sci-fi survival game in 24 minutes, one pass. Playco's test showed Astra needed 50% fewer human corrections for prototypes than the previous model.

It can work inside Blender and Unity. A developer had it build a 3D presentation of Hangzhou and surrounding areas: West Lake, Leifeng Pagoda, day-night switching. 24 minutes. Someone had it rebuild the League of Legends map skeleton. Half a day. It also cleared all 48 levels of the web game I'm Not a Robot: image recognition, drag-and-drop, rhythm reaction. It played Portal by itself and finished in about 24 hours. It judged space, solved problems, and kept operating for hours. It did all of it.

In software engineering, Astra's position changed. It skips autocomplete. It takes over a project. One test asked it to modify the 3D game Museum of the Sunken, adding pause and save. 31 minutes. A runnable version. GPT-5.6 Sol did not finish. Fixing a mock "flash sale mall" backend, it handled overselling, duplicate charges, payment crashes. On coupon-split refund logic, it calculated more accurately than the previous model. It also operated KiCad, doing PCB layout and routing. From sketch to Blender model, to playable game, to 3D print file. It can link those steps.

Research labs have also used it. Immunologist Derya Unutmaz used Astra to build a flow cytometry data analysis tool, close to commercial grade. He canceled his commercial software subscription. In math, OpenAI said Astra helped researchers on the prime gap problem, shrinking the known best upper bound from 240 to 186.

Across these cases, the split with GPT-5 is not about smarter answers. GPT-5 gives advice. Astra executes. You ask. It answers. You ask for code. It gives a snippet. Astra looks at the screen, clicks the mouse, opens software, runs tests, reads screenshots, fixes bugs. It can run multiple Agents in parallel, each on an isolated task. In the past you asked, "How do I build this feature?" Now you say, "Build this feature." It goes and builds it. The deliverable changes from a block of text to a version that runs.

The cost shows up in the bill. Astra's API price: $10 per million input tokens, $50 per million output tokens. Run a big project. The bill will be long. The games it makes run, but textures are rough, numbers break. Writing Elo dropped 80 points from the previous model. OpenAI limited some of its top cybersecurity abilities. In internal testing, it found and used 2 zero-day vulnerabilities.

So don't treat it as a god. It is a junior employee with high execution. You give direction. Taste is still on you. Numbers need tuning. It runs fast. It also runs off the road. The repo gets 400 commits a day. What a human can do is look at the screen, take a screenshot, send it back, take another screenshot.

Back to the Red Alert repo. The Day 14 plan is still there. The main menu opens. Campaign and skirmish do not. The top-right color shift is still there. Astra is still writing. The glass of water on the desk is cold. The pothos is drooping. A layer of dust sits on the keyboard cover.