Four kart racers from one prompt
Alexey Fateev ran one Mario Kart–style prompt four ways: DeepSeek and GPT-6 Astra alone, then each with the other as a subagent. Four builds, each from scratch.
Alexey Fateev wanted a comparison, so he built the comparison four times. Same kart-racer prompt. DeepSeek with an Astra subagent, Astra with a DeepSeek subagent, then each model alone, because a pair of results was not enough.
Astra stayed on medium reasoning. DeepSeek stayed on max. Every version started from scratch. A Mario Kart–style game is a useful instrument here: a track, a camera and a lap are easy to see and hard to fake for ninety seconds.
The post is less about any one winner than about the spread. Four garages, one assignment, and a minute and a half of them driving.

Made with
Alexey Fateev
@superalesha
I gave DeepSeek an Astra subagent. Then I gave Astra a DeepSeek subagent. Then I made both work alone, because apparently I needed 4 versions of Mario Kart. Same game prompt, each build from scratch. Astra on medium reasoning, DeepSeek on max throughout. I just wanted to see https://t.co/R7CUNbhu8Y
1.2K likes
- likes
- 1,162
- replies
- 115
Captured when this was curated, not live. —