Same prompt · many models · raced side by side
Benchmarks
Little races I run most weeks: give every model the same prompt, the same machine, the same tools — then watch them build. The club gets them first.
Grok 4.7 xhigh → a table in Blender
Grok’s terminal beside Blender, from the splash, sped up. This render shows metal, glass, and a gold ring on the table.
Watch the race →
Grok 4.7 vs Opus 5 → a solar system
A new race. The browser stayed open. Grok 4.7 low built the sun, then the planets, in 362 seconds. Opus 5 high did the same in 468, and the last frame has the date and the rings.
Watch the race →
GLM-5.2 vs Kimi K2.7 → who builds the better game?
Two frontier models, two fresh Linux desktops, one prompt: build the best browser game you can. GLM-5.2 spent 291s on a maximalist survival shooter; Kimi K2.7 shipped a lean one in 71s — ~4× faster. Games side by side.
Watch the race →
A blank desktop → an interactive multi-page space site
GLM-5.2 vs MiniMax-M3 vs Kimi 2.7, each on its own fresh Linux desktop: build a home page with an animated starfield hero plus a /destinations page, linked by a real nav. The fastest wrote the least code — and still shipped both pages.
Watch the race →
A blank desktop → a cyberpunk landing page
Kimi 2.7 vs MiniMax-M3 vs GLM-5.2, each on its own fresh Linux desktop: terminal on the left, the page building live on the right. Three premium pages, a 2.5× spread in time.
Watch the race →
A blank Linux desktop → a playable game
GLM-5.2 vs Kimi-K2.6 vs MiniMax-M3, each on its own fresh Linux desktop, build a cyberpunk Asteroids game from scratch. One did it in 51 seconds.
Watch the race →
One prompt → an interactive solar system
Three coding agents, one prompt, one single-file interactive solar system — raced side by side and sped up.
Watch the race →