Case study: an overnight free-model marathon
13 single-file HTML tasks - games, generative art, data viz, an interactive dashboard - each sent to 4 free models: cohere/north-mini-code:free, poolside/laguna-xs-2.1:free, tencent/hy3:free, nvidia/nemotron-3-nano-30b-a3b:free. 52 runs total. Every screenshot below is the actual final rendered output - click any thumbnail to open the live page.
2%
Succeeded on the first try, no fix needed
100%
Succeeded after auto-fix (mimo-v2.5 fixer)
$0.189
Total cost across all 52 runs
Biggest finding
On this kind of open-ended, creative single-file task, free models landed it on the first try in only 1 of 52 runs (2%). Almost every run either timed out at 300s with nothing, or announced an intent to build something and then didn't. The auto-fix chain (xiaomi/mimo-v2.5 as the fixer) rescued the other 51, taking the effective success rate to 100%. Without that fixer, this batch of "free models building fun stuff" would have looked like a near-total failure instead of a complete success.
Space parallax scene
Memory match game
Synthwave music player
Whack-a-mole
Dodge game
Tic-tac-toe vs AI
Interactive dashboard
Fireworks simulator
Kaleidoscope generator
Gradient blob playground
Solar system simulator
Animated bar chart race
Fake audio visualizer
Every screenshot and HTML file is the actual model output, rendered as-is - click any thumbnail to open the live page.
Full testing history and methodology: /stats.