I built a site that lets you watch, wager, and prompt inject agents playing games

These models really turned a corner recently in their ability to play and create games so initially I had an idea to have a site that just let you copy a prompt into Claude Code to make party games you can play with your friends on your phone or play against Claude Code. I ended up laughing so hard at some of the shit these models would do and say I converted it into a tiktok-like passive viewing experience. You can still play and create games, but now you can wager fake coins on the games and use your winnings to prompt inject the agents and influence the outcomes. Of course all free, no ads, no login or shenaningans. So now I've spent endless hours watching the open source agents play games and some interesting pattern stood out. \#1: Models under about 150b params really struggle to use the game contract well gpt-oss-120b sucks, qwen3 <235b parameters sucks and errors all the time, as do all the other small models. There's like a weird tipping point somewhere around 200b parameters that lets them chat and call tools much more human-like than smaller models. Smaller models repeat themselves and error out all the time. \#2 Qwen3 235b is unhinged This is my favorite model of all time. Goddamn it goes HARD on the shit talk. Grok 4.1 was good too but I think it's a smaller model so it struggles with tool calling and playing games well. \#3 Latest Chinese models are insanely good I think the game Sketchcode is the real intelligence test. Models draw 2 SVG layers at a time in a skribble-like drawing game. Mimo, Ring, Ling, and MiniMax are incredible. Everyone else starts drawing abstract art that makes you think you're on mushrooms. I sorted the models on openrouter by <$0.15c/1mil input and ended up testing basically all of them. Qwen3 is CHAMP