Gummy
Bears.

Eight AI models got the exact same prompt: build squishy, see-through gummy bears you can grab, stretch, squash, throw and pour. One HTML file each. Every build below is untouched. Click one and play with it.

Play the top build

Names and scores are hidden and the order is shuffled. Play them all, pick your favourite, then turn blind mode off to see which model built it.

Read the exact promptWord for word what every model got. Adapted from a Reddit post.
 

  






  
    

  
    
    
    
      
    How it was run
    
      
      
        
    • Same prompt, word for word, in an empty folder, with nobody answering the model's questions.
      • 
        
    • Every model on its highest setting: extra-high for Claude, GPT and Grok, high for Gemini (Google's top level). Kimi K3.1 Pro was set to extra-high, which OpenRouter may have ignored.
      • 
        
    • Claude models ran in Claude Code, OpenAI models in Codex, Grok in Cursor, both Gemini models in Gemini CLI, and Kimi K3.1 Pro in Claude Code over OpenRouter.
      • 
        
    • One run each, scored out of 10 by playing the build. Small gaps between scores don't mean much.
      • 
        
    • Costs are what each run would cost at the model's list API price, worked out from the tokens it used. The GPT models actually ran on a ChatGPT plan. Grok ran in Cursor, which doesn't report tokens, so it has no cost.
      •