In this video I wanted to answer a very practical question for agent builders: does Gemini 3 “think” better with images or with text? DeepSeek’s OCR paper triggered this for me—if an image can compress a ton of information, maybe we should shift more agent inputs from text to vision. So I upgraded my 3D chess arena and ran a clean experiment: the same chess setup, but one match where the model only sees screenshots of the board (plus a small history window), and another match where it only sees a text representation of the board state. What shocked me was the trade-off. In image mode, Gemini 3 showed stronger spatial reasoning and played better chess overall (better ELO/ACPL in my run), but the hallucination/illegal-move rate jumped to ~26%, which forced a retry system to keep the game legal. In text mode, legality was basically perfect (under 1% hallucination), latency was way better, and the whole run was much cheaper—but the model made more tactical/strategic mistakes because it had to “reconstruct” the board mentally from text. My takeaway for real agent use cases: multimodal inputs can boost spatial understanding, but you pay in reliability, latency, and cost—so only use images when you truly need that visual signal, and consider combining image + text when possible.