It leads on arc-agi-1 with Gemini 3.0 Deep Think, which uses "tool calls" according to google's post, whereas regular Gemini 3.0 Pro doesn't use "tool calls" for the same benchmark. I am unsure how significant this difference is.