Google just shipped what they’re calling Gemini 3.1 Pro (still preview), and despite the “.1” label this is a major upgrade. In this video I break down the release, the benchmark story, and what I think Google’s strategy is versus the other frontier labs. The big theme: Google is pushing harder toward a generalized model while others have been optimizing heavily for coding, and they’re back to dominating what I call the “Pareto frontier” of performance vs cost. On benchmarks, Gemini 3.1 Pro is leading on a lot of the major reasoning and agentic coding evaluations. The standout for me is ARC-AGI-2 because it’s a clean signal for reasoning capability—and the jump is more than double the previous iteration while staying at essentially the same price. I also cover Humanity’s Last Exam (strong even without tools), improvements Google claims around tool calling and hallucinations, and the fact that token efficiency seems better than some recent releases (cough… token-hungry models). Finally, I walk through the new AI Studio “Build” experience where Google’s Antigravity agent is now integrated directly into AI Studio. Your app runs in a sandbox, you can pick stacks like React/Next/Angular, set secrets, push to GitHub, and publish deployments. My take: raw model intelligence isn’t enough anymore—the wrapper/scaffolding (Deep Think, generator-verifier loops like Althia, and now agentic build environments) is where the real leverage is.