Gemini 4 Argon构建的脑控可玩游戏

Gemini 4 Argon built a playable game you control with your brain through a Bluetooth meditation headband. One prompt. Google is back. I went through a week of real testing from early users. Here's what Argon actually does: 1. Ranks #1 on 13 of 19 benchmarks. 77.9% on Deep Sweep, 91.9% on Vibe Code Bench. Behind Opus 5.5 on Terminal Bench but not by much. 2. Built a Fall Guys clone with 10 playable levels. Took iteration, but the final version is actually fun. 30 minutes. 3. Built a working Excel clone in the browser in 20 minutes. Formulas, drag-fill, SUM across ranges. It actually opened Excel on the tester's computer first to study it. 4. Built a cyberpunk 3D city in Three.js running at 60fps with neon bloom controls, auto director, cockpit HUD. 5. Found a webhook exposure bug in code written by Opus 5.5 that nobody caught. 6. Massive jump on knowledge work. Box benchmarks went from 71% on Gemini 3.1 to 85% on Argon. Spreadsheet eval, PDF extraction, documents at scale. 7. Spends about 30 minutes on a task by default when you don't set a budget. It's not just a response pattern, the model has a work rhythm. 8. Stays grounded in documents you give it. Doesn't drift, doesn't fill gaps, doesn't invent what's not there. Big deal for legal and research work. Where it loses: > 3D graphics still behind Opus 5.5 and Astra. Noticeable in Unreal Engine tests. > Needs more prompting than Opus to make games look good. Opus one-shots the aesthetic, Argon needs direction. > Antigravity as a coding tool still feels rough next to Codex or Claude Code. Strongest model, weakest wrapper. Available now for Google AI Pro and Ultra. API through Google AI Studio. Four frontier labs now shipping competing models in the same month. We're the ones who win.