NEW MODEL: Google enters the in-context prompting race! Gemini 4 Argon is Google's latest frontier model, offered through the Interactions API in AI Studio, the paid Gemini API tier and Antigravity. The robotics claim attached to the launch is a pipeline Google calls Video-to-Sim-to-Real: Argon watches a video of a person doing a task, reconstructs the room and the objects in it as a MuJoCo scene, writes Python that generates and tests candidate motions inside that scene, then runs the result on hardware. The launch demo is cornhole, the beanbag-into-a-hole lawn game, on Apptronik's Apollo humanoid. Google claims the loop compresses weeks of manual simulation modeling into a few hours. What i find interesting is that the video Argon learned from was shot by the robot. The operator's prompt is legible on screen in the console, the telemetry panel labels the reference footage, and a log line records the details. So the input is an egocentric capture the robot took of a colleague standing in its own lab, next to the board it was about to throw at, from a head camera whose extrinsics the system already knows.