Google DeepMind's Gemini 4 Argon scores 87% on embodied reasoning test and goes on sale to developers
Google said in a developer blog post on Oct 13 that Gemini 4 Argon, a general physical reasoning model for robotics, scores 87% on the Embodied Reasoning (ERQA) benchmark using reasoning alone, which the company says beats every other frontier model. The post shows Argon controlling a Boston Dynamics Spot quadruped that rehearses a foot-pedal trash bin task in a MuJoCo digital twin before doing it for real, steering a robot arm zero-shot, building calibrated simulation scenes from robot camera feeds, human video or paper sketches, and steering the Gemini Omni video world model. Argon is available from launch in Google AI Studio and Antigravity for Google AI Ultra subscribers and on the paid tier of the Gemini API as gemini-4-argon. A day later NEURA Robotics introduced its own Neuron AI foundation model, so two robot-brain releases from the companies on MEKAWIRE's Google DeepMind and NEURA pages landed within 24 hours.
