Google Awakens to Physics: Gemini 4 Argon Hands-On Tests Pour In

Google's Gemini 4 Argon has actually arrived, and Google is seen as back in the race after once being considered to be falling behind. On benchmarks it reportedly trades blows with GPT-6 Astra and Claude Opus 5.5 in text and logic. Its price is reportedly one-fifth that of rivals, and it is several times faster. It is especially notable as the first general physics reasoning model that appears to combine image, video, language and physical intuition. Wharton professor Ethan Mollick said Gemini 4 feels markedly different from other models. In hands-on tests, Argon grasped the geometric principle of Fourier series approximating closed curves, generating a pedaling pelican riding a bicycle. Its 3D Yingxian Wooden Pagoda was detailed but imperfect, recognizable as four of the original five stories. Its explanatory video with narration was usable, though the Chinese narration pronounced words oddly. A 10-second deep-sea video recreating the Titanic's sinking was of acceptable quality. In scientific reasoning tests, it coded a 3D stellarator magnetic field solver and a simulation of three-phase AC induction motor principles, though the Chinese pronunciation remained awkward. Google describes Argon as a general physics reasoning model, reporting 87% on the ERQA embodied reasoning benchmark. Reportedly, a robot learned throwing skills from ordinary video and hit targets seven times in a row on the Apollo humanoid.