Google Gemini 4 Argon Officially Launched: Teach Robots to Throw Sandbags via Video, Scoring 7 Consecutive Hits

Insight Beating AI Flash News: Google has officially rolled out Gemini 4 Argon to all global paid users, after the model was previously only available to select cybersecurity partners. The most eye-catching demo features a robot use case: Argon watches a video of a human throwing sandbags at a target, identifies the field location, calculates full-body joint movements, and writes its own physics simulation program. After simulation and real-device migration, Apollo, the humanoid robot from U.S. robotics firm Apptronik, successfully hits the target 7 times in a row on the actual field. Google notes this workflow cuts manual simulation modeling—originally taking weeks—down to just hours. In other demos, Argon built a simulation environment for Boston Dynamics’ Spot robot dog, letting it practice opening trash can lids and tossing in cans. It can also generate a simulation scene from a single photo of a desk, synthesizing over 4,500 training data segments. After fine-tuning the robot action model with this data, the real-device task success rate rose from 56% to 75%, while time per task dropped from 35 seconds to 24 seconds. In the ERQA benchmark, which tests 3D spatial and physical understanding capabilities, Argon scored 87%, outperforming other leading models. Currently, Google AI Pro and Ultra users can access Argon in the Gemini app; access in Antigravity is limited to Ultra users only, and developers can also call it via a paid API. The API launch price is $2 per million input tokens and $10 per million output tokens, on par with Claude Sonnet 5.5 and GPT-6.1 Sol, and just one-fifth the price of Claude Fable 5.1 and GPT-6 Astra. However, this discount is only valid until November 30, after which prices will double, matching Claude Opus 5.5’s rates.