Google Gemini 4 Argon officially opens: Watch a video and you can teach a robot to toss a sandbag, hitting 7 times in a row.

动察 Beating AI News Flash: Google has officially opened Gemini 4 Argon to paid users worldwide, after previously offering the model only to select cybersecurity partners. The most eye-catching moment was a robot demonstration: Argon watched a video of a person throwing beanbags, identified the venue location, calculated full-body joint movements, and wrote its own physics simulation program. After simulation and real-machine transfer, Apptronik's Apollo humanoid robot, from the U.S. robotics company Apptronik, made 7 consecutive successful throws on the real court.

Google says this process shortens manual simulation modeling that originally took weeks to just a few hours. In other demonstrations, Argon built a simulated environment for Boston Dynamics' Spot robot dog, letting it practice stepping open trash can lids and dropping in cans. It can also generate a simulation scene from a single photo of a tabletop, synthesizing more than 4,500 pieces of training data. After fine-tuning the robot motion model with this data, the real-machine task success rate rose from 56% to 75%, and the time required dropped from 35 seconds to 24 seconds. On the ERQA benchmark, which tests 3D spatial and physical understanding, Argon scored 87%, surpassing other frontier models.

Currently, Google AI Pro and Ultra users can use Argon in the Gemini App, but in Antigravity it is available only to Ultra users, and developers can also access it through the 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 only one-fifth of Claude Fable 5.1 and GPT-6 Astra. However, this discount lasts only until November 30, after which the price doubles, putting it on par with Claude Opus 5.5.