Google announced that its Gemini 4 Argon model is now available for developers who subscribe to Google AI Ultra or the paid Gemini API tier. Argon is meant to be a single AI that can write robot control code, build virtual practice worlds and check training videos. The company showed the model making a Boston Dynamics Spot robot toss soda cans after practicing in a MuJoCo simulation, and improving a simple robot arm’s speed from 59 seconds to 17 seconds over three tries.
How Argon works in demos
In the Spot demo, Argon first creates a virtual copy of a trash bin, lets the robot practice balancing on three legs, and only then lets it try the real task. Two extra AI agents watch the robot’s foot and gripper positions before the can is dropped. Another demo starts from a single photo of a table, generates 4,500 simulated training episodes and raises the robot’s success rate from 56 % to 75 % while cutting execution time from 35 seconds to 24 seconds. Argon can also turn a hand sketch of a humanoid robot knocking down boxes into a simulation that calculates a stable throwing motion.
Performance on benchmarks
Google ran Argon on the ERQA benchmark, which asks models to answer 400 physical‑reasoning questions without extra tools. Argon scored 87 %, beating Claude Opus 5.5 (82.3 %) and GPT‑6 Astra (81 %). In a different test, the AA‑Robotics benchmark, which lets language models directly control a real robot arm, Argon placed third with 55 points, behind GPT‑6 Astra (69) and GPT‑6.1 Sol (65).
Why it matters
For robotics developers, Argon promises to reduce the amount of hand‑crafted simulation work by generating training data from a single image or sketch. That could speed up prototyping of new robot tasks, though the model still needs careful prompting and is only accessible through a paid subscription.
