Google releases Gemini 4 Argon, a physical-reasoning model for robot control and simulation

The model scores 87% on a reasoning benchmark, but Google's robotics results rely on heavy task-specific setup and few published failure rates.

Google has released Gemini 4 Argon, which it positions as a general-purpose physical-reasoning model covering robot control, simulation construction from video, labeling of robot footage, and iterative policy improvement rather than a single robot policy. The company reports 87% on the ERQA embodied reasoning benchmark using pure reasoning without helper tools, and says that leads other frontier models on the test. In one case, Google synthesizes domain-randomized simulation episodes from a single photo to fine-tune a vision-language-action policy, reporting a rise from 56% to 75% success on the matching real setup. Access runs through Google AI Ultra and the paid Gemini API under the model id gemini-4-argon. Google's robotics examples rely on long task-specific prompts and scaffolding, and it does not publish per-demo failure rates beyond that single case. Separately, Google DeepMind's earlier Gemini Robotics 2.5 release reported tasks completed in Mercedes-Benz factories and a humanoid tying a bow, though those claims concern a different model generation.