Google DeepMind has shown that its new Gemini 4 Argon model can teach robots physical skills by watching human video. The company calls this approach “Video-to-Sim-to-Real.” It compresses weeks of manual simulation modeling into just a few hours. Google first announced Argon on September 30 and then released it globally on October 13, 2026.
In one demonstration, Argon watched a video of a person tossing beanbags into a cornhole board. It then mapped the arm motion and release point onto a humanoid robot’s joints. After that, it built a full physics simulation in MuJoCo, and the robot practiced until the throw was clean. As a result, Apptronik’s Apollo humanoid sank seven consecutive shots in a real-world trial with zero prior task-specific training.
The model also works with quadruped robots. In another demo, Argon controlled a Boston Dynamics Spot with a six-degree-of-freedom arm. The robot stepped on a pedal bin, held a soda can, used its wrist camera to center the can, and dropped it in. Before attempting this physically, Argon built a MuJoCo simulation to test the robot’s balance on three legs.
Furthermore, Argon can create digital twins from a single photograph. One snapshot of a table generates a full simulation environment with over 4,500 domain-randomized training episodes. This approach raised real-world task success rates from 56% to 75% and cut completion time by 11 seconds. On top of that, the model scored 87% on the Embodied Reasoning benchmark, which Google says outperforms every other frontier AI model.
Argon uses a sparse mixture-of-experts transformer architecture with a unified 1,048,576-token context window. Google offers it at an introductory price of $2 per million input tokens and $10 per million output tokens through November 30. After that, the price doubles to $4 and $20 respectively.
Beyond robotics, the model has already achieved results across multiple fields. It scored 77.9% on the DeepSWE software engineering benchmark, which sets a new state of the art. It also ranked first on the Vals Index for finance, coding, legal, and tax work. Additionally, it scored 91.7% on LVBench for long-video understanding and 68% on CWE-bench for vulnerability remediation.
Google’s own teams have used Argon extensively as well. The Chrome team used it to refactor a legacy folder inside Chromium’s roughly 40-million-line codebase. That task would have taken months manually. Meanwhile, other teams are migrating the 800,000-line Fuchsia Zircon kernel from C and C++ to Rust using Argon agents. In one case, the model replaced 32,000 lines of SIMD code and made the libgav1 video decoder run 2.7 times faster.
On the cybersecurity front, Google’s partner Wiz used Argon to find a critical healthcare vulnerability. That flaw exposed a master key and granted administrator access to systems holding hundreds of thousands of patient records.