Gemini 4 Hydrogen scores 58 on a robot arm test, Google says
Google says Hydrogen controls a real robot arm in eight tabletop tasks and scores 58, almost the same as larger Argon at a lower cost.
Google launched a Robotics plugin for Antigravity, its coding tool. Google says a developer can go from a new robot to a working application in an afternoon. The plugin pairs with Gemini 4 Hydrogen, a fast model that Google introduced last week. Google says Hydrogen does better than its Gemini Robotics ER 2 models on almost all axes.
How it works
You give Hydrogen a goal in plain words and a camera view. It plans the task, writes and runs control code, and adjusts until the work is done. The plugin gives skills for robot setup, simulation and control.
The numbers
On the Artificial Analysis Robotics 0.1 test, Hydrogen controls a real arm with six axes through eight tasks, such as a pour of a drink. It scores 58, almost the same as Argon. Google says it also does better than ER 2 on the public ERQA test of spatial reasoning.
What is not known
Google gives no cost numbers for the robot work. It does not say how Hydrogen does outside tabletop tasks. All claims come from Google.
Insight: One cheap model is spreading into every Google tool
In November 2026, Hydrogen was already in GitHub Copilot and in the Antigravity coding tool, at $1.50 for each million input tokens. It scored 51 on a general index, above larger Argon at 50. Now it drives robots. This seems a move to make one low cost model the default across code, agents and robots, so rivals must match its price. Compare it with your model by cost for each finished task.
Sources
- Gemini 4 Hydrogen: a faster model for general purpose roboticsdevelopers.googleblog.com
Posted