Google's newest model, Gemini 4 Argon, is also its pitch to robotics teams: a general-purpose physical reasoning model for the whole robotics stack, available to paid Gemini API users and Google AI Ultra subscribers. The robotics case is written up by Google principal software engineer Kanishka Rao. Google says Gemini 4 Argon scores 87% on the Embodied Reasoning (ERQA) benchmark using reasoning alone, ahead of every other frontier model.
ERQA tests whether a model can solve spatial problems in 3D environments, and that score is the headline claim. Google published no side-by-side table of rivals, so the comparison is its own. What the company does show is range: demos grouped under robot control, simulation building from real footage, physical understanding, and self-improvement.
The simulation work leans on the model's coding. Per Google, it can turn a robot teleoperation recording, an egocentric human video or even a photographed pen sketch into a runnable simulation, including a 76-degree-of-freedom humanoid throwing a ball at a stack of boxes. On the data side, Google reports that simulated episodes generated by the model, more than 4,500 of them, raised a real robot's task success rate from 56% to 75% and cut completion time from 35 to 24 seconds. Most demos come with the prompt Google used, some as templates, which makes this a rare vendor post that others can try to reproduce.
The control demos are the most concrete. Google says the model drove a Boston Dynamics Spot to step on a foot-pedal trash bin after practicing the move in a MuJoCo physics twin it built itself, and steered a simple robot arm to stand a fallen mug upright and drop a napkin inside. In another, it directed a video world model, Gemini Omni, to rearrange Scrabble tiles into the word ARGON by writing and refining timestamped prompts from the footage it got back.
Labeling is a quieter use. Given a 136-second overhead video of a bimanual ALOHA rig folding origami, Google says the model produced a four-level annotation hierarchy and flagged fumbles and retries so training could weight them down, which is the kind of cleanup robot-learning teams still do by hand.
Access is split by plan. Google AI Studio and Google Antigravity are open to Google AI Ultra subscribers, while the Gemini API offers the model on its paid tier under the ID gemini-4-argon in the Interactions API, with day-one support across the google-genai SDKs. The post lists no per-token pricing, and every performance figure in it comes from Google's own tests, so developers evaluating hardware work will want to rerun their own tasks before trusting the 87%.
