October 13, 2026
gemini-4-argon: Google's model for robots costs $2 per million tokens until Dec 1
Argon is already available in Antigravity and the Gemini API. From Dec 1, a million input tokens will go up from $2 to $4.
A two-armed Franka FR3 now succeeds in 75% of attempts instead of 56%. Its training data was generated by the model.
From a single top-down snapshot, Argon builds a physical replica of the scene, writes its own motion generator with a pose spread of ±10 cm and checks every attempt in a simulator. The output is 4,500 fine-tuning episodes, and the time per operation on the real robot dropped from 35 to 24 seconds.
These episodes used to be recorded by human operators, one at a time. Now the model prints them, and the hardware learns on synthetic data.
Where to turn it on. Argon is open through three doors at once: Google AI Studio with an AI Ultra subscription, the Gemini API on a paid tier, and Google Antigravity, where Argon can be set as the active model for planning and multi-agent runs. In the API the model is called gemini-4-argon, and the google-genai SDK has supported it from day one. The model is not available on the free API tier or in the Google AI Pro plan.
A price with a date. Until Nov 30, a million input tokens cost $2 and output tokens $10, with output counted together with thinking. From Dec 1 the price doubles to $4 and $20. Cached input goes from $0.10 to $0.20.
The model's window holds a million input tokens and up to a million output tokens. On ERQA, which tests understanding of the physical world, Google claims 87% and calls it the best result among frontier models.
Argon drives hardware without fine-tuning: Boston Dynamics' Spot quadruped throws an object along a trajectory the model synthesizes at 50 Hz. On a six-axis arm, a self-improvement loop sped up a pick-and-place task from 58.9 to 17.4 seconds in three rounds.
The next milestone is Dec 1, 2026, when Argon's pricing doubles.