Google's Gemini 4 Argon targets full robotics pipeline: 87% on ERQA, but third on AA-Robotics at 55/100

Google DeepMind is pitching Gemini 4 Argon as a single model for the full robotics pipeline: controlling robots, generating simulated training environments, and evaluating training data. Google claims Argon scores 87% on the ERQA spatial-reasoning benchmark, the highest of any frontier model tested. An independent test published a day later ranks it third.

What Google showed

Google's demos follow one pattern: Argon uses simulation to prepare, verify, or generate data for a task.

• Spot trash-bin task: Argon builds a virtual copy of a step-on bin in a MuJoCo simulation, and a Boston Dynamics Spot robot practices balancing on three legs before attempting the real task. Two additional AI agents verify foot and gripper position before the can is dropped.
• Photo to training data: From one photo of a table with party props, Argon built a virtual replica and proposed 77 tasks. It generated more than 4,500 training episodes for a control model. According to Google, real-robot success rose from 56% to 75% and execution time fell from 35 to 24 seconds.
• Other inputs: Teleoperation recordings, first-person human grasping video, and a ballpoint-pen sketch can all be turned into simulations. In the sketch demo, a humanoid throws a ball to topple six boxes without falling.
• Directing a video world model: Argon writes timestamped instructions for Gemini Omni, a video world model. It checks the generated frames and adjusts the next command. Five scattered Scrabble tiles were arranged to spell "ARGON" in 33 steps.
• Self-improvement: In a Lego-brick task with a simple robot arm, Argon reviewed its own logs and camera images between attempts. Execution time fell from 58.9 to 17.4 seconds over three tries.
• Research and data grading: Google says 68 browser-based sandboxes recreating published robotics experiments have been built so far. Argon also segments videos into sub-steps, such as a dual-arm robot folding an origami fox, and flags pauses and botched grips so those segments are weighted less in training.

Benchmarks

On ERQA (400 questions, no helper tools), Google reports:

Model	ERQA score
Gemini 4 Argon	87%
Claude Opus 5.5	82.3%
GPT-6 Astra	81%
Claude Fable 5.1	74%

These figures come from Google. On AA-Robotics from Artificial Analysis, where language models directly control a real robot arm across eight tasks, Argon scored 55 out of 100. GPT-6 Astra scored 69 and GPT-6.1 Sol scored 65.

Caveats in the demos

Google published the example prompts, and they show substantial scaffolding. The "zero-shot" cup-righting demo gave Argon a camera image with a centimeter grid overlay and pre-built commands such as "grasp" and "lower until contact." The simulation prompts include long prohibition lists to stop the model from taking shortcuts, such as copying measured sensor values. Google does not give failure rates for individual demos, and it does not say how many attempts underlie the 75% success figure.

Availability

Argon is available in Google AI Studio and the Antigravity development environment, but only for Google AI Ultra subscribers. It is also available through the paid tier of the Gemini API. Context window, per-token pricing, and parameter count were not disclosed in the source material. Google had already released Gemini 4 Argon globally days earlier, and Gemini Robotics-ER 1.6 is an earlier robotics-specific model.

What this means

The ERQA and AA-Robotics results measure different things. ERQA tests reasoning about physical scenes in question form. AA-Robotics tests closed-loop control of real hardware. Argon's lead on the first and third-place finish on the second suggest its strength is spatial understanding and simulation generation rather than direct control, which is where the OpenAI models lead.

The more significant claim is the simulation-to-real data loop. The 56% to 75% improvement is the most concrete evidence for it, but without attempt counts or baselines it is hard to assess. The published prompts show that production use will likely require similar scaffolding. Teams evaluating Argon should test it on their own hardware and tasks, not rely on the demo results.