Gemini 4 Argon Debuts for Robotics With Mixed Scores

Gemini 4 Argon scored 87% on ERQA for robotics but trails rivals on AA-Robotics and is mixed on later tests.

Google DeepMind introduced Gemini 4 Argon on Oct. 13, 2026, as a general-purpose physical-reasoning model for robot control, simulation, and training-data review. The company says pure reasoning scores 87% on the Embodied Reasoning (ERQA) benchmark; The Decoder specifies a 400-question test and scores of 82.3% for Claude Opus 5.5, 81% for GPT-6 Astra, and 74% for Claude Fable 5.1. Demonstrations include Boston Dynamics Spot mobile manipulation practiced in MuJoCo, zero-shot arm control, closed-loop steering of the Gemini Omni world model, Real2Sim twins from camera, egocentric, or sketch input, and iterative improvement; The Decoder adds that one photo produced 77 tasks and more than 4,500 episodes, lifting real-robot success from 56% to 75% and cutting execution from 35 to 24 seconds. Sources disagree on overall leadership: Google and early posts say Argon beats other frontier models on ERQA, while The Decoder reports third place on AA-Robotics at 55, behind GPT-6 Astra at 69 and GPT-6.1 Sol at 65. Later coverage is mixed: Argon scored 97.5% on ARC-AGI-3 with a Provider Adapter Harness, second to Astra’s 99.9% and at a reported $4,800 cost, but fell to 49.8% under a stricter Standard Harness and did not finish every one of 55 games; Epoch AI rates it 165, tying Fable 5.1, and estimates eighth in math and sixth in software engineering. On independent StarSkirmish Bench v0.2, Argon, Opus 5.5, and Fable 5.1 clustered at win rates of 37–39 writing Protoss StarCraft: Brood War bots, behind Astra, with Argon showing the widest 18-point run spread.

• Google DeepMind introduced Gemini 4 Argon on 2026-10-13 as a robotics model for control, simulation, and training-data review.
• Argon scored 87% on the 400-question ERQA benchmark; The Decoder lists Claude Opus 5.5 at 82.3%, GPT-6 Astra at 81%, and Claude Fable 5.1 at 74%.
• From one photo, Argon generated 77 tasks and more than 4,500 episodes, raising real-robot success from 56% to 75% and cutting execution from 35 to 24 seconds.
• On Artificial Analysis AA-Robotics, Argon scored 55, behind GPT-6 Astra at 69 and GPT-6.1 Sol at 65.
• On ARC-AGI-3, Argon scored 97.5% with a Provider Adapter Harness at a reported $4,800 cost, versus GPT-6 Astra at 99.9%; a Standard Harness low-reasoning score was 49.8%, and API failures left games unfinished among 55.
• Epoch AI’s Capabilities Index places Argon at 165, tying Claude Fable 5.1, with early estimates of eighth in math and sixth in software engineering.
• On StarSkirmish Bench v0.2, Argon clustered with Opus 5.5 and Fable 5.1 at 37–39 and had an 18-point run spread; GPT-6 Astra led.