Gemini 4 Argon scores 87% on embodied reasoning but takes 52 minutes per real-arm task

Google released Gemini 4 Argon worldwide on Oct 13, and Physical Intelligence opened the beta of an API that fine-tunes its π0 and π0.5 robot models on infrastructure Pi runs.

Google is entering robotics as the slow layer around the machine, building simulations, labelling data, grading progress and writing control code, while Pi's API is a bid to own the customer's collect-train-test loop before general models make that loop cheap. The disagreement is falsifiable: if a general model gets fast enough to close the control loop itself, the specialist position shrinks to deployment work.

Argon's jobs sit beside the control loop

Google's developer post lists four robotics uses for Argon: robot control, Real2Sim, physical understanding such as hindsight labelling and progress estimation, and research and self-improvement. Most of Google's examples are the slow work. Google says Argon watched a raw video of a person playing cornhole, mapped the environment, calculated whole-body joint movements and wrote simulation code; on Apptronik's Apollo humanoid it then sank seven consecutive shots. Google says its Video-to-Sim-to-Real pipeline compresses weeks of manual simulation modelling into a few hours. From a single RGB photo of a tabletop, Google says, Argon built a calibrated MuJoCo twin, defined 77 tasks and generated more than 4,500 domain-randomised episodes for VLA fine-tuning, lifting real-robot success from 56% to 75% and cutting completion time from 35 seconds to 24. On a 6-DoF SO-101 arm, Google says pick-and-place fell from 58.9 seconds to 17.4 seconds across three iterations.

Humanoids Daily reports that the cornhole workflow took feedback from physical attempts on the real robot; if the streak or the 75% figure came after many uncounted tries, both say less about reliability than seven clean shots suggest. The control demo makes the same point: Google's "zero-shot control" mug task hands Argon a wrist-camera image with a projected 2 cm world-coordinate grid and pre-built primitives such as grasp and place_until_contact, so the model chooses among primitives instead of commanding joints. Google also already shipped Gemini Robotics-ER 1.6 for spatial perception and task planning.

Frontier models run about 100 times slower than a person

Artificial Analysis's AA-Robotics v0.1, published Oct 14, drove a real 6-axis Hiwonder NexArm, five joints plus a gripper, through a simple HTTP API with four cameras and eight tabletop tasks, zero-shot. On a 100-point scale, GPT-6 Astra scored 69.4, GPT-6.1 Sol 64.6, Gemini 4 Argon 55.3, Claude Opus 5.5 48.8 and Claude Fable 5.1 46.3.

The timing is the sharper result. Artificial Analysis measured the models at roughly 100 times slower than a human, about half an hour per task against a 20-second human average. Argon was slowest at 52 minutes, or 3,120 seconds, about 156 times the human benchmark; the 30-minute average is 90 times. Argon also spent an average of 21 minutes observing and reasoning before its first action, nearly twice as long as any other model. Artificial Analysis puts cost per task at $16.36 for GPT-6 Astra and $1.93 for GPT-6.1 Sol, and says frontier language models are "far too slow and expensive for any kind of production usage" today while expecting that to change quickly.

Google's 87% on ERQA, its Embodied Reasoning benchmark of 400 multiple-choice questions about spatial reasoning, sits beside those numbers. Google says that result comes from pure reasoning and beats every other frontier model; runner-up Claude Opus 5.5 scored 82.3% on Google's chart. ERQA measures spatial reasoning, not whether a robot completes a task, and Argon placed third once Artificial Analysis put this class of model in front of a real arm.

Pi sells the loop, and the loop is the wager

Pi's beta hands over an iteration loop, not a controller. Developers upload data, launch fine-tuning of the π0 and π0.5 family on compute and infrastructure Pi manages, and run the result through a hosted inference gateway. Pi describes the loop as collecting demonstrations, training a model, testing it on the robot and using the failures to choose the next data, and writes, "We built the API to let developers run this iteration loop themselves, using our latest models." Pi says it plans to keep enabling access to its latest models through the API, and adds, "The more our models are used, the more we learn how to improve them."

The early proof is small but named. Pi says Almond's dual-arm Axol robot deburred spacers used in its own assembly with a model obtained after four days of iterating through the API, and that Actor's autonomous wheel loader scooped material into a dump truck in a commercial deployment in Utah with no operator inside. Pi lists AgRobotics, CloudChef, a dairy partner, Ultra and Weave among the others. Most of the applications ran on cloud-hosted inference through the API, and Pi names network connectivity as a challenge for some.

The strongest case against the split sits in Google's own post. Robot control is the first of its four areas, Argon writes control code, and the SO-101 result is self-improvement on the very loop Pi sells. If the general model gets fast and cheap enough to run that loop on its own, hosted iteration becomes a commodity and Pi is left selling deployment work. Argon finishing third behind two GPT-6 models shows the generalist race is still open, so the fast-controller scenario may arrive from a competitor. The number to watch is latency in the next AA-Robotics release: if per-task time drops from 52 minutes toward the 20-second human bar, the case for a separate training-layer company weakens with it.