Gemini 4 Argon for Robotics: the 87% ERQA score and the 56% to 75% pick-and-place gain, checked

Google's Gemini 4 Argon for Robotics, published October 13, 2026 on the Google Developers Blog, reports an 87% score on the ERQA embodied reasoning benchmark and a real-robot pick-and-place success rate that rose from 56% to 75% after training on simulation-generated data. Both figures are Google's own, and the post offers no independent run.

What Google released

The post, bylined to Kanishka Rao, a principal software engineer, presents Argon as a general-purpose physical reasoning model for the robotics lifecycle. It covers robot control, Real2Sim (turning real scenes into simulations), physical understanding, and research and self-improvement use cases.

Access is split. In Google AI Studio and Google Antigravity it is for Google AI Ultra subscribers. Through the Gemini API it is on the paid tier, using the model name gemini-4-argon in the Interactions API, with day-one support in the google-genai SDKs. The main model launch, covered in Google's announcement by Koray Kavukcuoglu, lists introductory pricing of $2 per million input tokens and $10 per million output tokens, rising on November 30, 2026 to $4 and $20. We covered that change in our note on the Argon price rise.

The two headline numbers

ERQA at 87%. The post says: "Using pure reasoning, Argon scores 87%, outperforming every other frontier model." It does not, in the text I could read, list the other models' scores, name the evaluation date, or link a leaderboard. The Gemini Robotics paper on arXiv (March 25, 2025) is where Google introduced Gemini Robotics-ER, its embodied reasoning model. The abstract page I opened does not mention ERQA, so I cannot describe the benchmark's size or categories from a source I read. The practical point: "outperforming every other frontier model" is a vendor claim with no visible table.

56% to 75%. On the SO-101 pick-and-place task, the post says real-world success rose from 56% to 75% and completion time fell from 35 to 24 seconds, using data generated in simulation. That is a more concrete result. A 19-point gain on a physical robot is worth noting, and the time saving is a useful second signal. But SO-101 pick-and-place is a simple task, and the post does not give trial counts, so the confidence interval is unknown. With a small number of trials, a 19-point swing can fall within noise.

Why the simulation angle is the real story

The more interesting claim is the pipeline. Per the post, Argon built simulations from real-world inputs and that synthetic data improved real performance. That is the Real2Sim loop robotics teams have wanted, because real-robot data is slow and expensive and simulated data is cheap but often does not transfer. If a model can write the simulation, check it against reality and iterate, the data bottleneck eases.

The main announcement gives a flashier example. In an internal test, a humanoid robot sank seven consecutive cornhole shots after Argon built its simulation from video. That is a demo. It is a vivid one, but seven makes is a count of successes in one test and does not tell you the miss rate over a long run.

The post also describes constraints in its prompts, such as rules to avoid "sim cheating" and safety rules referenced from an AGENTS.md file. It does not explain them. That detail is worth keeping in mind, because a model that builds its own simulator has an obvious way to flatter itself, by making the simulation easier than the world. The fact that Google felt the need to write rules against it is informative, and the post does not show how well they hold.

The "physical AGI" line

The closing section states: "Gemini 4 Argon marks an important milestone accelerating the path toward physical AGI." As a forecast, it is not evidence. Our reporting on a separate test found how far frontier models are from reliable robot work: see our coverage of the RobotWorld benchmark, where frontier models solved a minority of tasks. Different benchmarks, different setups, so the two are not directly comparable, but the contrast is a fair check on the rhetoric.

What to look for next

• Third-party ERQA runs, with other models listed side by side
• Trial counts and variance for the SO-101 result
• A test on more than one robot and more than one task family
• A safety section covering what happens when an agent writes control code that moves hardware

How to read a robotics benchmark claim

Robotics numbers need more context than language-model scores. Ask what the robot, the objects and the lighting were, how many trials were run, and whether the test conditions matched training. A pick-and-place task with known objects on a clean table is the easiest family of manipulation problems, so success there says little about clutter, soft objects or long multi-step jobs. The 35 to 24 second improvement is useful because it measures something other than a pass rate, but it too depends on the trial count. Until Google or an outside group publishes those, treat the gain as encouraging evidence for the simulation approach, not as a measure of general capability.

Our take

This is a sensible, narrow robotics release wrapped in a large claim. The simulation-to-real result is the part worth following, because it addresses a real bottleneck, and 56% to 75% on a physical arm is a respectable first data point. The 87% benchmark line and the physical AGI sentence are marketing until someone else reproduces them. If you build robots, try the API on your own simulator and measure transfer on your hardware. If you track the race, file this as a lab claiming the lead in embodied reasoning, with the evidence still to come.