Google reports 87 percent for Argon on a robotics reasoning test

Watch

Lab claim

On October 13, 2026, Google posted a robotics note for Gemini 4 Argon. Google says pure reasoning scores 87 percent on Embodied Reasoning QA.

How to read this claim

Measurement — Other models' ERQA scores are not printed next to the 87 percent in the claim. There is no robot to deploy from the…
Bill — List input and output per million tokens on the vendor page — a loop pays both again.
Door — Status on this card: Shipped.

Keep the habit · field card with prices and worked checks

What changed

Argon had been a price and an API door. This post adds a robotics reasoning score, 87 percent on ERQA, measured by Google.

What it introduces

A lab measurement of physical reasoning for a model that is already on the paid API. The robot sequences are demonstrations.

Who benefits Anyone following robot policies, and anyone who might confuse a model demo with a system they can deploy.

Try this

Keep the API price and this 87 percent in separate sentences. Do not plan a robot deployment from the Spot clip. The score is Google's.

What we do not know

Other models' ERQA scores are not printed next to the 87 percent in the claim. There is no robot to deploy from the post.

Evidence

Primary Google DeepMind. Reported by Google. The ERQA score is Google's. The robot sequences are demonstrations in the post.

Independent The 87 percent ERQA figure is Google's. The post says it is ahead of other frontier models and does not print those other scores in the opening claim.

Sources

Google DeepMind Primary · October 13, 2026

Still unclear? · Start here · ranks, prices, unknown Other models' ERQA scores are not printed next to the 87 percent in the claim. There is no robot to deploy from the…

The record

What happened

On October 13, 2026, Google published Gemini 4 Argon for Robotics. The post says that using pure reasoning, Argon scores 87 percent on the Embodied Reasoning QA benchmark and that this is ahead of other frontier models. It shows prompts for a Boston Dynamics Spot practicing a trash-bin task in MuJoCo, and for a tabletop arm. Those sequences are demonstrations in the post.

Why it matters

Argon was already the price story. This page is the physical-reasoning claim that came with it. Eighty-seven percent is a number a reader can remember, and it belongs to Google. Nothing on the page is a robot runtime you install. Odyssey-3 is the other preview in that position this week.

The post also describes real-to-sim reconstruction and closed-loop steering of video world models. Those are workflows in the writeup, not a shipped robot stack with a price.