• The verdict is Wait: the 87% ERQA score comes from Google's own post and has not been independently reproduced.
• On Artificial Analysis's AA-Robotics v0.1, a real 6-axis arm across eight tabletop tasks, Gemini 4 Argon scores 55.3, behind GPT-6 Astra at 69.4 and GPT-6.1 Sol at 64.6.
• Google's post says simulation data raised the party-props success rate from 56% to 75% and cut completion time from 35 to 24 seconds, but it gives no number of real-world attempts.
• The Spot demo is a trash-bin task, and its published prompt invokes a ready-made 'spot-mobile-manipulation skill', so the model is not starting from nothing.
• The most testable parts for engineering teams are simulation generation and scoring training data, where errors are cheap.

The verdict on Gemini 4 Argon is Wait. Google's launch post claims 87 percent on the ERQA embodied-reasoning benchmark and says that outperforms every other frontier model. The independent number is weaker: Artificial Analysis's AA-Robotics v0.1 puts Argon third at 55.3 of 100. The pitch of one model for control, simulation and data evaluation is plausible, but the published evidence does not yet show it. This piece reads the primary pages; we did not run the model ourselves.

The 87 percent is a vendor number

Google's post states that, using pure reasoning, Argon scores 87 percent on ERQA. The Decoder reports the chart's comparison scores as 82.3 percent for Claude Opus 5.5, 81 percent for GPT-6 Astra and 74 percent for Claude Fable 5.1, on 400 questions without helper tools. The benchmark repository is public, but the run and the model selection are Google's. Until a third party reproduces the score, treat it as a vendor claim.

The independent robot test is less flattering

Artificial Analysis describes AA-Robotics v0.1 as a research preview in which each model controls a physical 6-axis arm, identified as a Hiwonder NexArm, through eight tabletop tasks. It reports GPT-6 Astra at 69.4, GPT-6.1 Sol at 64.6 and Gemini 4 Argon at 55.3. The test is early and narrow, but it is the only hardware result we found that Google did not run itself.

The demos come with scaffolding

Google publishes full prompts for each demo. The Spot prompt begins by invoking a 'spot-mobile-manipulation skill', so a pre-built capability already exists. According to The Decoder, the cup-righting demo supplies a camera image with a centimeter-grid overlay and commands such as grasp and lower until contact, and the simulation prompts carry long prohibition lists. That shows the demos need engineering support. It does not show how the model performs without it.

Simulation gains lack an attempt count

For the party-props demo, Google's post says simulation data lifted real-world success from 56 percent to 75 percent and cut completion time from 35 to 24 seconds. It does not state how many real attempts sit behind those rates. The 4,500-plus episodes it cites are simulated. A percentage without a denominator is a lead, not evidence, so reproduce it before relying on it.

The bottom line: Wait

Wait. Test Gemini 4 Argon in simulation and on training-data scoring, where mistakes are cheap, using your own tasks and logging every prompt aid you add. Do not move it into closed-loop control on real hardware until a third party reproduces the ERQA result and its AA-Robotics score of 55.3 improves. Google's post lists access through Google AI Studio and the paid Gemini API tier, so a bounded trial is possible. Re-check when AA-Robotics or an independent group publishes new runs.