# Google says Gemini 4 Argon scores 87% on its own embodied-reasoning benchmark and found new fusion-plasma solutions

The model became generally available in the Gemini API on October 13, and Google is pitching it as a single model for robot control, simulation building and scientific research, though no independent reliability data exists yet.

- Published: 2026-10-16T06:26:08.965Z
- Canonical: https://polylog.news/ai/2026-10-16/google-says-gemini-4-argon-scores-87-on-its-own-embodied-rea
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Polylog editors](https://polylog.news)

Google DeepMind reports that [Gemini 4 Argon scored 87% on the Embodied Reasoning Question Answering (ERQA) benchmark](https://t.me/ai_machinelearning_big_data/11140), a test of physical-world reasoning for robots, without using external tools. Google says that result beats every other frontier model. The [Google Developers Blog](https://developers.googleblog.com/gemini-4-argon-for-robotics/) presents Argon as "a general-purpose physical reasoning model" with four functions: robot control, Real2Sim (turning real footage into physics simulations), physical understanding of video, and research and self-improvement.

The benchmark needs context: Google DeepMind created ERQA itself when it [launched Gemini Robotics](https://arxiv.org/pdf/2503.20020) in 2025, so the 87% figure is a vendor result on a vendor-built test, not an independent replication. The demonstrations follow the same pattern. According to [The Decoder](https://the-decoder.com/google-wants-gemini-4-argon-to-be-the-only-ai-model-robots-need/), Google showed mobile manipulation practiced in MuJoCo, zero-shot robot control, and simulation twins rebuilt from robot camera feeds, egocentric human video or hand-drawn sketches. Google also says Argon helped Apptronik's Apollo humanoid [land seven consecutive cornhole shots](https://humanoid.guide/google-demonstrates-gemini-4-argon-robotics-on-apollo-humanoid/) after watching video of a person playing. One review described reliability as [unclear](https://superpowerdaily.com/posts/google-opens-gemini-4-argon-to-paying-users-with-tools-for-robot-control-and-training).

The research claim concerns stellarator physics. According to [AI Post](https://t.me/aipost/8493), Argon agents helped researchers find new exact solutions for 3D magnetic plasma configurations, with promising candidates appearing within hours. Google says its team and University of Maryland professor Matt Landreman [found a broader class of counterexamples](https://interestingengineering.com/ai-robotics/google-gemini-4-argon-quantum-fusion-robotics) to Grad's conjecture. That 1967 hypothesis held that smooth 3D plasma equilibria require special symmetries. The solutions can serve as reference cases for checking stellarator simulation codes.

Credit for the September results is contested. Javier Gómez-Serrano, Lukas Liehr and Mitchell Taylor posted [rigorous counterexamples](https://www.theneuron.ai/news/gpt-6-astra-fusion-math-breakthrough-grads-conjecture/) on September 21. The Neuron reported that Landreman's explicit families from September were found with help from OpenAI's GPT-6 Astra Pro. Google's announcement describes a further class found with Argon. Readers should treat the Argon contribution as an extension of work already underway, not as the original breakthrough.

## What this means

Google wants robotics companies to treat one frontier API as the perception, planning and simulation layer instead of training separate vision-language-action models. If developers accept that, Google gets recurring API revenue from hardware makers such as Apptronik, and specialist robotics model startups lose their reason to exist. What is verified so far is a vendor score on a vendor benchmark plus curated demos. Success rates over many trials have not been published. The research claim points the same way, since two labs are now claiming AI-assisted progress on the same plasma problem within weeks, which makes the result partly a public dispute over scientific credit. The likelier outcome over the next few months is that robotics firms use Argon for Real2Sim data generation and offline planning, while closed-loop control stays with smaller on-robot models. The reason is latency: a separate paper in this edition shows reasoning models losing to faster models when the environment does not pause. That read would be proven wrong if Google or a partner publishes multi-trial success rates for real-time manipulation driven directly by Argon at production volume.

## What to watch

- Whether any third party reproduces Argon's ERQA score or publishes success rates over many trials for its robot control. Either would turn a vendor demo into evidence.
- Whether the Landreman and Google counterexamples appear as a refereed paper with clear attribution between Gemini and GPT-6 Astra Pro, which would settle the credit dispute.
- Whether humanoid makers announce production contracts built on Argon rather than research collaborations, which would show the single-model robotics pitch has buyers.