Gemini 4 Argon hands-on: strong physics reasoning at low cost
Google's Gemini 4 Argon is positioned as a general-purpose physical reasoning model, and a hands-on review republished by 36Kr says it trades blows with GPT-6 Astra and Claude Opus 5.5 on text and logic benchmarks while costing roughly one-fifth as much and running faster. The piece shows Fourier-series pelican-on-a-bicycle generation, 3D scenes and narrated video, an 87% score on the ERQA embodied-reasoning benchmark, and a demo where Argon learned a beanbag toss from ordinary video and drove an Apollo humanoid robot to seven straight hits. It also flags weaknesses: weak performance on real-world production code, a clunky Antigravity workspace, and limited access for $20-a-month Pro subscribers.
If the physics and embodied-reasoning gains hold at that price, Google could turn its vast YouTube video library into a data moat for physical AI and reset frontier-model price-performance expectations.