Google releases Gemini 4 Argon, a frontier AI model trained from scratch on Tensor Processing Units

Google has released Gemini 4 Argon, a new frontier AI model developed from scratch over a six-month period using Tensor Processing Units (TPUs) and the JAX Python library. The model utilizes a sparse mixture-of-experts transformer architecture and features a knowledge cutoff of March 2026. In performance evaluations, Argon achieved top rankings on the DeepSWE software engineering benchmark and the Vals Index, which measures accuracy across finance, legal, and coding sectors based on their share of the US GDP. It also demonstrated high proficiency in long-video analysis, scoring 91.7% on LVBench. Google is positioning Argon as a cost-effective alternative to competitors like Claude 5.5 and GPT-6 Astra, offering introductory pricing of $2 per million input tokens and $10 per million output tokens through November 2026. Despite its benchmark success, the model shows varying results in subjective human rankings, placing first in text tasks but falling to ninth in web development coding. Technical tests involving Three.js 3D rendering revealed that while Argon produces a sophisticated user interface, its 3D modeling accuracy may lag behind rivals in specific geometric tasks. The model was initially rolled out to a select group of cybersecurity professionals via the Fairwind Program before becoming available to Google AI Ultra subscribers.

Tech4K StaffVideo report

Google has released Gemini 4 Argon, a new frontier AI model developed from scratch over a six-month period using Tensor Processing Units (TPUs) and the JAX Python library. The model utilizes a sparse mixture-of-experts transformer architecture and features a knowledge cutoff of March 2026. In performance evaluations, Argon achieved top rankings on the DeepSWE software engineering benchmark and the Vals Index, which measures accuracy across finance, legal, and coding sectors based on their share of the US GDP. It also demonstrated high proficiency in long-video analysis, scoring 91.7% on LVBench. Google is positioning Argon as a cost-effective alternative to competitors like Claude 5.5 and GPT-6 Astra, offering introductory pricing of $2 per million input tokens and $10 per million output tokens through November 2026. Despite its benchmark success, the model shows varying results in subjective human rankings, placing first in text tasks but falling to ninth in web development coding. Technical tests involving Three.js 3D rendering revealed that while Argon produces a sophisticated user interface, its 3D modeling accuracy may lag behind rivals in specific geometric tasks. The model was initially rolled out to a select group of cybersecurity professionals via the Fairwind Program before becoming available to Google AI Ultra subscribers.

Gemini 4 Argon was trained from scratch over six months using Google's Tensor Processing Units and the JAX library. The model uses a sparse mixture-of-experts transformer architecture rather than being a fine-tune of previous versions.

Argon achieved a 91.7% score on the LVBench benchmark for long-video understanding and multimodal tasks. Introductory pricing for the model is set at $2 per million input tokens and $10 per million output tokens.

The model's knowledge cutoff is March 2026, suggesting a training completion date shortly thereafter. Argon ranks first on the Vals Index, which weights performance based on US GDP contributions from various industries.

Subjective human evaluations on Arena AI place Argon at the top for text tasks but significantly lower for web development.

Chapter guide

Introduction to Gemini 4 ArgonOverview of the new model release and its context in the competitive AI landscape.Naming and BackgroundDiscussion of the 'Argon' name and the model's initial rollout via the Fairwind Program.Architecture and TrainingTechnical details on the sparse mixture-of-experts transformer and TPU-based training.Benchmark PerformanceAnalysis of Argon's scores on DeepSWE, Vals Index, and LVBench compared to competitors.Pricing and AvailabilityDetails on introductory token costs and the Google AI Ultra subscription requirement.Multimodal and 3D CapabilitiesTesting Argon's ability to generate video and render interactive 3D graphics using Three.js.Macroscope SponsorshipOverview of the Macroscope code review tool and its automated pull request approval features.

Worth noting

  • The video is set in a fictionalized future (October 2026) and references hypothetical models like GPT-6 Astra and Claude 5.5.
  • The 'TrustMeBro' benchmark mentioned is a satirical reference and not a real industry standard.
  • The video includes a paid sponsorship for the code review tool Macroscope.
  • The 3D rendering test results are subjective and presented with a comedic tone.

Watch the original video ↗