Hardware Ready Ready to Bench?
Google has officially announced the global debut of Gemini 4 Argon, including in Italy, a new artificial intelligence model designed to excel in deep work and the management of complex multi-stage tasks. Designed to support professionals, developers, and researchers in high-complexity fields such as programming, finance, law, cybersecurity, and science, the system ranks first on the Vals Index, surpassing major competitors in sector tests and launching a direct challenge to rivals like GPT-6 Astra and Claude Opus 5.5.
The release is active from these hours for developers and businesses via the Gemini platform APIs in AI Studio and Android Studio, as well as for subscribers to Google AI Pro and Ultra plans through the Google Antigravity environment, the Gemini application, and the integrated AI mode in the search engine.
Gemini 4 Argon launch: why it matters
At the core of Gemini 4 Argon’s operational capabilities is a unified window capable of handling up to 1 million tokens between input and output via API. The service is offered at a launch price of $2 per million input tokens, with a 95% discount for cached tokens, and $10 per million output tokens. These capabilities have already generated measurable impact within Mountain View’s software development workgroups.
The use of Argon on Google Antigravity allowed for an increase of over 35% in code delivery via autonomous agents, while simultaneously reducing rollbacks due to bugs or malfunctions by nearly one-third. A concrete case involves the Chrome browser, where the model enabled isolating the management of main web pages from a legacy folder of approximately 6 million lines of code embedded within the project’s over 40 million lines in Chromium.
The operation, which would have required months of manual work, was completed in a few weeks by modifying hundreds of files in a coordinated manner: for end users, this translates into the ability to move active web pages between tabs, side panels, and mobile windows without reloading them, eliminating an entire category of program anomalies.
In parallel, working autonomously overnight on the open-source Python ADK framework, Argon reduced long-session in-memory preparation times by 40% and sped up cold imports by 21%. On the computational efficiency front, the model also accelerated Gemma 4 E2B inference by 2.85 times on A100 accelerators via vLLM, bringing processing flow from approximately 37,900 to over 108,000 tokens per second in a few days without any loss of quality, while also demonstrating strong skills in creating three-dimensional visualizations and interactive animations executable directly in the browser.
Beyond purely code-related areas, Gemini 4 Argon has been refined to communicate clearly, directly, and free of unnecessary repetitions or clichés, strictly respecting stylistic guidelines in high-criticality workflows. Several international organizations and startups that participated in the early access program, including Deloitte, LexisNexis, Ambience Healthcare, and Isomorphic Labs, are already employing the new architecture for document synthesis, historical data series analysis, and inspection of large computer archives.
Another fundamental pillar of the offering concerns digital defense, a sector where Google has begun extending the Fairwind program to hundreds of newly accredited partners, offering them access to the model without usual operational blocks for defensive purposes. The specialized company Wiz, through its free public interest initiative named “Scan for Good,” used Argon to identify logical flaws and unknown vulnerabilities.
In collaboration with a U.S. federal agency, the system proactively fixed a flaw that risked exposing confidential government data, while in the healthcare sector it enabled the discovery and blocking of a chain of vulnerabilities that allowed bypassing public download controls and exposing a main administration key, thereby securing hundreds of thousands of private patient clinical records.
The contribution of Gemini 4 Argon extends decisively to academic and experimental research, to the point where Google has opened applications to offer the highest usage level of Google Antigravity free for 6 months to 10,000 scientists.
What Changes and What Are the Effects
In the field of quantum computing, the model translated an internal version of the AlphaQubit error decoder into a single high-efficiency hardware kernel, nearly doubling real-time processing speed, and wrote a formally verified implementation in Lean for the arcsin function that cuts space-time cost by over 100 times compared to reference standards. In nuclear fusion physics, researchers from Google and the University of Maryland used Argon to identify a new and broader class of mathematical counterexamples to Grad’s 1967 conjecture, providing valuable parameters for designing complex three-dimensional magnetic fields in stellarator reactors dedicated to clean energy.
Notable results also come from quantum chemistry, where the renowned open-source xTB package was made up to 2.3 times faster in predicting the three-dimensional structure of molecules and 1.8 times faster in conformer search, as well as from life sciences. Researchers at Gamow Labs indeed employed the AI to examine the genomes of newborns affected by rare lung diseases, uncovering a chromosomal rearrangement that had eluded traditional software for years and recording a score of 57% on the RareBench diagnostic test and 52.2% on LabBench.
In the field of industrial biochemistry, however, it took Argon only a few hours to select data and design five new enzymes of the serine esterase class, whose catalytic efficacy on the 4MU-acetate molecule was then confirmed by laboratory tests. This ability to understand the physical world finally finds application in advanced robotics: after achieving a record score of 87% in the ERQA three-dimensional spatial reasoning benchmark, Argon demonstrated that it can learn physical tasks simply by watching a video.
Analyzing raw video footage of a person engaged in the game of cornhole, the model mapped environmental coordinates, calculated necessary joint movements, and generated simulation code in a few hours which, once transferred to Apptronik’s Apollo humanoid robot, allowed the machine to make seven consecutive successful throws in the real world.
This entire ecosystem rests on a solid infrastructure of security and automatic real-time monitoring, capable of halving violations in adversarial tests, containing risks under 1% in biological tests, reducing agent attempts to take unauthorized shortcuts during programming to 2%, and lowering the success rate of indirect prompt injection attacks to 0.7% across 15 adaptive attempts certified by the Gray Swan benchmark.
Source and further reading on Gemini 4 Argon launch: original article.
* Content created with the assistance of artificial intelligence systems.