Alphabet-owned Google introduced Gemini 4 Argon on Wednesday, describing it as its most capable AI model with initial access restricted to selected cybersecurity partners. The company reported gains in software development, security and demanding professional tasks, with a broader release dependent on further safeguards.
Access is being provided through Fairwind, Google’s security program. According to the company, Argon’s training focused on cyber defense, enabling it to identify serious software flaws, verify them and produce fixes without human direction.
Inside Google, the model is already supporting both infrastructure operations and employees’ routine engineering work. The company said its use in data center memory optimization made hundreds of terabytes available without requiring purchases of more hardware.
Employees have also applied Argon to debugging and moving codebases, while quantum computing researchers have used it in their work. Google said the model can interpret visual material, including charts and extended video footage.
To support its performance claims, Google pointed to tests covering several distinct fields. It reported a record result for practical software engineering, a shared top ranking in cybersecurity and the highest score on a separate assessment spanning finance, law and other professional activities.
A Google blog post also presented benchmark results that it said placed Argon substantially ahead of OpenAI’s GPT-6 Astra and Anthropic’s Fable and Opus. The company cited the AI model index from Vals, an increasingly popular benchmarking startup, as another measure on which Argon held the top position.
Those comparisons come amid competing releases from leading AI developers, which have continued introducing more powerful systems while also warning about the possibility of losing control over the technology. OpenAI recently presented Astra as its strongest offering, while Anthropic promoted Fable in similar terms when it launched earlier this year.
For Google, the introduction follows recent progress with Gemini after earlier perceptions that the company was trailing competitors. In August, Google reported more than one billion monthly users for its app. OpenAI has also recently put ChatGPT’s monthly audience at one billion, leaving the two services competitive on that measure.
Safety preparations remain part of Google’s rollout plans. On Tuesday, chief executive Sundar Pichai joined President Donald Trump and other major technology executives in signing a voluntary agreement at the White House after a luncheon focused on mounting AI safety concerns.
Google is coordinating with the U.S. government on safety assessments before release and intends to expand access in stages, beginning with trusted security partners. Before making Argon publicly available, it said Wednesday, it wants to strengthen protections across four cybersecurity priorities, such as misuse and prompt injection.
