Pakyong, October 2: Google Gemini 4 Argon, the technology giant’s newest flagship artificial intelligence model, was unveiled on Wednesday, September 30, with the company describing it as its most advanced system so far for coding, cybersecurity defence and complex professional work. Access has initially been restricted to vetted cybersecurity partners and Google’s own teams, and no date has been announced for general public availability.
The model is Google’s first flagship release since February and arrives nearly a year after Gemini 3. It also introduces a new naming scheme for the Gemini line.
Limited rollout through Fairwind Program
Google said the model is first being made available to trusted testers and cyber defenders through its Fairwind Program, the company’s security initiative. These partners can use the model to find and fix flaws in their software before a wider release.
According to the company, the model was trained specifically for defensive cyber work and can “autonomously find, validate, and patch critical software vulnerabilities.”
Tulsee Doshi, Google’s Gemini model product lead, told CNBC that the approach was deliberate. “Starting this rollout in this way gives us more confidence, but also enables us to put a model that is trained and strong in cyber defense in the hands of defenders as soon as possible,” she said.
Google said broader access will follow in phases, beginning with Google AI Ultra subscribers and paid API customers. The model is eventually expected to underpin most of Google’s services.
What the new model offers
Google has raised the model’s output limit to one million tokens from 64,000, which it said allows longer and more complex tasks to be completed in a single run. The company also highlighted improvements in coding, reasoning and the analysis of long videos and charts.
“Built to sustain deep reasoning across complex, long-horizon workflows, Argon is fundamentally changing the way we work and build at Google,” the company said in its announcement.
Google said thousands of its employees are already using the model. In one example cited by the company, a team of Argon agents identified memory optimisations across its data centres that freed more than 300 tebibytes of memory. The model is also being used to migrate large C and C++ codebases to the Rust programming language, with the rewrites undergoing automated and manual review before deployment.
Benchmark claims
Google has claimed leading results on several industry benchmarks. On DeepSWE v1.1, a software engineering test, it reported a score of 77.9 per cent, against 74.2 per cent for Anthropic’s Claude Opus 5.5 and 74.1 per cent for OpenAI’s GPT-6 Astra.
The company also reported a score of 91.7 per cent on LVBench, which measures long video understanding, and 51.3 per cent on AutomationBench, which tests the execution of business tasks. On CWE-bench v1, which evaluates the fixing of security vulnerabilities, it said the model tied for first place at 68 per cent.
These figures are the company’s own. Independent testing firm Artificial Analysis has reportedly rated the model level with GPT-6 Astra on its intelligence index, while Bloomberg reported ahead of the launch that some within Google were concerned it may not match rival offerings.
Pricing
Google said the model will launch at an introductory price of 2 US dollars per million input tokens and 10 US dollars per million output tokens. After the introductory period, the price will rise to 4 US dollars and 20 US dollars respectively.
Safety measures
Before the wider release, Google said it is focusing on preventing misuse in cyber, chemical, biological, radiological and nuclear attacks, strengthening resistance to prompt injection attacks, hardening its testing environments, and monitoring the model’s reasoning and actions so that execution can be stopped when necessary.
The company is also participating in the United States government’s voluntary pre-release review of advanced AI models. The launch came a day after Google chief executive Sundar Pichai reportedly signed a voluntary AI safety agreement at the White House along with other technology executives.
Intensifying AI race
The release comes amid strong competition among leading AI developers. OpenAI launched GPT-6 Astra in early September, while Anthropic released Claude Fable 5 in June. Google said in August that its Gemini app had crossed one billion monthly users.