Google has unveiled Gemini 4 Argon, the first model in its next generation of Gemini, and it is betting the new system can put the company back at the front of the AI race. Announced on September 30 in a blog post by Koray Kavukcuoglu, Google’s chief AI architect and senior vice president of Google DeepMind, Argon is built for long and complex work in software engineering, finance, legal research and cybersecurity defence. There is a catch for almost everyone reading this: for now, only a small group of trusted cyber defenders can use it.
Access runs through Google’s Fairwind Program, a limited scheme for governments and trusted partners that the company launched in early September. Google says releasing frontier capabilities at this level “requires a phased approach,” and that it is taking part in the US government’s voluntary process for pre-release model access. Paid API customers and Google AI Ultra subscribers will be next in line, followed by developers, businesses and consumers, but Google has not given a date. When it does open up, Argon will cost an introductory $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 once the introductory period ends.
Google is making bold claims about performance. Argon scores 77.9% on DeepSWE v1.1, a test of lengthy software engineering tasks, which the company calls a new state of the art. Axios reported that this compares with 74.2% for Anthropic’s Claude Opus 5.5 and 74.1% for OpenAI’s GPT-6 Astra. Google also says Argon leads the Vals Index, which weights finance, coding, legal and tax work by their contribution to the US economy, ranks first on Zapier’s AutomationBench with 51.3%, and scores 91.7% on LVBench, a test of understanding long videos. On CWE-bench v1, which measures how well a model fixes security vulnerabilities, it ties for first place at 68%.
The headline technical change is how much the model can produce in one go. Google has raised Argon’s output limit to 1 million tokens, up from 64,000. “When the model has the headroom to think deeply and generate hundreds of thousands of tokens in a single trajectory, it adds a new level of depth in reasoning to solve tough problems in one go,” the company wrote.
Google says thousands of its own staff already use Argon every day, and it shared some striking examples. A team of Argon agents analysed profiling data from across Google’s data centres and applied memory optimisations that freed more than 300 tebibytes once rolled out, with total savings estimated at between 500 tebibytes and 1 pebibyte. Argon agents are also rewriting C and C++ code in the memory safe Rust language across Google, including more than 800,000 lines for the Zircon kernel at the heart of its Fuchsia operating system, with those rewrites going through automated and manual audits before reaching production. In one project, the model produced a video decoder that runs 2.7 times faster than the previous Rust version.
Cybersecurity is where Google is pushing hardest. It trained Argon to find, validate and patch serious software flaws on its own, and it will release the model to trusted defenders and its internal teams without cyber guardrails, so they get its full capability. Security firm Wiz is already using Argon through its Scan for Good initiative, which protects public infrastructure for free. Google says the model uncovered a critical vulnerability exposing sensitive personal information in healthcare software used by hospitals worldwide, a flaw earlier frontier models had missed.
For everyone else, Google is adding safeguards before a wider launch. It says Argon refuses requests that could help with cyber or chemical, biological, radiological and nuclear attacks, leads Gray Swan’s benchmark for resisting prompt injection, and runs under monitoring that watches its reasoning and actions and halts execution if it steps beyond what a user intended. Google is also sealing its sandboxed test environments before high risk training runs. Those measures arrive in a week dominated by agents slipping their limits, from OpenAI’s training pause to Nvidia’s new agent safety platform.
Argon also carries the weight of a difficult year for Google. It is the company’s first new generation of flagship models since Gemini 3 last November. A promised Gemini 3.5 Pro was expected in June but never arrived, and Axios reported in July that poor morale inside Google DeepMind contributed to the delay. “Argon is a well-rounded model that has frontier capabilities across several domains,” Tulsee Doshi, head of Gemini products at Google DeepMind, told Axios.
The real test comes when people outside Google can use it. Every benchmark figure so far comes from Google itself, and Claude Opus 5.5 and OpenAI’s GPT-6 models are already in customers’ hands while Argon sits behind an invitation list. If Google can move quickly from cyber defenders to paid developers without the delays that dogged Gemini 3.5, Argon could reset the leaderboard. If the wait drags on, its strongest numbers will belong to a model most people still cannot try.
Get Techord in your inbox
The week’s biggest AI, science and robotics stories, explained in plain English.
