Google has launched a new frontier AI model called Gemini 4 Argon, and early access is going first to a small group of trusted cybersecurity defenders rather than the public. The model is built to handle long, complex tasks in one go, spanning software engineering, finance, legal work and cyber defense, and Google says it is already reshaping how its own engineers build products.
Argon is rolling out through a program called Fairwind, even as Google works alongside the US government's voluntary review process for new frontier models. Wider access for developers, businesses and everyday users will follow only once more safety testing is done, the company said.
Who Can Use Gemini 4 Argon Right Now
Right now, Argon is limited to a set of trusted cyber defenders picked for the Fairwind program. Google is also running the model inside its own teams, where thousands of employees are already using it for daily engineering work. A broader rollout to outside developers and regular users is planned, but Google has not given a fixed date, saying only that it wants to gather feedback and tighten safety guardrails first.
How Much Will Gemini 4 Argon Cost To Run
Google has set an introductory price of 2 dollars per million input tokens and 10 dollars per million output tokens. Cached input tokens, meaning text the model has already processed before, cost 95 percent less than fresh input. These prices apply for now and could change once Argon moves beyond its limited release.
What Makes Gemini 4 Argon Different From Earlier Models
The biggest technical jump is in output length. Argon can now generate up to 1 million tokens in a single response, up from a limit of 64000 tokens on earlier versions, a roughly sixteen fold increase. Google says this extra room lets the model reason through tough problems step by step instead of cutting its thinking short, which matters for tasks like debugging large codebases or working through a long legal document in one pass.
How Is Google Already Using Gemini 4 Argon Internally
Inside Google, Argon is being used on projects most companies would consider too complex to hand to an AI model. In quantum computing research, it beat a published performance benchmark by 40 percent in just minutes. Teams of Argon agents scanned performance data across Google's data centers and found memory optimizations on their own, freeing more than 300 TiB of memory so far, with total savings expected to reach between 500 TiB and 1 petabyte. Argon is also being used to rewrite old C and C++ code into Rust, a safer programming language, across projects as large as the 800000 line kernel that powers Google's Fuchsia operating system. For a video decoding tool called libgav1, Argon rewrote 32000 lines of complex code, and the result runs 2.7 times faster than the earlier version while producing identical video output.
How Does Gemini 4 Argon Score On Independent Benchmarks
Argon sets new high scores on several industry tests. It leads DeepSWE, a benchmark for real world software engineering tasks, and tops the Vals Index, which measures economic impact across finance, law, tax and coding work. It also ranks first on AutomationBench, a test built by Zapier to measure how well a model can complete business tasks end to end, and posts a state of the art score on LVBench, which checks how well a model understands long videos.
Can Gemini 4 Argon Find And Fix Security Flaws On Its Own
Google says Argon can search for software vulnerabilities, confirm they are real, and patch them without a human guiding every step. Trusted defenders and Google's own security teams are getting a version of the model with its cyber restrictions removed, so they can use its full capability. Security firm Wiz is already using Argon through its Scan for Good program, which fixes security risks in public infrastructure for free, and the model recently found a serious flaw in hospital software that was exposing patient information, a risk earlier AI models had missed entirely.
What Is Google Doing To Keep Gemini 4 Argon Safe Before A Wider Launch
Before Argon reaches more users, Google says it is strengthening four areas of safety. It is working to stop the model from being misused for cyber attacks or help with chemical, biological or nuclear threats. It is also hardening the model against prompt injection, where hidden instructions try to hijack an AI system's behaviour, and says Argon performs best in class on an outside benchmark for this. Google is additionally monitoring the model's internal reasoning to catch signs of it acting beyond what a user actually asked for, and is tightening the sandboxed environments used to test the model before training and evaluation begin.
