Google has unveiled Gemini 4 Argon, its most powerful AI model, but is releasing it first only to a small group of trusted cybersecurity defenders.
The company said the model performs at the frontier in software engineering, legal and financial work, and cyber defence.
Google said releasing capabilities at this level safely required a phased approach.
It is taking part in the US government's voluntary scheme for reviewing models before release, and will gather feedback from early testers before opening Argon to developers, businesses and consumers.
Paid API customers and subscribers to Google AI Ultra, its top consumer plan, will get access first.
Same price as OpenAI's Sol
Argon will launch at an introductory price of $2 per million input tokens and $10 per million output tokens.
Cached input, text the model has already seen, will be 95% cheaper, at $0.10 per million tokens.
That matches the pricing of GPT-6.1 Sol, the near-flagship model OpenAI released this week.
No cyber guardrails for defenders
Google trained Argon to find, confirm and fix serious software vulnerabilities on its own.
Trusted defenders and Google's internal teams will get a version without cyber guardrails, so they can use its full capabilities.
Wiz, the cloud security company, is already using Argon in a free programme that protects public infrastructure.
Google said the model found a critical flaw exposing sensitive personal data in healthcare software used by hospitals worldwide, which earlier frontier models had missed.
Benchmarks
Google said Argon set a new high score of 77.9% on DeepSWE v1.1, which tests long-running software engineering tasks.
It ranked first on AutomationBench, Zapier's test of business tasks, with 51.3%, and tied for first on CWE-bench v1, which measures how well models fix security flaws, with 68%.
The company also claimed the top spot on the Vals Index, which weights performance in finance, coding, legal and tax work by each sector's share of US economic output.
All the figures come from Google.
Google is raising the model's output limit to 1 million tokens from 64,000, allowing it to work through much longer problems in a single run.
Already at work inside Google
Thousands of Google staff are already using Argon, the company said.
Teams of Argon agents found memory savings across Google's data centres that freed up more than 300 tebibytes once rolled out.
The agents are also rewriting code from C and C++ into Rust, a programming language designed to prevent common security bugs, including more than 800,000 lines for Google's Fuchsia operating system kernel.
Safety measures
Google said Argon is designed to refuse requests that could help cyberattacks or chemical, biological, radiological or nuclear attacks.
It is also monitoring the model's reasoning and actions so it can stop it if it goes beyond what a user intended.
Google urged other AI companies to keep their models' reasoning visible, so warning signs of misbehaviour can still be spotted.
