Google has announced Gemini 4 Argon, its latest frontier model which the firm says delivers industry-leading performance across enterprise use cases and cybersecurity capabilities.
Argon is Google’s first pro model in over seven months and aims to compete with the likes of OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5. Its AI research lab Google DeepMind said Argon is capable of autonomous vulnerability discovery, penetration testing and patching.
On CWE-bench v1, a benchmark that tests AI models on their ability to patch real-world vulnerabilities in software, Argon is tied with SpaceXAI’s Grok 4.7 and Astra at 68 per cent.
Google said it has already given the US government access to the model and from 30 September this would expand to “trusted cyber defenders” enrolled in the Fairwind Program.
This is Google’s early access track for vetted companies such as Accenture, McKinsey & Company and Palo Alto Networks and a direct competitor with Anthropic’s Project Glasswing and OpenAI’s Daybreak.
Argon is also intended to be a leap for Google’s AI multimodal enterprise capabilities across tasks such as legal and finance work and visual understanding. It is the new leader at LVBench, a benchmark that tests AI long video understanding.
The independent AI benchmarking platform Artificial Analysis ranks Gemini 4 Argon as the new leader at its AutomationBench-AA benchmark, which tests how well an AI model can complete business tasks using simulated applications such as Gmail, Google Sheets, HubSpot, Slack and Zendesk.
Argon scored 77.5 per cent, comfortably ahead of Claude Sonnet 5.5’s 71.3 per cent and DeepSeek v4.1 Flash’s 98.9 per cent.
The firm said its own engineers have been using Argon for daily work and more advanced tasks. It gave examples including a team using Argon agents to optimise Google’s centre memory use, freeing more than 300 TiB of memory.
Engineers have also used Argon agents to migrate Google’s C and C++ codebases to the memory safe programming language Rust, with internal teams using the AI model to replace hundreds of thousands of lines of code.
Argon has a one million token output limit, meaning it can produce up to 30,000 lines of code or text equivalent to four or five novels in one go. Google noted the real purpose of the expanded limit, however, is to enable Argon to apply additional layers of reasoning to complex tasks, so that it can solve complex problems in a single prompt.
Google has not provided a rollout timeline for the model but said it is working with early access testers to “strengthen our systems” before making it available via the Gemini API and to Google AI Ultra subscribers.
It will make Gemini 4 Argon available at $4 per million input tokens and $20 per million output tokens, with an introductory discount that reduces these prices to $2 and $10. The firm did not specify the dates for this promotion.
