The AI model is aimed at complex, multi-step work across software, security, finance, law and science, and arrives with introductory pricing that will double from December 1.
Google has released Gemini 4 Argon, a new artificial intelligence model built for demanding, multi-step work in areas ranging from software development and cybersecurity to finance, legal analysis, scientific research and robotics. The company says the model improves on earlier versions in reasoning across several steps, interpreting visual material and carrying out tasks with little human direction.
Developers and enterprises worldwide can use the model through the Gemini API. Subscribers to Google AI Pro and Ultra can also reach it in the Gemini app and in AI Mode in Search, depending on eligibility. Google Antigravity users who qualify have access as well, while developers can work with it in Google AI Studio and other supported tools. Business customers can use it through eligible enterprise subscriptions and the API.
An introductory rate of $2 per million input tokens and $10 per million output tokens applies for now, with cached input tokens discounted by 95 per cent. From December 1, 2026, those prices rise to $4 and $20 respectively. The API supports a context window of one million tokens, which lets developers take on larger tasks and longer workflows.
Coding, robotics and research claims
Google positions Gemini 4 Argon as suited to jobs that demand extended reasoning. In software work, it can restructure large codebases, spot security weaknesses and tune application performance. Among Google’s own engineering teams using the model alongside Antigravity, the company reported a 35 per cent rise in agentic code submissions and a drop of nearly one-third in code rollbacks.
The company pointed to further examples, including gains in the speed of its open-source Agent Development Kit, improvements to scientific computing software and support for researchers studying complex engineering problems.
In robotics, Google said the model scored 87 per cent on the Embodied Reasoning benchmark, which tests how well a system understands and solves spatial problems in three-dimensional settings. In an internal demonstration, the model watched a video of a person playing cornhole, built a simulation from it and then passed the learned task to an Apollo humanoid robot.
Google also cited uses in quantum computing, nuclear fusion and enzyme design. Researchers, it said, used the model to propose enzyme candidates that later showed catalytic activity in laboratory tests.
Safety and security measures
Google says the model carries protections meant to limit harmful outputs, resist prompt injection attacks and stop AI agents from taking actions they are not authorised to perform. It has also added monitoring systems designed to flag potentially unsafe behaviour while the model carries out tasks on its own.
The performance figures and research examples come from Google’s own reporting and have not been independently verified.
