Google has made Gemini 4 Argon available worldwide for developers, enterprises, and research teams through multiple access channels. The model ranks first on a benchmark that measures AI performance across critical economic sectors including finance, coding, legal, and tax work.
Engineering and software optimization
At Google, teams have been using Argon to tackle complex engineering problems. Software engineers report shipping code faster, with agentic code submissions up more than 35% and code rollbacks down nearly a third when using Argon in Antigravity. In one example, Argon helped the Chrome team refactor a six-million-line legacy folder within Chromium's forty-million-line codebase, completing months of work in weeks. For users, this enables live webpages to move between tabs and windows without reloading.
Argon also optimized Google's open-source Agent Development Kit (ADK), a Python framework for building enterprise AI agents, with changes now merged into the google/adk-python repository.
In open-source contributions, Argon accelerated xTB, a widely used quantum chemistry package for predicting molecular structures, making it up to 2.3 times faster at structure prediction and 1.8 times faster at conformer search. These improvements have been contributed to the official xTB repository on GitHub.
Argon also achieved a 2.85 times inference speedup on Google's Gemma models via vLLM.
The model shows strong capabilities in visual and spatial reasoning, enabling complex 3D visualizations, games, and interactive animations written in vanilla code that runs directly in browsers. Examples include autonomous generation of a detailed multi-building temple complex in Blender and creation of a playable miniature racing game built from photographs of a desk.
Enterprise users in early access report performance at the frontier across complex workflows. Teams in finance, legal, healthcare, and cybersecurity testing found Argon excels at multi-step reasoning and maintaining context across long document analyses, with some evaluations showing it completes research tasks 1.5 times faster than peer models.
The model launches at an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input tokens at 95% off.