Google has launched the Gemini 4 Argon model, which it presents as its latest and most capable flagship model to date, with a focus on executing complex tasks that span multiple steps, rather than simply producing short answers or executing single commands. The model is specifically targeted at software engineering, specialized knowledge work, and cyber defense.
One of the most notable technical improvements is raising the output limit to one million tokens, compared to 64,000 tokens in the previous model, according to Google. The company says that increasing the available space for the model allows it to continue with long inferences, executing hundreds of thousands of tokens in a single path, which targets problems that require sequential analysis rather than a short series of instructions.
This means, in practical terms, that the model can handle large software projects or multi-stage research tasks within a single context, while retaining more of the information needed to continue the task. Google emphasizes that this approach is a fundamental part of Argon's design, not simply an increase in the size of the context window.
From writing code to rebuilding projects
Google places programming at the forefront of the model's intended uses, saying its engineers use it for debugging, designing algorithms, and migrating massive codebases between programming languages.
The company notes that Argon agents are working on migrating code bases written in C and C++ to Rust, with one migration reaching over 800,000 lines in the Fuchsia Zircon kernel. It also says the model achieved a score of 77.9% in the DeepSWE v1.1 test, which is designed to measure model performance in real-world, long-term software engineering tasks.
In another example, Google says that Argonne agents rewrote 32,000 lines of SIMD code in the libgav1 project, resulting in a REST version that was about 2.7 times faster than the previous REST version, while maintaining the same output. The company emphasizes that large-scale rewrites are subject to automated and manual review and testing before being adopted in production environments.
Financial and legal work
The model is not limited to programming; Google designed it to perform specialized knowledge tasks in areas such as finance, law, and taxation. The company says that Argon performed well in tests that included multi-step financial research, legal research and drafting, and business tasks that require the execution of a full series of actions.
Google also points to its multimedia capabilities, including graph analysis and understanding of long videos, where the model scored 91.7% in the LVBench test designed to measure understanding of long videos.
Tags:
technology