Google unveiled Gemini 4 Argon for complex professional and technical tasks

Gemini 4 Argon, a new artificial intelligence model designed for complex programming, research and knowledge processing tasks, was presented by Google.

According to the company, the new model can be leveraged in real working conditions, in areas such as software development, cyber defense, finance, law, and taxation.

One of the key capabilities of the Gemini 4 Argon is the production of up to 1 million tokens. Tokens are, simply put, the small sections in which an artificial intelligence system divides the text in order to read it, process it and produce answers. The increased threshold allows the model to complete very large tasks and produce correspondingly large volumes of text. Google says this is the highest production threshold on the market.

In the field of programming, Gemini 4 Argon recorded 77.9% performance in DeepSWE v1.1, an evaluation tool that examines the performance of models in real-life and complex software engineering tasks of long duration. According to Google, this performance is a new top result.

In cyber defense, the model scored 68% on the CWE-bench and tied in first place. This tool assesses whether an AI model can detect and restore security vulnerabilities in software.

Argon also ranked first in the Vals Index, which measures the economic impact of AI models on knowledge-driven tasks. The evaluation includes areas such as finance, planning, legal and taxation.

Already, according to the company, thousands of Google employees are using the new model in specialized programming and research tasks. Its applications include transferring code libraries from C and C++ languages to Rust, as well as identifying ways to better utilize memory in data centers.

Google estimates that these optimization interventions can free up more than 300 TiB of memory in its infrastructure.

The disposal of Gemini 4 Argon will take place gradually. Initially, the model is tested by selected partners through the Fairwind program. It will then be made available more widely, initially to paid API customers and Google AI Ultra subscribers.

The company points out that phasing in is part of the process of testing the safety of more advanced AI models as their capabilities expand.