Google Unveils Flagship Gemini 4 Argon AI Model With an Output Limit of Up to 1 Million Tokens

Google has unveiled its new flagship artificial intelligence model, Gemini 4 Argon, designed specifically for complex tasks requiring in-depth analysis and prolonged work with large amounts of data. Unlike standard systems designed for quick, short responses, the new model is optimized for multistage programming, large-scale scientific research, detailed document analysis, and comprehensive workflow automation.

Argon’s key technological breakthrough is an unprecedented increase in its output token limit: the model can now generate responses of up to 1 million tokens, many times higher than the previous limit of 64,000. This allows the neural network to reason about a single task for much longer, sequentially analyze large projects, identify hidden errors in program code, and immediately propose ready-made fixes without having to restart the context.

The new model’s high effectiveness is confirmed by independent benchmarks. In DeepSWE v1.1, a specialized test that evaluates an artificial intelligence system’s ability to autonomously solve complex web development tasks, Gemini 4 Argon achieved a score of 77.9%. This result allowed the new model to surpass leading market competitors—GPT-6 Astra with 74.1% and Claude Opus 5.5 with 74.2%. Argon also delivered outstanding results in analyzing long videos and conducting complex legal and financial assessments.

The developers paid particular attention to cybersecurity: the model was specifically trained to find and eliminate critical vulnerabilities, enabling it to score 68% on the specialized CWE-bench v1 test and match the performance of the world’s best solutions in this field. Within Google itself, Argon-based agents are already being actively used to automatically optimize server infrastructure and migrate legacy software code from C/C++ to the safer Rust on a large scale.

Despite the official announcement, the company is initially deploying the system under strict precautions. At the first stage, only Google’s internal teams and trusted experts will receive full access to specialized cyberattack protection features, while the standard version of Gemini 4 Argon will be rolled out gradually and become available to Google AI Ultra subscribers as well as paying users through the API.