The most substantial technical change in this release is the expansion of the output token limit from 64,000 to 1,000,000 tokens. This increase allows the model to handle multi-step tasks and large-scale codebase migrations without interruption. Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect, stated that this capability fundamentally changes how the company builds and works internally. The model is already powering internal workflows, with thousands of employees utilizing it for specialized coding, deep research, and writing tasks.
Performance metrics indicate that Gemini 4 Argon sets new state-of-the-art results in several key benchmarks. On DeepSWE v1.1, which measures long-horizon software engineering tasks, the model achieved a score of 77.9%. In enterprise applications, it leads the Vals Index, a metric that weighs economic impact across finance, coding, legal, and tax work by their contribution to U.S. GDP. Additionally, on AutomationBench, which tests end-to-end execution in business functions, Argon ranked first with a score of 51.3%. The model also demonstrates strong multimodal capabilities, scoring 91.7% on LVBench for long video understanding, allowing it to identify details and take actions based on visual data.
Cybersecurity is a primary focus of the Argon release. Google trained the model to autonomously find, validate, and patch critical software vulnerabilities. To support this, the company is initially rolling out the model to a select group of trusted cyber defenders through its Fairwind Program. For these specific users and internal Google teams, the model will be released without standard cyber guardrails to allow for full frontier-level defense capabilities. This phased approach is part of a broader strategy to safely release frontier capabilities, with Google actively engaging with the U.S. government’s voluntary pre-release access process.
While the initial rollout is limited, Google has confirmed pricing for when the model becomes generally available to developers, enterprises, and consumers. The introductory price is set at $2 per million input tokens and $10 per million output tokens. Cached input tokens will be priced at a 95% discount off the standard input rate. The company plans to gather feedback from early testers to iterate on guardrails before expanding access. This release follows months of internal testing and development, where the model was internally codenamed “argon” before its public announcement.
The development of Gemini 4 Argon reflects a strategic push by Google to compete in the frontier AI market, focusing on autonomous agent workflows and complex code generation. By bypassing traditional release cycles for a more targeted deployment to security professionals, Google aims to establish the model as a robust tool for defense before broader commercial release. The model’s ability to handle diverse workloads, from financial research to legal drafting, positions it as a versatile enterprise solution rather than a single-purpose tool.
Google will continue to monitor performance and safety protocols as it expands access beyond the initial Fairwind cohort. The company has not provided a specific date for general availability but indicated that broader access is being prepared as soon as possible. As the AI landscape continues to evolve, the introduction of Gemini 4 Argon signals an intensification of competition among major AI developers, each seeking to define the standards for frontier model capabilities in high-stakes professional environments.
