CALIFORNIA / RankWire.AI / – Google has introduced Gemini 4 Argon, its latest flagship AI model designed to handle complex professional tasks. The company unveiled this model on Sept. 30, positioning it as the cornerstone of the Gemini 4 lineup. Argon is tailored for applications in software engineering, financial analysis, legal research, and cybersecurity defense. Google stated that this model is capable of sustaining more in-depth reasoning across long, multi-step processes. Currently, access is being granted to a select group of cybersecurity specialists via its Fairwind Program.
With Gemini 4 Argon, the maximum token output has been increased to 1 million, a significant jump from the previous cap of 64,000 tokens. This expanded limit enables the model to process and carry out extensive tasks within a single run. The introductory API pricing is set at $2 per million input tokens and $10 per million output tokens. Cached input tokens are offered at a 95% discount compared to the input rate. After the initial period, pricing is expected to increase to $4 and $20 per million tokens, respectively.
Google reports that thousands of its employees are already utilizing Argon for specialized coding, research, and writing projects. Internal teams have also applied the model to optimize data center memory and facilitate large-scale code migrations. One project involved using Argon agents to transition C and C++ codebases to Rust. Another employed the agents for memory profiling across Google’s data centers. These efforts have resulted in freeing more than 300 tebibytes of memory, with ongoing work identifying further savings.
Enhanced capabilities for advanced professional tasks
Google shared that Gemini 4 Argon achieved a score of 77.9% on DeepSWE v1.1, a benchmark for evaluating long-duration software engineering performance. The company also highlighted positive results across benchmarks related to finance, legal work, and automation. Argon supports multimodal reasoning, complementing its coding and enterprise functionalities. Developed by Google DeepMind as part of the broader Gemini family, the model’s increased output capacity enables it to effectively manage workflows that involve numerous consecutive reasoning and execution steps.
Cybersecurity is another vital focus in the model’s initial deployment. Google explained that Argon can identify, verify, and patch software vulnerabilities within controlled defensive environments. Through its Scan for Good initiative, Wiz is utilizing Argon to detect security vulnerabilities in public infrastructure. Google noted that Argon scored 68% on CWE-bench v1, a benchmark designed for vulnerability remediation. Access is being provided to selected cybersecurity defenders without the usual guardrails, specifically for approved defensive tasks.
Initial limited release sets the stage for broader Gemini 4 deployment
Google has not announced a definitive date for the widespread public release of Gemini 4 Argon. The company is currently employing a phased rollout, gathering feedback from early testers to refine the model. Additionally, it is participating in a voluntary process with the U.S. government to facilitate pre-release access to the model. Future plans include making the model available to developers, enterprises, and consumers. The initial rollout will target paid API customers and Google AI Ultra subscribers, although no specific launch date has been provided for these groups.
Google also clarified that it has no plans to release Gemini 3.5 Pro, which was initially scheduled for June. This decision positions Gemini 4 Argon as the company’s latest flagship model for demanding reasoning and professional workloads. Meanwhile, Google continues to offer other Gemini variants tailored to different performance levels and cost structures. Argon distinguishes itself with its larger output capacity, enhanced coding features, and specialized cybersecurity capabilities. Currently, access remains limited to trusted testers and select security partners involved in defense applications.