Gemini 4 Argon is Google’s AI built for the big jobs

Google has unveiled Gemini 4 Argon, its latest artificial intelligence (AI) model, boasting advanced complex reasoning and handling of lengthy tasks such as software engineering and enterprise work. This new model is tailored to tackle workflows that necessitate multiple steps, rather than merely addressing individual prompts. Google initially plans to introduce Argon to a select group of cybersecurity professionals via its Fairwind Program before expanding its availability to developers, enterprises, and consumers.

One of the most significant enhancements in Argon is its expanded output token limit, reaching an impressive 1 million tokens, a substantial increase from the previous 64,000 tokens. This allows Argon to process and generate extensive amounts of information in a single interaction, facilitating the resolution of intricate problems, the writing of substantial code, and the completion of lengthy research endeavors without the need to fragment tasks across multiple sessions.

This feature proves particularly beneficial for long-term projects requiring continuous planning, execution, verification, and ongoing work.

While Argon excels in coding and software engineering, Google envisions it as a versatile enterprise-level AI rather than solely a programming assistant. The company identifies Argon as leading in the Vals Index, which evaluates AI performance across finance, coding, legal, and tax domains. Furthermore, Argon showcases strong multimodal capabilities, with the ability to analyze professional charts, comprehend extended video content, and derive insights from multiple documents simultaneously.

Google highlights Argon's capability for autonomous vulnerability detection, validation, and patching within software. The company also announces its collaboration with Wiz, a cybersecurity firm, to leverage Argon through its Scan for Good initiative, aimed at identifying and addressing vulnerabilities in critical infrastructure.

Prior to rolling out Argon more broadly, Google is reinforcing safety measures around the model, particularly for its initial rollout to trusted cybersecurity defenders and internal teams. Google is enhancing protections against misuse, including risks related to cyber and biological threats. Additionally, the company is improving the model's resistance to indirect prompt injection attacks, where malicious instructions embedded in external content could manipulate an AI agent.

Google also deploys systems to monitor Argon's reasoning and actions for signs of misalignment, ensuring that the AI's behavior aligns with user expectations. Should misalignment be detected, the system can halt execution as needed. Google assures that similar monitoring was employed during the model's training, with any potential issues forwarded to a dedicated response team.

Importantly, Google has taken care not to feed back these findings into the training process to minimize the risk of Argon learning to circumvent these safeguards.

The phased release of Gemini 4 Argon begins with trusted cyber defenders and testers, with plans to subsequently make it available to developers, enterprises, and consumers. Initially, Argon will be accessible to paid API customers and Google AI Ultra subscribers. Pricing for Argon is set at an introductory rate of $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95% below the standard input rate.