Sundar Pichai on Google launching Gemini 4 Argon, its most powerful AI model yet

Google has introduced Gemini 4 Argon, its most advanced AI model to date, in a bid to compete in the rapidly evolving AI landscape. The powerful model is geared towards long-running tasks in software engineering, enterprise knowledge work, and cybersecurity. Initially, Argon will be accessible to Google's "trusted cyber defenders" via its Fairwind program, with broader availability slated for developers, enterprises, and users later.

Premium API subscribers and Google AI Ultra customers will be the first to gain access to Gemini 4 Argon, though the company has not yet set a release date for a wider public rollout.

Google CEO Sundar Pichai highlighted the model's strong performance in complex workflows, cyber defense, and software engineering, noting that teams within Google are extensively using it for tasks ranging from coding to quantum computing with positive feedback. The model boasts an impressive 1 million-token output limit, compared to the previous 64K-token limit, enabling it to handle more extensive reasoning tasks in a single run.

Google engineers have been utilizing Argon for routine tasks, including debugging and large-scale codebase migrations, setting a new state-of-the-art record in DeepSWE v1.1 performance metrics (77.9%), which assesses a model's proficiency in real-world software engineering tasks.

Argon's capabilities extend beyond coding, as it is designed to excel in knowledge-intensive fields such as finance, legal research, and drafting. According to Google Deepmind tests, the model demonstrates strong performance across diverse real-world tasks, including professional knowledge work, coding, scientific problem-solving, and managing vast amounts of information. Argon's ability to work with large volumes of data and discern relevant details is particularly noteworthy.

In terms of cybersecurity, Argon takes center stage by autonomously identifying, validating, and patching software vulnerabilities. Currently, the model is offered without its usual cyber guardrails to early users and Google's internal teams. Cybersecurity firm Wiz has already incorporated Argon into its "Scan for Good" initiative.

Early demonstrations have shown Argon successfully identifying a critical vulnerability in healthcare software used by hospitals worldwide, exposing sensitive personal information. Google is adopting a phased approach for the rollout due to the model's cybersecurity features, working on safeguards to prevent misuse, including cyberattacks and CBRN-related threats, as well as defenses against indirect prompt injection attacks.

The intensifying AI race among tech giants, including Google, Anthropic, and OpenAI, has led to the release of several advanced models. In September, OpenAI unveiled GPT-6 Astra, its most capable model yet, featuring a 1.05 million-token context window and a maximum output of 128,000 tokens. Anthropic recently launched Claude Opus 5.5, which they claim is comparable to its more powerful Claude Fable 5.1 on most tasks while being 40% less expensive to operate.

Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, with a 60% reduction in cache-read costs. Google is positioning Gemini 4 Argon as the ideal choice for agentic coding, computer use, and knowledge work. Additionally, the company has been expanding its AI efforts in India, selecting four Indian startups for the inaugural Google DeepMind Accelerator: AI for the Planet, providing them with access to Google's AI stack, technical support, and mentorship to develop AI-driven solutions for environmental issues.