Google has introduced Gemini 4 Argon, a new frontier AI model designed for long, complex tasks in software development, professional research, and cybersecurity. The company is initially making the model available to a small group of trusted cyber defenders through its Fairwind Program, rather than releasing it broadly at launch.
Google says Argon can handle up to 1 million output tokens in one response, a sharp increase from the 64,000-token limit of its previous models. In practical terms, that capacity could allow the system to work through large codebases, lengthy reports, document collections, or multi-step research tasks without losing as much context along the way.
The company is taking a gradual approach because of the model’s cybersecurity capabilities. Argon can reportedly find, verify, and help patch serious software vulnerabilities. Google says trusted defenders will receive access without the model’s usual cybersecurity restrictions, while the company continues to test safeguards before a wider launch.
CNBC reports that Google is also participating in a voluntary U.S. government process for pre-release model access. Tulsee Doshi, Google’s product lead for Gemini models, told the publication that the limited rollout is intended to put advanced defensive tools in the hands of cybersecurity teams while building confidence in the model’s safety measures.
Claims of gains in coding and professional work
Google positions Argon as a general-purpose model for demanding work, not only cybersecurity. It claims the model reached 77.9% on DeepSWE v1.1, a benchmark for long-horizon software engineering tasks. The company also says Argon leads the Vals Index, which evaluates work in fields including finance, coding, legal services, and tax.
For business automation, Google reports a score of 51.3% on Zapier’s AutomationBench. It also cites a 91.7% result on LVBench, a benchmark for understanding long videos. These results are company-reported and reflect benchmark performance, which may not directly predict results in every workplace.
Internally, Google says teams already use Argon for coding, research, and writing. The company reports that agents using the model identified memory improvements across its data centers that freed more than 300 tebibytes of memory after deployment. It also says the model helped quantum researchers improve a published baseline for one optimization problem by 40%.
Google describes another use case involving libgav1, its open-source video decoder. According to the company, Argon agents replaced 32,000 lines of SIMD code in a Rust port. Google says the resulting decoder ran 2.7 times faster than the prior Rust version while retaining identical video output.
Safety work before wider access
The company says it is strengthening protections against misuse involving cyberattacks and chemical, biological, radiological, and nuclear risks. It also highlights defenses against indirect prompt injection, where malicious instructions hidden in documents or other external content attempt to redirect an AI system.
Google further says it monitors model reasoning and actions for signs that an agent may go beyond a user’s stated intent. The company acknowledges that it is still refining these protections before expanding access to developers, enterprises, and consumers.
Argon is due to reach paid API customers and Google AI Ultra subscribers after the initial testing phase. Google lists an introductory price of $2 per million input tokens and $10 per million output tokens. After the introductory period, those prices are set to rise to $4 and $20 respectively. Cached input tokens will receive a 95% discount from the standard input rate.
Sources
Stay up to date
AI for content creation: the latest tools, tips and trends. Every two weeks in your inbox:
