Google has unveiled Gemini 4 Argon, a new AI model that kicks off the fourth generation of Gemini systems. The device stands out with an industry-leading output limit of one million tokens in a single trajectory and has been designed with complex workflows in mind.
A new generation of models from Google DeepMind
According to the official announcement published on September 30, 2026, by Google, a model designated as Gemini 4 Argon is hitting the market. The developers point out that this solution marks the beginning of the fourth generation of advanced AI systems. The model's architecture focuses on handling long-term and multi-stage tasks that have previously posed a significant challenge for digital infrastructure.
If you work with large software projects, financial data analysis, or security systems, you are likely aware of how quickly standard tools hit their memory and context limitations. Google has decided to address these issues by introducing radical changes to the model's input and output architecture.
One million output tokens in practice
The most important technological differentiator of the new version is undoubtedly the output generation limit. As Google states in its official post, Gemini 4 Argon allows for the processing and generation of up to one million tokens within a single trajectory. For comparison, previous versions had a limit of just 64,000 tokens.
Such a technological leap changes the way the system handles deep reasoning. The model can generate massive blocks of coherent source code, extensive reports, or full technical documentation in a single query. It is worth checking how these capabilities translate into everyday work environments, especially since these changes directly affect the IT industry.
Benchmark results and applications in cybersecurity
The model's performance in software engineering tasks has been confirmed in independent tests. As industry analyses indicate, Gemini 4 Argon achieved a score of 77.9% on the deepswe v1.1 benchmark, securing a leading position. High results were also recorded on the Vals Index, where the model scored 68.9%.
It is worth noting the area of cybersecurity. The model has undergone training in autonomous vulnerability detection and patching. Google partner, Wiz, used these mechanisms to detect a critical bug in medical software as part of the Scan for Good initiative. Due to the *dual-use* nature of the technology, access to raw defensive capabilities is initially limited to a select group of specialists under the special Fairwind program.
Internal deployments and pricing policy
Before the system reaches a wider audience, thousands of Google employees are using it for daily tasks. Examples include memory optimization in data centers or mass code migration in large projects. The company has also prepared a pricing plan for developers using the API:
- The promotional introductory rate is two dollars per million input tokens and ten dollars per million output tokens.
- Cached tokens are heavily discounted, reducing their processing cost to ten cents.
- After the promotional period ends, the rates will increase to four and twenty dollars, respectively.
Those interested in new technologies can follow the further development of the platform, keeping in mind that the market for advanced models is constantly accelerating.
Sources
- https://x.com/Google/status/2105388143902175529
- https://www.marktechpost.com/2026/09/30/google-deepmind-unveils-gemini-4-argon-with-1m-output-tokens-for-coding-knowledge-work-and-cyber-defense/
- https://pulse2.com/google-gemini-4-argon/
- https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/
Comments