10:10Machine-translated from Azerbaijani by AIRead in Azerbaijani
Google Introduces the Gemini 4 Argon AI Model
Google has introduced the new Gemini 4 Argon AI model for complex, long-running tasks. The model can handle work in programming, research, law, finance, writing, and cybersecurity. According to the company, it is the most powerful model it has developed to date. After the Gemini 3.5 Pro project was discontinued, the focus shifted to 3.8 Flash and the new Argon model.
The response limit for Gemini 4 Argon has increased from 64,000 tokens to 1 million tokens. A token is a small piece of text processed by the model. The higher limit means the model can generate longer responses. In DeepSWE v1.1, a test measuring software engineering skills, the model scored 77.9%. Claude Opus 5.5 scored 74.2%, while GPT-6 Astra scored 74.1%. During the initial phase, Gemini 4 Argon was made available only to trusted cyberdefense partners under the Fairwind Program. Google says the model can independently discover, validate, and fix critical software vulnerabilities. It tied for first place with a score of 68% in CWE-bench v1, a test measuring the remediation of software vulnerabilities. The process of giving the US government access to the model before its official launch is also underway. Google is working to prevent the model from being used for malicious purposes before making it widely available. This includes protecting it from fake instructions intended to deceive the model, strengthening system security, and monitoring inappropriate behavior. The company will monitor the processes occurring within the model and the steps it takes and will stop its operation if a risk arises. The security of isolated environments will also be strengthened separately before high-risk training and testing. The model also achieved high scores in tests involving financial research and the drafting of legal texts. It ranked first with a score of 51.3% in AutomationBench, a test measuring the completion of tasks from start to finish. It scored 91.7% in LVBench, a test evaluating the ability to understand long videos. The model can also analyze information in images, diagrams, and long videos.
Thousands of Google employees use the Argon model for code fixes, research, and migrating large codebases. AI agents have freed up more than 300 TiB of storage space in data centers. The total savings are expected to range from 500 TiB to 1 PiB. The model is also helping migrate the C and C++ code of the Fuchsia OS Zircon kernel, which consists of more than 800,000 lines, to Rust. In the libgav1 project, a version that updated 32,000 lines of code ran 2.7 times faster than the previous Rust version. Gemini 4 Argon will soon become available, initially to Google AI Ultra subscribers and paying API customers. The initial price will be $2 (approximately 3 AZN) per 1 million input tokens. The price for output tokens has been set at $10 (approximately 17 AZN). Later, the price of input tokens will rise to $4 (approximately 7 AZN), while the price of output tokens will increase to $20 (approximately 34 AZN). Input tokens stored in the cache, or temporary memory, will initially be sold at a 95% discount.
The response limit for Gemini 4 Argon has increased from 64,000 tokens to 1 million tokens. A token is a small piece of text processed by the model. The higher limit means the model can generate longer responses. In DeepSWE v1.1, a test measuring software engineering skills, the model scored 77.9%. Claude Opus 5.5 scored 74.2%, while GPT-6 Astra scored 74.1%. During the initial phase, Gemini 4 Argon was made available only to trusted cyberdefense partners under the Fairwind Program. Google says the model can independently discover, validate, and fix critical software vulnerabilities. It tied for first place with a score of 68% in CWE-bench v1, a test measuring the remediation of software vulnerabilities. The process of giving the US government access to the model before its official launch is also underway. Google is working to prevent the model from being used for malicious purposes before making it widely available. This includes protecting it from fake instructions intended to deceive the model, strengthening system security, and monitoring inappropriate behavior. The company will monitor the processes occurring within the model and the steps it takes and will stop its operation if a risk arises. The security of isolated environments will also be strengthened separately before high-risk training and testing. The model also achieved high scores in tests involving financial research and the drafting of legal texts. It ranked first with a score of 51.3% in AutomationBench, a test measuring the completion of tasks from start to finish. It scored 91.7% in LVBench, a test evaluating the ability to understand long videos. The model can also analyze information in images, diagrams, and long videos. Thousands of Google employees use the Argon model for code fixes, research, and migrating large codebases. AI agents have freed up more than 300 TiB of storage space in data centers. The total savings are expected to range from 500 TiB to 1 PiB. The model is also helping migrate the C and C++ code of the Fuchsia OS Zircon kernel, which consists of more than 800,000 lines, to Rust. In the libgav1 project, a version that updated 32,000 lines of code ran 2.7 times faster than the previous Rust version. Gemini 4 Argon will soon become available, initially to Google AI Ultra subscribers and paying API customers. The initial price will be $2 (approximately 3 AZN) per 1 million input tokens. The price for output tokens has been set at $10 (approximately 17 AZN). Later, the price of input tokens will rise to $4 (approximately 7 AZN), while the price of output tokens will increase to $20 (approximately 34 AZN). Input tokens stored in the cache, or temporary memory, will initially be sold at a 95% discount.