# Google introduced the new AI model Gemini 4 Argon

Google has introduced a new artificial intelligence model, Gemini 4 Argon. It is designed for complex tasks in programming, corporate work, and cybersecurity.

Currently, the model is being tested by cybersecurity specialists as part of the Fairwind program. Google plans to gradually open access to it for developers, companies, and regular users after additional testing.

Gemini 4 Argon features a context window of 1 million tokens, compared to 64,000 in the previous version. This allows the model to work longer on complex tasks without needing to restart the process.

Google is already using Argon internally. The model assists engineers with programming, research, and system optimization. In one project, it helped improve a quantum algorithm by 40% compared to previously published results.

In another project, AI agents analyzed the operation of Google's data centers and identified ways to optimize memory usage. After implementing these changes, more than 300 TiB of memory was freed. The company estimates potential savings at 500 TiB to 1 PiB.

Argon also helps Google migrate large codebases from C and C++ to Rust. For example, the model refactored 32,000 lines of SIMD code for the libgav1 video decoder. The resulting Rust version runs 2.7 times faster than the previous Rust implementation while maintaining the same output.

The model also demonstrated high performance in tests for programming, finance, legal work, and other tasks. In AutomationBench from Zapier, which evaluates the ability of AI to execute business tasks from start to finish, Argon scored 51.3%.

Google separately highlights Gemini 4 Argon's capabilities in cybersecurity. The model can find, verify, and fix vulnerabilities in software. In the CWE-bench v1 test, it scored 68%, tying for first place with another model.

Before the broad launch, Google will continue testing Argon's protection against abuse, prompt injection, and other potential risks.

Initially, Gemini 4 Argon will be available to paid API users and Google AI Ultra subscribers.

Source: [blog.google](<https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon>)
