Google has introduced Gemini 4 Argon, the first model in its Gemini 4 generation, with an initial rollout limited to selected cybersecurity defenders, trusted partners and Google’s internal teams.
Announced on September 30, 2026, through Google’s Fairwind Programme, the model is designed to handle advanced software engineering and cybersecurity tasks, including identifying, validating and fixing critical software vulnerabilities.
However, Google is taking a cautious approach to its release, citing the potential security risks associated with making such powerful capabilities widely available.
What makes Gemini 4 Argon different?
According to Google, Argon can autonomously identify software vulnerabilities, validate potential weaknesses and develop patches with limited human intervention.
The model also recorded a 77.9% score on the DeepSWE v1.1 software engineering benchmark. Google reported scores of 74.2% for Anthropic’s Claude Opus 5.5 and 74.1% for OpenAI’s GPT-6 Astra on the same test.
On CWE-bench v1, which evaluates vulnerability identification and remediation, Argon scored 68%, matching GPT-6 Astra, according to Google’s evaluations.
Argon also supports up to one million tokens in a single response, allowing it to process and generate significantly longer outputs for complex coding and research tasks.
These figures are based on Google’s reported evaluations and will need broader independent testing to establish how the model performs across real-world applications.
How Google is already using Argon
Google says thousands of its employees are already using the model for coding, research and specialised engineering tasks.
One notable application is migrating large C and C++ codebases to Rust, a programming language known for its memory-safety features. The projects range from smaller software libraries to the Zircon kernel used by Google’s Fuchsia operating system, which contains more than 800,000 lines of code.
The company is also using Argon to improve data-centre efficiency. Google estimates that its AI agents could free up between 500 tebibytes and one pebibyte of memory through optimisation work.
Its quantum computing researchers have also used the model to improve an algorithm, beating a published baseline by 40% in one reported experiment.
Why Google is limiting access
Argon’s cybersecurity capabilities are central to Google’s decision to restrict its initial release.
Although the model can help organisations discover and fix security weaknesses, similar capabilities could potentially be misused to identify vulnerabilities for cyberattacks.
Through the Fairwind Programme, Google is giving selected cybersecurity partners early access to test the model and strengthen digital defences before a wider rollout.
The company is also participating in the US government’s voluntary pre-release assessment process as it continues refining its safeguards.
Google has announced an introductory API price of $2 per million input tokens and $10 per million output tokens. After the introductory period, these prices will increase to $4 and $20 respectively. Cached input tokens will receive a 95% discount on the input price.
For now, the model is available to selected cybersecurity defenders. Google says paid API customers and Google AI Ultra subscribers will be among the first groups to receive access when the rollout expands, although no specific date has been announced.
The launch comes as Google, OpenAI and Anthropic continue developing increasingly capable AI models for software engineering, enterprise operations and cybersecurity.
But Gemini 4 Argon introduces another important consideration in the AI race: as models become capable of identifying and fixing complex security vulnerabilities, deciding who gets access to those capabilities becomes just as important as improving the technology itself.
