Gemini 4 Argon: Google's new flagship model is ready for development and cyber defense

What's new

On September 30, 2026, Google announced Gemini 4 Argon, which the company calls its new flagship model. It was first rolled out to a small group of selected cyber defenders through Google's Fairwind program. On October 13, Google expanded access: the model can now be used in the Gemini API (paid tier, the name is gemini-4-argon), in AI Studio, in the Gemini app and in AI Mode in Google Search for AI Pro and Ultra subscribers. Ultra subscribers can also use it in the Antigravity tool.

Google has also raised the limit on how long a response from the model can be to 1 million tokens (small chunks of text), up from 64,000 previously. Google calls this industry-leading. The combined limit for input and output in the API is 1,048,576 tokens.

The sources do not say whether the model is available in Denmark or the EU. The rollout is described as gradual.

What's smart

Google says Argon is built for long, deep workflows in software development, knowledge work such as law and finance, and cyber defense. According to Google, thousands of employees have used it internally for debugging, large code migrations and algorithm design.

On the security side, Google says the model can independently find, validate and fix serious security vulnerabilities in software. Selected defenders and Google's own teams get the model without the usual cyber safeguards. The security firm Wiz is already using it in the Scan for Good program, which provides free protection for critical public infrastructure.

The model supports code execution, function calling, file search and URL context, among other things.

Cheaper or better?

In the API, Argon costs $2 per million input tokens and $10 per million output tokens during an introductory period. Cached input tokens, meaning text the model has seen before, are 95 percent cheaper. After the introductory period, the price rises to $4 per million input tokens and $20 per million output tokens. One source states that the increase takes effect on November 30, 2026. Monthly prices for AI Pro and Ultra are not given in the sources.

Google's own benchmarks produce impressive numbers: 77.9 percent on DeepSWE v1.1 (long-running software tasks), 51.3 percent on Zapier's AutomationBench (ranked No. 1) and 91.7 percent on LVBench for long-video understanding. Google also says the model leads on Vals Index, which measures economic impact in finance, coding, law and tax. The figures are Google's own and have not been independently verified.

The picture is not entirely rosy. According to reports, Argon trails competitors on the FrontierSWE v2 and Terminal-bench 4.0 tests. According to TechNews Taiwan, Bloomberg has reported that Google employees, citing anonymous sources, doubt its real-world performance, especially in frontend design for apps and websites. Some believe Anthropic's Claude and OpenAI's Astra are advancing faster, but opinions are divided. Competitors are also on the move: Anthropic launched Claude Opus 5.5 on September 22, and OpenAI released discounted GPT-6 variants.

What it's good at

According to Google and the internal sources, Argon is strongest at cybersecurity and multimodal tasks, such as extracting information from video. For long-running work in programming, law and finance, Google's figures also point in a positive direction.

For businesses, the low introductory price means it can be cheap to try the model now, but the price doubles later. The coverage encourages users to test it on their own tasks, especially when it comes to designing apps and websites. Google says it continues to collect feedback and adjust safety safeguards as access expands.