Google unveils top-tier AI 'Gemini 4'... general access restricted
- Input
- 2026-10-01 09:52:48
- Updated
- 2026-10-01 09:52:48

[Financial News] Google, which had released only lightweight artificial intelligence (AI) models, has finally unveiled a top-tier AI model. However, it has restricted public access for safety reasons.
Google announced on the 30th (local time) that it had launched its next-generation AI model, Gemini 4 Argon.
The model is Google's top-tier offering, released approximately seven months after the company introduced Gemini 3.1 Pro in February 2026.
Google said the model can understand and reason through entire complex, long-term workflows, allowing it to perform in areas ranging from coding to specialized knowledge-based work in finance, law and taxation, as well as cybersecurity.
In fact, Gemini 4 Argon scored higher than models from competitors OpenAI and Anthropic on coding benchmarks such as DeepSWE v1.1 and VibeCode Bench.
It also surpassed or outperformed competing models by a wide margin on benchmarks measuring financial and legal capabilities. On tests such as CWE-Bench v1, which assesses cybersecurity skills, it performed at roughly a tie for first place with GPT-6 Astra.
To capitalize on its strength in solving complex problems, Google expanded the output-token limit from the previous 64,000 tokens to 1 million, the highest level in the industry.
This allows the model to complete complex coding or analytical tasks spanning hundreds of pages with a single prompt.
Artificial Analysis, which evaluates AI models, gave Gemini 4 Argon an Artificial Analysis Intelligence Index (AAII) score of 53, placing it in a tie for third with Claude Fable 5.1 and GPT-6 Astra.
However, Google said that the risk of misuse had also increased along with the model's improved performance and decided to block access for general users for the time being.
Google plans to give the model priority access to vetted cybersecurity experts through the Fairwind Program, a security initiative led by the company.
Koray Kavukcuoglu, senior vice president (SVP) leading Google DeepMind, said, "A phased approach is necessary to safely release cutting-edge capabilities at this level," and explained, "We are gradually expanding access and voluntarily participating in the U.S. government's model pre-submission process."
In recent months, Google had repeatedly postponed the release of its Pro model after determining that the AI model's coding capabilities were inadequate.
Meanwhile, as Google came to be seen as falling behind OpenAI and Anthropic in the AI race, it carried out a leadership reshuffle under the direction of co-founder Sergey Brin, transferring de facto management control of DeepMind from Demis Hassabis to Koray Kavukcuoglu.
Shares of Google's parent company, Alphabet, surged more than 3% within minutes after the regular session ended when news of the model's launch emerged. They later gave back some of those gains, fluctuating around $348 as of 5:50 p.m. Eastern Time, up 1.2% from the closing price.
[email protected] Lee Seok-woo, international affairs correspondent Reporter