Google announces Gemini 4 Argon, a model that generates responses of up to one million tokens, aimed at coding, finance, legal and cybersecurity tasks. Before opening up to businesses and users, it is reaching cyber defense specialists.
Google has announced Gemini 4 Argon, its new high-end artificial intelligence model, designed to sustain long reasoning on complex tasks. It is aimed at software engineering, professional work in areas such as law and finance, and cybersecurity. For now, only trusted cyber defense specialists can use it, through the Fairwind Program, while Google takes part in the U.S. government's voluntary process for accessing models before their release.
The main technical change is the output limit, the amount of text the model can generate in a single response, which rises from 64,000 to one million tokens.
In the tests published by Google, which compare it with GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1, Argon gets the highest score in most evaluations. On DeepSWE v1.1, which covers real-world software engineering tasks, it scores 77.9%, compared with 74.2% for Claude Opus 5.5 and 74.1% for GPT-6 Astra. On AutomationBench, focused on business processes, it reaches 51.3%, almost nine points ahead of the next model. However, it falls behind on other coding tests, such as FrontierSWE v2, where GPT-6 Astra scores 65.5% against its 55%. On CWE-bench v1, which measures the fixing of vulnerabilities, it ties at 68% with GPT-6 Astra and Grok 4.7.
Google already uses it internally, for example to migrate C/C++ code to Rust, including the Zircon kernel of Fuchsia, with more than 800,000 lines. Trusted defenders will receive Argon without the usual cybersecurity restrictions. The company Wiz uses it in its free Scan for Good initiative, where it detected a critical vulnerability exposing personal data in healthcare software used by hospitals worldwide.
Before expanding access, Google is strengthening safeguards against misuse, malicious prompt injection and actions that go beyond what the user asked for. On the Gray Swan test, prompt injection attacks succeed in 0.7% of cases with Argon, compared with 1% for Claude Opus 5.5 and Claude Fable 5.1 and 8.5% for GPT-6 Astra. The introductory price will be $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 after the introductory period. The first to get access will be paying API customers and Google AI Ultra subscribers.
Gemini is Google's artificial intelligence assistant developed by DeepMind. Works with text, images, audio, video, and code. Generates content, answers questions, and connects with Gmail, Calendar, ...
Google AI develops advanced platforms that improve people's lives. Its Gemini ecosystem integrates models, products, and APIs, driving responsible innovation and enabling developers and businesses to ...
05/11/2026
OpenAI introduces GPT-6 Cyber, a cybersecurity-focused model that identifies vulnerabilities, develops exploits and helps fix security flaws. Its ...
29/10/2026
Mistral AI publishes on Hugging Face the weights of its trillion-parameter model, trained from scratch on European infrastructure, under a license ...
27/10/2026
Google DeepMind introduces Genie 4, a model that generates interactive worlds in real time with realistic images, movement, sound and dialogue. It ...
21/10/2026
Anthropic opens a research preview of Claude Duet in the US by waitlist, an app for iOS, Android and claude.ai that checks connected apps, flags what ...