Google has launched Gemini 4 Argon, its latest most advanced model, and it is built for the kind of work that pushes AI beyond quick answers. The model is designed to sustain deep reasoning through complex questions, bringing finance, software engineering, coding, creative writing, and cybersecurity defense into one powerful system.
That ambition puts Argon directly into the race against frontier models from other companies. Google sees it as comparable with OpenAI’s GPT-6 Astra and Anthropic’s Opus on key benchmarks, while the model also scored one point ahead of OpenAI’s GPT-6.1 Sol.
Aiming for Frontier Performance at a Lower Cost
Argon’s benchmark position matters, but its pricing could make the launch just as important for developers and businesses. Gemini 4 Argon matches GPT-6 Astra’s score on the Intelligence Index at 60 percent of the cost per task at current discounted prices.
Google lists an introductory price of $2 per million input tokens and $10 per million output tokens. Astra costs $10 per million input tokens and $50 per million output tokens, giving Argon a clear price advantage for workloads that process large amounts of information or generate long responses.
The model also has an output token limit of 1 million tokens. That is far beyond GPT-6 Astra’s 128,000-token output limit, creating room for long reasoning chains, extensive code work, detailed document tasks, and other projects that need sustained output in one session.
Argon has also posted a 15 percent hallucination rate, the lowest among leading models in the listed comparisons. GPT-6 Astra and GPT-6.1 Sol each have a 54 percent hallucination rate, making accuracy a central part of Google’s case for its new model.
From Long Documents to Critical Vulnerabilities
Google says Argon can analyze charts professionally, identify details from long-form videos, and perform tasks based on documents. Those abilities give the model a broad working range, especially when a complex assignment depends on multiple types of information rather than a single text prompt.
The company has also trained Argon for cybersecurity defense. The model can autonomously find, validate, and patch critical software vulnerabilities, connecting detection with the work needed to address a problem.
During an early demonstration, Argon spotted a critical vulnerability in healthcare software used by hospitals around the world that exposed sensitive information. On the CWE-bench leaderboard for cybersecurity capabilities, the model tied for first place with Grok 4.7 and GPT-6 Astra.
That result gives Argon a role in security work that reaches beyond code generation. Finding a weakness, confirming that it is real, and creating a patch demand different capabilities, and Google says the model can handle each part of that process.
Built for Google’s Most Demanding Work
Google is already using Gemini 4 Argon for quantum computing research and codebase migrations. The model has also helped free up 300 TiB of memory in Google’s data centers, showing how its impact extends into the systems that support the company’s own operations.
Safety controls sit alongside those capabilities. Google designed Argon to be resilient to prompt injections, a key defense when outside instructions try to redirect a model away from its intended task. Google is also deploying misalignment mitigations to prevent Argon from acting on its own without prompting from the user.
Those protections matter because a model with a 1 million-token output limit and the ability to work through complex technical tasks can operate across a huge amount of information. The goal is to preserve its ability to reason and execute while keeping user direction at the center of its behavior.
The Rollout Starts With Early Users
Gemini 4 Argon follows Gemini 3.5 and is now rolling out to members of Google’s Fairwind Program. The model will eventually reach developers, enterprises, and general users, starting with paid API customers and Google AI Ultra subscribers.
That rollout will put Argon’s claims under pressure from real workloads. Developers will test its price and output limits, enterprises will examine its reliability on long-running tasks, and security teams will measure how well it finds and patches vulnerabilities outside an early demonstration.
Google has positioned Gemini 4 Argon as a model for deep reasoning, broad multimodal work, cybersecurity defense, and demanding internal research. With frontier-level benchmark results, a lower listed price than GPT-6 Astra, and access to a 1 million-token output limit, Argon enters the next stage of AI competition with a very large target in sight.
Based on
