Google Unveils Gemini 4 Argon: 1 Million Token Output and Cyber Defense
Originally published in Turkish. Read the original →
Google unveiled Gemini 4 Argon, its new flagship model with a 1 million token output capacity and advanced cybersecurity capabilities.
As competition among frontier models in the AI ecosystem continues unabated, Google has made its long-awaited move and officially announced Gemini 4 Argon, which it describes as its most capable AI model to date. Designed especially for software engineering, autonomous research processes, finance, law and long-running workflows that require complex reasoning, Argon stands out for its cybersecurity and code migration capabilities.
Given the model's high operational capacity, Google is adopting a phased rollout. Accordingly, Gemini 4 Argon has initially been made available to trusted cyber defense teams in the company's Fairwind Program. The goal is to push the model's limits with real cyberattack and defense scenarios before it opens to general use.
The Capacity to Generate 1 Million Output Tokens in One Go
Among the most radical technical innovations of the Gemini 4 Argon architecture is the output limit on the generation side. While earlier Gemini versions could generate at most 64 thousand tokens in a single response, with Argon this ceiling has been raised roughly 16-fold to a 1 million token output capacity.
This huge output window makes it possible for the AI not just to read long documents but to build huge software projects from start to finish in a single stream, produce technical documentation hundreds of pages long and carry out multi-step reasoning chains without interruption. Developers can get results with high consistency in a single session, without wasting time splitting complex system architectures into small pieces and stitching them back together.
Internal Use at Google and Migrating C/C++ Code to Rust
Before opening its new model to external users, Google began testing it extensively on its own massive corporate infrastructure. Thousands of engineers within the company are using Argon-powered agents for code review, debugging and large-scale refactoring operations.
The most striking example of these operations is Google's project to migrate its historical C and C++ codebases to the memory-safe Rust language. Critical codebases ranging from core libraries tens of thousands of lines long to the Fuchsia Zircon kernel, made up of more than 800 thousand lines, are being converted autonomously by Argon agents. These migrations, which go through multi-layered automated and manual reviews before reaching production, are setting a new standard in software security.
The open source libgav1 video decoder stands out as concrete proof of this. Argon agents converted the 32 thousand lines of SIMD (Single Instruction, Multiple Data) code in the Rust version of this software into modern, safe Rust code that the compiler can vectorize flawlessly. Measurements confirmed that the new version runs a full 2.7 times faster than the previous Rust implementation and produces bit-for-bit identical video output.
Memory Optimization in Data Centers and Quantum Computing
Gemini 4 Argon's problem-solving ability isn't limited to converting source code; it also feeds directly into hardware and infrastructure efficiency. Analyzing the profiling data from Google's data centers, the model automatically detected memory leaks and unnecessary memory allocations.
As a result of the first fixes, more than 300 TiB of memory was freed up directly in the data centers. Google engineers expect total memory savings to reach 500 TiB to 1 PiB once these optimizations are integrated into all systems globally.
A notable success was also recorded on the quantum computing side. A core subroutine that forms a critical bottleneck in the practical implementation of quantum algorithms was optimized by Argon. Within minutes, the model discovered and presented to researchers an innovative mathematical approach that uses 40 percent fewer computational resources than the best baseline result published in the literature.
A New Threshold in Cybersecurity: Defense-Focused Agents
The area where the model is positioned first is undoubtedly cybersecurity. It can automatically discover software vulnerabilities through static and dynamic analysis, prove a vulnerability's exploitability in isolated environments and generate a security patch instantly.
Wiz, the global cloud security platform included in the early testing process, validated the model's capabilities in a real-world environment. According to the information shared, Argon managed to detect a critical vulnerability in healthcare software actively used in hospitals around the world that could have exposed sensitive patient records. This flaw, which earlier frontier models had missed, was reported and patched thanks to Argon's deep context analysis.
Sharing industry leadership with a score of 68 percent on the independent CWE-bench v1 test, which measures the repair of security vulnerabilities, the model can successfully catch zero-day vulnerabilities in complex repos written in 20 different programming languages.
Benchmarks and the Performance Balance
The benchmark data shared by Google shows that Gemini 4 Argon sits at the top of its class in many enterprise and technical tasks:
- DeepSWE v1.1: A 77.9 percent success rate on real-world software engineering tasks.
- LVBench (Long Video Understanding): Superior comprehension in long-duration visual and video tracking with a score of 91.7 percent.
- AutomationBench: A 51.3 percent completion rate in end-to-end business processes and automation scenarios.
On the other hand, the model doesn't hold absolute leadership in every area. It stays at 57.4 percent on Terminal-Bench 4.0, which measures terminal- and shell-focused tasks, and at 55 percent on FrontierSWE, putting it a step behind its rivals in some independent tasks. Google says this stems from the model's strict focus on security and enterprise reliability constraints in particular.
API Pricing and Rollout Schedule
To help developers adopt the Gemini 4 Argon ecosystem quickly, Google announced an attractive limited-time introductory rate:
- Introductory Price: $2 per 1 million input tokens and $10 per 1 million output tokens.
- Standard Rate: After the introductory period, $4 per 1 million input tokens and $20 per 1 million output tokens will apply.
After the cybersecurity reviews under the Fairwind Program are complete, the model will first be offered to Google AI Ultra subscribers and paid API developers, then gradually rolled out to a wider user base.
