A limited rollout of Google’s newest AI model, Gemini 4, is underway for security experts through a program called Fairwind. The company reports that the model has already changed internal processes, ranging from quantum computing to codebase migrations, and is now being offered to trusted cyber defenders before a wider release.
A staged launch strategy accompanies the model’s release. Google has entered into a voluntary arrangement with the U.S. government regarding pre-release access to the model, as it increases availability step by step. As the company moves through its rollout, it will take in feedback from initial users before refining safety controls. That process of improvement precedes making Argon available to developers, enterprises, and consumers.
The Price and Token Limits
The Argon launch comes with a starting rate of $2 for each million input tokens and $10 for each million output tokens. Input tokens kept in cache carry a discount of 95% off the regular input token price.
Google has increased the model’s token limit substantially. The new cap is 1 million tokens, compared with the prior 64K tokens. According to Google, this expansion lets the model produce far more text in a single chain of thought, which enables it to reason through problems with greater depth and generate hundreds of thousands of tokens at once.
Inside Google’s Internal Workflows
The company points to three major projects where the model has made notable contributions. Across those projects, Googlers are already using Argon for specialized coding tasks, deeper research, and improving writing quality.
Quantum Algorithmic Optimization
Google’s quantum computing team used Argon to make more efficient use of the spacetime resources within subroutines that hold up key applications. One particular case saw the model outperform a published baseline by 40%, accomplishing the improvement in a matter of minutes.
Memory Efficiency
Argon agents ran through fleet-wide profiling telemetry on their own to spot memory optimizations that could be applied across Google’s data centers. Once the rollout happened, it freed up over 300 TiB of memory, with an estimated total savings of 500 TiB to 1 PiB.
Large-Scale Codebase Migrations
Google’s Argon agents are moving C/C++ codebases over to Rust. The work covers everything from core libraries such as re2 and libgav1, which carry tens of thousands of lines of code, all the way up to the Fuchsia OS Zircon kernel, which has 800K+ lines.
Given how essential many of these systems are, such wide-ranging changes are being put through strict checks. Machines and people alike are looking them over, running tests against copies of the old systems, and going over them carefully before they go live.
For libgav1, Google’s open-source software for decoding video, Argon agents took an existing Rust port and replaced 32K lines of SIMD code by running many rounds of profile-guided experiments, studying the compiler’s output, and producing safe Rust so the compiler would vectorize it automatically. The result is a memory-safe video decoder that runs 2.7x faster than the Rust port, with identical video output, bringing it closer to the optimized C++.
The State of the Art Across Domains
DeepSWE v1.1 has a new leader, and it is Argon, which scores 77.9%. The benchmark judges a model’s skill at handling real-world software engineering tasks that span long periods.
The Vals Index measures economic impact across finance, coding, legal, and tax work, with each sector weighted by its contribution to U.S. GDP. Argon leads on that ranking. The model also takes top spot on Vals Finance Agent v2 (multi-step financial research) and Harvey’s Legal Agent Benchmark (legal research and drafting).
Zapier’s benchmark, AutomationBench, measures end-to-end execution across core business functions. Argon sits at position #1, with a score of 51.3%.
When visual comprehension matters most, Argon holds its own among knowledge work tools. It handles professional chart analysis, spots details across lengthy videos, and acts upon a chain of documents. The test known as LVBench gauges long video understanding, and Argon sits atop it with a score of 91.7%, marking the current high point.
Cybersecurity Defense Capabilities
Google built Argon with a strong focus on helping cyber defenders face the new wave of attacks. The model was trained to work by itself in finding, checking, and fixing serious software weaknesses.
The security-focused Argon release will omit any cybersecurity guardrails for trusted defenders and Google’s own internal teams, allowing them to take advantage of its full frontier-level capabilities.
Argon has been put into service by Wiz for cybersecurity protection through its Scan for Good initiative, which is a project aimed at safeguarding essential public infrastructure without charge by identifying and correcting serious security weaknesses. As proof of its effect, the system detected a major flaw that exposed private data from healthcare systems relied upon by hospitals around the world.
CWE-bench v1 is an evaluation of the model’s ability to remediate security vulnerabilities, Argon ties for first place with a top score of 68%, building on 3.8 Flash Cyber’s frontier performance against CWE-bench v0.
Key Figures From the Announcement
- Argon launches at $2 per million input tokens and $10 per million output tokens, with cached input tokens at 95% off the input token price
- Output token limit: 1 million tokens, up from the previous 64K tokens
- Quantum optimization: beat a published baseline by 40% in minutes
- Memory savings: over 300 TiB freed, with an estimated 500 TiB to 1 PiB in total
- Video decoder speedup: 2.7x faster than the Rust port
- DeepSWE v1.1: 77.9%
- Vals Index: lead position
- AutomationBench: 51.3%
- LVBench: 91.7%
- CWE-bench v1: tied for first at 68%
The company is proceeding cautiously, working with the U.S. government and collecting input from people who tried the product first. Four areas show genuine technical advance: the healthcare security problem found through Wiz’s Scan for Good effort, the quantum calculation work, the reductions in memory use, and the changes made to the video decoder.
The plan shows trust in the project. Argon is being rolled out, and the company’s own successes point to a useful tool for developers, enterprises, and consumers alike.
Source material: “Gemini 4 Argon,” Google.
Get the Notebook.
The day's best stories and every fresh verdict, in plain English, in your inbox by seven. One email a day, no more.