Google Gemini 4 Argon Is Here: Everything You Need to Know

Google announced Gemini 4 Argon on September 30, 2026, calling it the company's new frontier model and, internally, its most performant yet. But "announced" is doing a lot of work in that sentence. Argon isn't showing up in your Gemini app, and there's no sign-up page for the general public. For now, it's rolling out to a narrow group: trusted cybersecurity defenders working through Google's Fairwind Program. If you're trying to figure out whether you can actually use Gemini 4 Argon today, what it's built to do, and when that might change, here's what's confirmed and what Google is still staying quiet about.

What Is Gemini 4 Argon?

Gemini 4 Argon is Google DeepMind's newest frontier model, introduced in a blog post by Koray Kavukcuoglu, SVP of Google DeepMind. Google describes it as built for "complex, long-horizon workflows," a deliberately broad phrase that covers three areas: real-world software engineering, enterprise knowledge work such as legal and financial analysis, and cybersecurity defense. It's not pitched as a chatbot upgrade for casual use; Google's own framing is about tasks that take many steps and a lot of context to get right.

The company says Argon is already running inside Google itself, with employees using it for specialized coding tasks, deeper research, and writing. A few of the internal use cases Google has disclosed: optimizing a quantum algorithm by 40% over a previously published baseline, freeing more than 300 TiB of memory across its data centers, and migrating parts of the Fuchsia operating system's Zircon kernel from C/C++ to Rust.

Why Gemini 4 Is Different

The single clearest technical change in Argon is output length. Google has raised the model's output token ceiling from 64,000 to 1 million tokens, which it calls an industry-leading number. In practice, that means tasks like threat-hunting across a large codebase or analyzing a long chain of logs, work that previously required stitching together multiple API calls, can now run in a single pass. Google also highlights improved multi-step reasoning, long-form video and chart analysis, and the ability to handle large-scale codebase migrations without losing track of earlier context.

None of this makes Argon a faster version of Gemini 3.8 for everyday questions. It's positioned, and priced, as a model for sustained, complex work rather than quick turnaround chat.

What Can Gemini 4 Argon Do?

Based on Google's own disclosures and early independent reporting, Argon's practical strengths fall into three buckets:

• Software engineering: large-scale codebase migrations, debugging, and autonomous vulnerability discovery and patching.
• Enterprise knowledge work: finance, legal, and tax-style reasoning tasks, where Google says Argon leads on the Vals Index, a benchmark built by the AI evaluation startup Vals.
• Long-context analysis: reviewing long videos or large charts in one pass, helped by the expanded 1-million-token output window.

One real-world example that's already public: the security firm Wiz used Argon through its "Scan for Good" initiative and says it found critical vulnerabilities in healthcare software that earlier frontier models had missed.

Gemini 4 Argon and Cybersecurity

Cybersecurity is the reason most people are hearing about Argon at all right now, because it's the only area where outside organizations can actually touch the model. Access runs through the Fairwind Program, Google's security-focused early access initiative. Fairwind opened in September 2026 with a smaller model, Gemini 3.8 Flash Cyber, and has since grown to more than 650 participating organizations, including named participants CrowdStrike and Palo Alto Networks.

Inside Fairwind, defenders get a version of Argon with its cyber guardrails removed so it can do the job properly: autonomously finding, validating, and patching software vulnerabilities. Google says it's pairing that capability with safety layers on its end, including internal activation monitoring, chain-of-thought misalignment checks, and sandboxed testing environments, and that it's engaged with the U.S. government's voluntary pre-release model access process. The company's own explanation for the staged rollout is straightforward: a model this capable at finding and exploiting vulnerabilities is also a model that could cause real harm if it reached the wrong hands before the guardrails were tested, so trusted defenders go first.

Gemini 4 Benchmarks and Performance

Google's self-reported numbers are strong. On DeepSWE v1.1, a real-world software engineering benchmark, Argon scores 77.9%. On CWE-bench v1, which tests vulnerability remediation, it ties for first place at 68%. On AutomationBench, a test of end-to-end business tasks, it ranks first at 51.3%. Google also cites a 91.7% score on LVBench for long-video understanding and says Argon leads the Vals Index across finance, coding, legal, and tax tasks.

These are Google's own benchmarks, run and reported by Google. For a second opinion, the independent evaluator Artificial Analysis ran Argon through its own Intelligence Index, a combined score across ten tests including coding, scientific reasoning, and agentic office work. There, Argon scores 52.6 points at its tested "High" compute setting, essentially level with GPT-6 Astra's 52.7 at its own highest setting, while Claude Opus 5.5 leads both at 57.6. On Humanity's Last Exam, a difficult knowledge and reasoning test, Argon's 57.1% beats GPT-6 Astra's 54.7% but trails Opus 5.5's 61.4%. On GDPval-AA, which scores real-world work across 44 occupations, Argon posts an Elo of 1,611 against Astra's 1,542 and Opus 5.5's 1,846.

That's a meaningfully different picture from Google's own marketing, which cites Vals to claim Argon scored "significantly higher" than both GPT-6 Astra and Anthropic's Fable and Opus models. The gap likely comes down to which benchmark suite and which compute setting each side is comparing. Worth knowing before you take either side's claim at face value.

Gemini 4 vs GPT and Claude

There's no single answer to "which model is best" here, and anyone claiming otherwise is picking their preferred benchmark. Google's own reporting, built around the Vals Index, positions Argon as the clear leader. Artificial Analysis's independent testing puts Argon roughly tied with OpenAI's GPT-6 Astra on overall intelligence, with Claude Opus 5.5 ahead of both by a noticeable margin, about five points on the Intelligence Index and over 200 Elo points on real-world work tasks. On the specific vulnerability-remediation benchmark CWE-bench v1, Google says Argon ties with GPT-6 Astra at 68%, rather than beating it outright.

For context on how Anthropic's current lineup stacks up on its own terms, see our breakdown of Claude Sonnet 5.5 vs Opus 5.5. The honest summary for Argon: it's a genuine step up for Google and closes the gap with OpenAI, but the independent data doesn't support a clean "Google is now winning" headline.

Can You Use Gemini 4 Right Now?

For almost everyone, no. Argon is not available in the consumer Gemini app, it doesn't have a general API key you can sign up for, and there's no Google AI Ultra bundling yet. The only way to use it today is through the Fairwind Program, and that program is specifically built around vetted cybersecurity organizations, not the general public or typical developers. If you've been using Gemini 3.8 Flash or comparing Gemini to ChatGPT day to day, that experience is unaffected by this announcement; Argon runs on a separate, restricted track for now. For a look at what is broadly available today, see our guide to Gemini 3.8 Flash and our ChatGPT vs Gemini comparison.

When Will Gemini 4 Be Available to Everyone?

Google hasn't given a date. The company's stated plan is to expand access to API customers and Google AI Ultra subscribers after the Fairwind phase, and it says it wants to move "as soon as possible." That phrase is not a commitment, and Google has not published a quarter, a month, or even a rough window. Given that Argon's cyber-focused guardrail testing is explicitly framed as ongoing, a cautious reading is that broader access depends on how that testing goes, not on a fixed calendar. Anyone telling you a specific release date for general availability right now is guessing.

What Gemini 4 Means for AI Users

If you're not a cybersecurity professional, the immediate, practical impact of this announcement is close to zero. Your Gemini app doesn't change today. What it does signal is where Google is pointing its most capable model first: not at chat, but at high-stakes, high-complexity defensive security work, with enterprise and developer access to follow later. It also tells you something about how frontier labs are now handling genuinely dangerous capabilities, autonomous vulnerability discovery chief among them, by staging access through vetted partners instead of a simultaneous public launch. Expect more "limited rollout before broad release" announcements like this one as models get better at tasks that are useful for defense and just as useful for attack.

Frequently asked questions

Sources

• Google: Gemini 4 Argon, our next era of frontier intelligence
• TechCrunch: Google releases Gemini 4 Argon
• SiliconANGLE: Gemini 4 Argon goes to cybersecurity defenders first
• The Decoder: independent benchmark analysis (Artificial Analysis)
• VentureBeat: Gemini 4 Argon's limited release

Related articles

Comments