← Back
LLMsNEW

Gemini 4 Argon, Explained: Google's Frontier Model You Can't Use Yet

On September 30, 2026, Google announced Gemini 4 Argon, the first model of the Gemini 4 series, claiming frontier performance in software engineering, enterprise knowledge work and cyber defense. For now it is available only to vetted cyber defenders through Google's Fairwind Program, with no API pricing and no public release date. Here is what Google says it can do, and why you cannot have it yet

5 min read
Updated Oct 1, 2026
QUICK ANSWER

On September 30, 2026, Google revealed Gemini 4 Argon, the first model of its Gemini 4 series and its first real attempt to retake the frontier in months.

Key Takeaways
  • This guide provides comprehensive, actionable information
  • Consider your specific workflow needs when evaluating options
  • Explore our curated LLMs tools for specific recommendations

What Google actually announced

On September 30, 2026, Google revealed Gemini 4 Argon, the first model of its Gemini 4 series and its first real attempt to retake the frontier in months. The Verge quotes Koray Kavukcuoglu — Google's chief AI architect and the DeepMind SVP who took over the lab in August — promising "frontier performance in complex workflows across real-world software engineering, enterprise knowledge work like legal and finance, and cybersecurity defense." Business Insider frames it as Google reasserting itself in high-stakes AI after a summer of smaller Flash releases while OpenAI and Anthropic pulled ahead on engineering and security.

The name itself carries context. Google promised Gemini 3.5 Pro in June, postponed it repeatedly over performance, and quietly never shipped it. As Ars Technica puts it: so much for Gemini 3.5 Pro. Argon is the model that actually made it out, and the skipped generation number is Google saying so.

77.9%
on the DeepSWE v1.1 software engineering benchmark, ahead of GPT-6 Astra, Fable 5.1 and Opus 5.5 in Google's chart
1M
output token limit confirmed, up from 64,000 in previous Gemini models, so whole jobs fit in one step
Fairwind only
initial access is limited to vetted governments and cyber authorities, no public date or API pricing

The benchmarks, all first-party for now

Every number below comes from Google's own announcement materials, published by Ars Technica and Business Insider. Nobody outside Google has run the model yet, and the AI community had already been buzzing on X over what appeared to be leaked Argon benchmarks earlier in the week.

On software engineering, Google claims 77.9% on DeepSWE v1.1, higher than GPT-6 Astra, Fable 5.1 and Opus 5.5. On security, Google says Argon ties GPT-6 Astra for the highest score on CWE-bench, which tests whether a model can find and patch real vulnerabilities. On long-horizon knowledge work, it points to an industry-leading score on the Vals Index economic analysis test. The pattern across all three is deliberate: this launch is aimed at exactly the workloads — coding, legal and finance, cyber defense — where Google had fallen behind.

One hardware-flavoured detail supports the engineering claim from a different angle: Argon reportedly used fleet-wide telemetry data to help Google save 300 TiB of memory across its data centers without new hardware, according to Ars Technica.

Why you can't have it yet

Access is deliberately narrow. Google is making Argon available to "a set of trusted cyber defenders" — partners in its Fairwind Program, an initiative for vetted governments and cyber authorities to test new models proactively and address cyber vulnerabilities, per Business Insider. Kavukcuoglu says Google is "actively engaged in the U.S. government's voluntary process for pre-release model access while we gradually expand access."

What that means in practice:

Public API
Not announced. Google has not published pricing, and there is no release date for general availability.
Output limit
Confirmed at 1 million output tokens, a large jump from the 64,000-token ceiling of previous Gemini models, so long tasks complete in one step instead of many chained calls.
Inside Google
Argon already runs internal workflows. Google says engineers use it for debugging and large-scale codebase migrations, including converting C and C++ to Rust: thousands of lines in the re2 and libgav1 libraries and more than 800,000 lines in the Fuchsia OS Zircon kernel.

The dogfooding is not just marketing. Migrating a kernel of that size is exactly the long-horizon engineering task the benchmarks claim, and it is the kind of work Google can verify end to end before anyone outside touches the model.

The safety story Google is telling

Google says it is strengthening "critical frontier safeguards" before a broader rollout, including defenses against misuse and prompt injection and monitoring for misalignment, per The Verge. Business Insider adds two specific mechanisms: systems that track Argon's chain of thought and can stop it from acting when necessary, and precautions around how the model is given feedback, so it cannot be trained — accidentally or deliberately — to evade its own monitoring.

The restricted release is part of the same posture. A model that only vetted defenders can use is a model whose failure modes can be observed in a controlled population first. Whether that reads as responsible staging or as an admission that the frontier is outrunning public deployment is a fair question, and the week's other news feeds it: OpenAI cancelled its planned GPT-6.1 Astra model over safety worries days earlier, and the FTC opened probes into OpenAI and Anthropic the same day Argon was announced.

Who should care

You run a security team at a large organisation
Fairwind is your path in. It is worth asking your Google account team whether your organisation qualifies, since vetted early access is the only access there is.
You are deciding which frontier model to build on
Do not plan around Argon yet. There is no date and no price, and every benchmark is first-party. Keep your current stack; revisit when independent runs and API pricing exist.
You are a Google Cloud customer
History suggests that when Argon opens up, it lands on Google's own platforms first. If your workloads already run there, the migration test will be cheaper for you than for anyone else.
You watch AI benchmarks for a living
The DeepSWE and CWE-bench claims are the ones to wait on. They are checkable, they matter commercially, and independent scores will say more about the frontier than any launch chart.

For the model Argon replaces in Google's lineup, see the Gemini 3 Ultra listing. For what Google shipped the last time it pushed Gemini forward, see Gemini 3.8 Live. For the wider frontier picture, the best LLMs in 2026 and LLM pricing: complete cost comparison keep track of the field. The announcement day is covered in the September 30 briefing.

FREQUENTLY ASKED QUESTIONS
What is Gemini 4 Argon, what can it do, who gets access to it, and when will it be publicly available?
On September 30, 2026, Google revealed Gemini 4 Argon, the first model of its Gemini 4 series and its first real attempt to retake the frontier in months.
EXPLORE TOOLS

Ready to try AI tools? Explore our curated directory:

SHARE THIS GUIDE

On September 30, 2026, Google revealed Gemini 4 Argon, the first model of its Gemini 4 series and its first real attempt to retake the frontier in months.

Share on X LinkedIn Reddit Email
Copied to clipboard