Google shipped a new frontier model on September 30, and you cannot sign up to use it — not today, not with a credit card, not through any waitlist form. Gemini 4 Argon is built for exactly the kind of work most people never see: patching real vulnerabilities in production codebases, mapping attack surfaces, and reasoning across security logs at a scale no human team could manage. Google is routing initial access through something it calls the Fairwind Program, open only to a set of trusted cyber-defense teams it has personally vetted.
That's a deliberate, unusual launch strategy for a model this capable, and it's worth understanding both what Argon can actually do and why Google is choosing to gate it this tightly before anyone else gets near it.
Quick summary: Google announced Gemini 4 Argon on September 30, 2026, built for coding, enterprise knowledge work, and cybersecurity defense. It ties GPT-6 Astra and Grok 4.7 for first place on CWE-bench v1 (all at 68%, a benchmark for patching real vulnerability classes), with Claude Opus 5.5 at 67% and Claude Fable 5.1 at 58%. On the DeepSWE v1.1 coding benchmark, Argon scores 77.9% versus Astra's 74.1% and Opus 5.5's 74.2%. Its output limit jumped to 1 million tokens (from 64,000), and on long-context tasks between 256K and 1M tokens it holds 84.2% accuracy versus Astra's 71.8%. Introductory API pricing is $2 per million input tokens / $10 per million output tokens. Access is currently restricted to vetted cyber-defense teams through Google's Fairwind Program, with paid API and Google AI Ultra access promised "as soon as possible" but no firm date given.
What Argon Actually Does Well
CWE-bench v1 tests whether a model can take a real, known vulnerability class and actually remediate it — not just describe the bug. Argon ties GPT-6 Astra (and Grok 4.7) for the top published score at 68%, ahead of Claude Opus 5.5 at 67% and Claude Fable 5.1 at 58% (Source: emergent.sh). A tie at the top of a security-remediation benchmark is a meaningfully different kind of result than a narrow single-point lead — it suggests Argon and Astra are currently operating in the same tier on this specific skill, not that either has a clear edge.
On DeepSWE v1.1, a broader coding benchmark, Argon pulls ahead more clearly at 77.9%, against 74.1% for Astra and 74.2% for Opus 5.5 — a real gap, if still a single-digit one on a benchmark covering general coding rather than security specifically (Source: redreamality).
The long-context numbers are where the gap widens most. Both Argon and Astra sit near the ceiling at shorter context lengths — 99.7% and 98.7% respectively up to 128K tokens — but between 256K and 1M tokens, Argon holds at 84.2% while Astra drops to 71.8%. Google also raised Argon's maximum output from 64,000 tokens to 1 million, useful for tasks like generating a full remediation report across an entire codebase in one pass. Google has not stated Argon's maximum input context window directly, so take any specific input-side number you see elsewhere as unconfirmed.
| Model | CWE-bench v1 | DeepSWE v1.1 | GraphWalks long-context (256K–1M) |
|---|---|---|---|
| Gemini 4 Argon | 68% | 77.9% | 84.2% |
| GPT-6 Astra | 68% (tied) | 74.1% | 71.8% |
| Claude Opus 5.5 | 67% | 74.2% | 66.8% |
| Claude Fable 5.1 | 58% | 67.4% | 65.0% |
Why You Can't Just Sign Up
Google is routing Argon's first access through the Fairwind Program — a vetted group of cyber-defense teams, not the general developer public. Argon also participates in the U.S. government's voluntary pre-release review process. Google says paid API customers and Google AI Ultra subscribers will get access "as soon as possible," but has not committed to a date (Source: neoteo). The practical read: a model this good at finding and patching vulnerabilities is also, by definition, good at finding vulnerabilities to exploit — and Google appears to be sequencing defenders ahead of everyone else on purpose.
Benchmark ties and narrow leads shift fast in this market — Argon tying Astra at 68% today doesn't mean it'll still be tied once either lab's next point release lands. Treat any single-digit gap on these benchmarks as a snapshot, not a verdict.
Who This Actually Matters For
- Security teams at organizations large enough to qualify for Fairwind vetting, who want AI-assisted vulnerability remediation across large codebases now, not in six months.
- Open-source maintainers and bug-bounty researchers watching for when broader API access opens, since Argon's long-context score is the most interesting number here for scanning sprawling repositories in one pass.
- Developers already comparing GPT-6 Astra against Claude Opus 5.5 for general coding work — Argon isn't available to you yet, but its DeepSWE score suggests it's worth benchmarking the moment it is.
It's not relevant yet to anyone who just wants a chat assistant — Argon isn't a consumer product, and Google hasn't signaled it will become one in the near term.
01Can I use Gemini 4 Argon right now?
Not unless your organization has been vetted into Google's Fairwind Program as a trusted cyber-defense team. Google says broader access for paid API customers and Google AI Ultra subscribers is coming "as soon as possible," but has not given a date.
02Is Gemini 4 Argon better than GPT-6 Astra?
It depends what you're measuring. The two are tied at 68% on CWE-bench v1 (vulnerability remediation). Argon leads on DeepSWE v1.1 (77.9% vs. 74.1%) and on long-context tasks between 256K and 1M tokens (84.2% vs. 71.8%). Neither model is unambiguously ahead across the board.
03How much will Gemini 4 Argon cost?
Google's introductory pricing is $2 per million input tokens and $10 per million output tokens, labeled explicitly as introductory — meaning it could change once the model reaches general availability.
04Why is Google restricting access instead of launching to everyone?
Argon is built partly for cybersecurity defense work, and a model skilled at finding and patching vulnerabilities carries real dual-use risk if it reaches bad actors before defenders have had a head start. Google is also putting Argon through the U.S. government's voluntary pre-release review process.
If you want to see how models like GPT-6 Astra, Claude Opus 5.5, and Claude Fable 5.1 actually perform on your own coding or research tasks — without waiting on any single lab's access list — LumiChats gives you one subscription with 40+ AI models in one place, so you can compare them directly instead of reading benchmark tables alone.
