Google announced Gemini 4 Argon on September 30. If you opened the Gemini app afterward hoping to try it, you didn't find it. That's not a bug. Right now Argon is only available to trusted testers and a group of cybersecurity partners in Google's Fairwind Program. Everyone else is still on Gemini 3.8 Flash.
So this isn't a hands-on review. I haven't tested Argon, and almost nobody outside Google has. What I can do is go through what Google actually announced, separate the numbers that matter from the marketing, and tell you whether it's worth paying attention to.
Google took a detour to get here. Gemini 3.5 Pro was cancelled over the summer, and the company spent the past few months shipping one Flash model after another: 3.5, 3.6, 3.7 and 3.8, all between May and September. Argon is Google's first new flagship since Gemini 3.1 Pro in February. It also starts a new naming scheme, a version number plus a name, which looks a lot like what OpenAI is doing with GPT-6 Astra, Sol and Luna.
The headline spec is output length. Argon can generate up to 1 million tokens in a single response, up from 64,000. If you mostly type questions into a chat box, that number means nothing to you. You will never ask for a million-token answer. It matters for long, autonomous work: a model that can keep reasoning and writing code for hours without hitting a ceiling. Google's pitch is basically "give it a hard problem and let it think."
Google is focusing on four areas: coding, knowledge work, cybersecurity defense and creative writing. Cybersecurity gets the most attention. Google says Argon can find, verify and patch serious software vulnerabilities on its own, and approved security teams will get a version without the usual cyber guardrails.
💡 Read this first: every number below comes from Google's own announcement. No independent testing has been published yet.
| Benchmark | What it tests | Argon | Competitors |
|---|---|---|---|
| DeepSWE v1.1 | Real-world software engineering | 77.9% LEADER | Claude Opus 5.5 74.2%, GPT-6 Astra 74.1% |
| AutomationBench | Business workflows across multiple apps | 51.3% (#1) | — |
| LVBench | Understanding long videos | 91.7% | — |
| CWE-bench v1 | Fixing security vulnerabilities | 68% (tied for #1) | — |
A 3.7-point lead on DeepSWE is real, but it's not a blowout. Claude Opus 5.5 and GPT-6 Astra are almost tied with each other, and both are within striking distance of Argon. If Claude or ChatGPT already works well for your coding, this table alone isn't a reason to switch.
On launch day, Bloomberg reported that some people inside Google are skeptical of how well Argon performs in key areas, including coding. That's an awkward fit with a launch that leads with a coding benchmark. It doesn't mean Google's numbers are wrong. Benchmarks and everyday use often disagree, and that's exactly why I run my own side-by-side tests on this site instead of just repeating scores.
The more interesting part of the announcement isn't the benchmark table. It's what Google says Argon has already done inside the company:
These are Google's claims about its own work, so read them with that in mind. But they're far more concrete than "state-of-the-art reasoning."
At launch, Argon's API will cost $2 per million input tokens and $10 per million output tokens. That's an introductory price. Once it ends, the price doubles to $4 and $20.
Here's what stood out to me: $2 and $10 is exactly what OpenAI charges for GPT-6 Sol, its mid-tier model, not its flagship. GPT-6 Astra costs $50 per million output tokens. If Argon is anywhere near as good as Google says, that intro price is aggressive. The catch is the word "introductory."
For regular users, Google hasn't said anything about the free plan. Argon will roll out first to Google AI Ultra subscribers and paid API customers.
When Argon is available to regular users, I'll run it through the same tests I've done with ChatGPT, Claude and Gemini: same prompts, real screenshots, results checked by hand. Until then, treat every number in this post as Google's claim, not a verdict.
Sources: Google's announcement as reported by 9to5Google and TechCrunch; Bloomberg (September 30, 2026).
Take our 5-question quiz and get a personalized recommendation for what you actually do every day.
🎯 Find My AI →