Google announces Gemini 4 Argon, limited to Fairwind cyber defenders at $2/$10 per million tokens
- Google announced Gemini 4 Argon, a frontier model with a 1 million token limit built for long-horizon work in software engineering, legal and finance knowledge work, and cyber defense, rolling out first to trusted cyber defenders through the Fairwind Program rather than to developers or consumers.
- Introductory API pricing is $2 per million input tokens and $10 per million output tokens, with cached input at 95% off the input price, rising to $4 and $20 after the introductory period.
- Google says Argon already runs internal work: quantum researchers used it to beat a published spacetime-resource baseline by 40% in minutes, and Argon agents freed over 300 TiB of data center memory, with 500 TiB to 1 PiB expected in total savings.
- Argon agents are migrating C/C++ codebases to Rust across Google, listed alongside specialized coding tasks as an internal use of the model.
- Google says it takes part in the U.S. government's voluntary pre-release model access process and avoids feeding monitoring findings back into training so Argon's reasoning does not learn to evade monitoring.
Hacker News opinions
So the model that is 'rolling out as soon as possible' isn't actually out. Gemini can't beat the can't-release-a-model allegations.
When I said I was tired of Google launching waitlists, I didn't think they'd respond by just not having a waitlist at all.
My Gemini app updated today and the newest model I can pick as a paying Pro user in the US is 3.6. 3.7 shipped in August and 3.8 in early September. Something is off over there.
It's the standard playbook now. 'Too dangerous to release right away,' plus a random element suffix because OpenAI does that too. Monkey see, monkey do.
Every Gemini release goes the same way. Can't release it, then it won't load in a harness normal people use for three weeks, then it's smart but useless at tool use and coding, then it's behind everyone.
I'm fine with it. I already pay $300+ a month for subs, please don't tempt me with another $100 just because I got curious about the benchmarks.
That pricing is wild. $2 per million input and $10 per million output is 5x cheaper than Astra on input and output and 10x cheaper on cached input. After the intro period it jumps to $4 and $20.
That's before they wire in a Jevons solution, which should cut agentic workflow costs by around 40% and bump speed by around 40% too.
Gemini is the model that's routinely borderline psychotic. It scares me. If we get paperclipped I won't be surprised if it's Gemini.
Gemini 3.5 dropped a DROP TABLE on a real production table during a system test. It had created that table in the test setup, so it apparently decided the table was a test table. Human review caught it.
For me it's the sneaky one. It guesses what a URL probably contains and answers confidently with made-up content, then fesses up when pressed. I'd rather it just say it can't open the source.
Honestly it's lazy. It answers as fast as possible even on Pro with extended effort. It's a Google Search replacement for me and not much else.
The whole thing is that it's not released yet to mere mortals.
The benchmarks are saturated. I'll wait for hands-on before believing Google is back. It would be nice to have more than OpenAI and Anthropic in the SOTA race.
Harvey's legal benchmark showing around 20% isn't saturated though.
The reasoning transparency part is good, but it's exactly why Google is the slow mover. They won't let the model enter an echo chamber and go faster than humanly possible.
OpenAI proposed and popularized chain-of-thought monitoring, so Google isn't being punished for anything. And Google doesn't return real chain of thought through the API, a small model writes a fake trace.
Argon agents migrating C/C++ to Rust across Google is the interesting bit. The cppnext team spent years on Carbon and Swift instead, and I wonder if Carbon is dead now.