“Google just shipped a model I can’t use. Cool. Great. Let me read all of it anyway.”
I was skimming Google’s announcement when one line made me stop scrolling. It wasn’t a benchmark. It was the output token limit.
I read it twice. Then I opened the old docs to check the previous number.
That’s how Gemini 4 Argon got my attention.
Gemini 4 Argon’s 1M Output Token Limit, Explained 🧠
“Output, not context. Did they really mean output?”
Yes. Output.
Google raised Argon’s output limit from 64K tokens to 1 million. Most of us talk about context windows, how much a model can read. This is about how much it can write and think in one go.
The so-what is simple. Hard problems need room. A model that can reason through hundreds of thousands of tokens in a single run doesn’t have to chop a tough task into pieces and hope the pieces line up. Google says this is what lets Argon solve difficult problems “in one go.”
For anyone building agents, that’s a bigger deal than another point on a leaderboard.
Gemini 4 Argon Benchmarks: DeepSWE, Vals Index and AutomationBench 📊
“Numbers are cheap. Let me see which ones actually mean something.”
Here’s what Google published:
- DeepSWE v1.1: 77.9%, which Google calls a new state of the art for long-horizon, real-world software engineering
- Vals Index: leads, covering finance, coding, legal and tax work
- AutomationBench (Zapier): #1 at 51.3%
- LVBench (long video understanding): 91.7%
- CWE-bench v1 (fixing security vulnerabilities): 68%, tied for first
Notice the pattern. These aren’t “write a to-do app” tests. They measure long, messy, multi-step work.
One caveat. These are Google’s reported numbers, and the model isn’t publicly available. Nobody outside the trusted-tester circle can check them yet. I’d treat them as strong claims, not settled facts.
Rust Migrations and 300 TiB Saved: What Argon Already Does Inside Google 🦀
“Okay, this part is either very impressive or very carefully curated.”
Benchmarks are one thing. Google also shared what Argon is doing on its own systems.
A team of Argon agents analyzed fleet-wide profiling data and applied memory optimizations across Google’s data centers. That’s over 300 TiB of memory freed, with total savings estimated at 500 TiB to 1 PiB.
In quantum research, Argon beat a published baseline for optimizing subroutine resources by 40%. In minutes.
The one I keep thinking about is the C/C++ to Rust migration. Agents are working on codebases from tens of thousands of lines up to 800K+ lines for the Fuchsia Zircon kernel. For libgav1, Google’s open source video decoder, Argon replaced 32K lines of SIMD code with safe Rust the compiler can vectorize itself. The result runs 2.7x faster than the existing Rust port, with identical video output.
Google also says these rewrites are going through automated and manual auditing, emulation testing and review before production. Good. A kernel rewrite shouldn’t ship on vibes.
I’ve shipped 7+ production apps on the same architecture. I know how large migrations go. They don’t fail on the first 80%. They fail on the last 20%. If Argon holds up there, that’s the real story.
Gemini 4 Argon for Cybersecurity: What the Fairwind Program Is 🛡️
“This is the part nobody gets to touch. Yet.”
Argon is launching first to trusted cyber defenders through Google’s Fairwind Program. Not developers. Not consumers. Defenders.
Google says Argon can autonomously find, validate and patch critical software vulnerabilities. For trusted defenders and Google’s own teams, it ships without cyber guardrails, so they get the full capability.
Wiz is already using it through its Scan for Good initiative, which finds and fixes high-risk exposures in critical public infrastructure for free. In an early result, Argon found a critical vulnerability exposing sensitive personal data in healthcare software used by hospitals worldwide. Google says earlier frontier models missed it.
That’s a strong sign. It’s also a reminder of why the rollout is gated.
Gemini 4 Argon Pricing in INR (and the Hike Coming After Launch) 💸
“Cheap now. Less cheap later. Classic.”
Introductory pricing, with rough INR conversions at about ₹96.5 per dollar:
USD per 1M tokens Approx. INR Input $2 ~₹190 Output $10 ~₹960 Cached input 95% off input price ~₹10
After the introductory period, it goes to $4 input and $20 output per million tokens. That’s roughly ₹385 and ₹1,930.
Output is the expensive side. And Argon’s whole pitch is long outputs. Do the math before you build anything that thinks for 500K tokens a run.
Gemini 4 Argon Safety Safeguards: What Google Says It’s Doing 🔒
“Every lab says ‘safety.’ I want to know what’s actually in place.”
Google lists four areas:
- Misuse defense: refusing harmful cyber and CBRN requests, plus monitoring the model’s internal activations to spot misuse. Internal and external red teams tested these safeguards.
- Prompt injection: Google calls Argon its most resilient model yet and says it leads Gray Swan’s indirect prompt injection benchmark.
- Misalignment monitoring: the chain-of-thought and actions are monitored, and execution stops when needed.
- Hardened sandboxes: isolated and sealed before high-risk training or evaluations begin.
The part I liked most is Google’s note that it took care not to feed monitoring findings back into training. The aim is to avoid teaching the model to hide its reasoning from the monitors. It also pushes the rest of the industry to keep reasoning transparent.
Safeguards described in a blog post are still a description. We’ll learn more once outside people can poke at it.
Should You Wait for Gemini 4 Argon? My Take as a Flutter Developer 🎯
“Hype is free. Access is not.”
Here’s where I land.
Impressive: the 1M output limit, the long-horizon coding results, the real migration work.
Missing: public access. Google says it’s still in the U.S. government’s voluntary pre-release access process and gathering tester feedback. When it opens up, it starts with paid API customers and Google AI Ultra subscribers. No date was given.
What I’m watching: how it handles real codebases in private corporate repos, which is where I work. Benchmarks don’t tell me that. Only trying it does.
Until then, I’m not rewriting my stack around a model I can’t run. I’ll wait, test it, and report back.
Tags: Gemini, Google AI, AI Coding, Flutter, Cybersecurity