Gemini 4 Argon Is Here: What Google’s New Frontier Model Actually Changes for You
Google released Gemini 4 Argon on October 13, 2026, and I spent this morning reading the launch post from Google DeepMind’s Koray Kavukcuoglu line by line. It is a big post, full of demos, benchmark claims and science stories. I wanted to pull out the parts that matter if you write code, run a team, or just pay for a Gemini subscription.
So this is my plain-English read. Every number below comes from Google’s own announcement, so treat them as Google’s claims until independent testing catches up.
What Google actually launched
Argon is the first model of the Gemini 4 family. Google says “the rest of the Gemini 4 family” is still to come, so this is the opening act, not the whole lineup.
It is available globally from day one, which is unusual for a frontier model. Here is where you can find it:
- Developers: the Gemini API paid tier in Google AI Studio and Android Studio. Google’s developer guide lists the model ID as
gemini-4-argon. - Enterprises: Gemini for business, the API, and Google Antigravity.
- Everyday users: the Gemini app and AI Mode in Search, for Google AI Ultra and Pro subscribers.
- Google AI Ultra subscribers can also try it in Antigravity and Google AI Studio.
Google pitches Argon at “deep work”: long, multi-step tasks in coding, finance, legal, cybersecurity and science. The headline benchmark claim is a number one ranking on the Vals Index, which Google describes as a measure of performance across sectors like finance, coding, legal and tax work.
The price, and the date to circle
Google states the price in two places, the launch post and the Gemini API pricing page, and the numbers match. Argon starts at an introductory $2 per million input tokens and $10 per million output tokens. Cached input gets 95% off, which the pricing page lists as $0.10 per million tokens.
The catch is the end date. The pricing page says the introductory rates run through November 30, 2026, and from December 1 they move to $4 per million input tokens and $20 per million output tokens (cached input goes to $0.20). That doubles every rate.
If you are testing Argon for a product, I would budget at the later price from the start. It is easy to fall in love with a model during a cheap trial month and then get a surprise on the December bill.
One more detail from the pricing page: Argon is not on the API’s Free tier, and the docs say it is not available on the Google AI Pro plan either. Yet the launch post says Pro subscribers get it in the Gemini app, so check which product you are paying for before you plan around it.
The other spec that caught my eye is what Google calls an “industry-leading 1M unified input-output token limit” in the Gemini API, plus a new continuation protocol for long jobs. The developer guide puts a number on it: input, thinking and output share one window of 1,048,576 tokens, and long answers are written in hops of up to 64K tokens that pick up from each other. In practice that means one budget shared by what you send, what the model thinks and what it writes back. Long agent runs will eat into it from both sides, so read the developer docs before you wire it into anything that runs overnight.
The coding claims, read carefully
Most of the launch post is about engineering work, and the numbers are specific enough to be interesting.
Inside Google
Google says Argon in Antigravity helped its teams raise agentic code submissions by more than 35% while cutting code rollbacks by nearly a third. That second number is the one I care about. More code is easy. Less broken code is the hard part.
The named examples
The post lists a few projects with real figures:
- Chrome: Argon helped separate core webpage handling from a roughly 6 million line legacy folder inside Chromium’s 40 million line codebase, finishing a refactor across hundreds of files in weeks instead of months.
- Agent Development Kit: given the team’s
.agents/skillsrulebook and a one-paragraph goal, Argon worked overnight and found changes that cut in-memory session prep time by 40% in long sessions and sped up cold imports by 21%. - Gemma inference: a 2.85x end-to-end speedup on an A100 via vLLM, from about 37,900 to over 108,000 tokens per second on Gemma 4 E2B, with 100% token parity.
- xTB quantum chemistry package: up to 2.3x faster at predicting molecular structures and 1.8x faster at conformer search, with the changes sent upstream.
I like that several of these point at public code you can inspect. That is a better habit than a chart with no source. Still, these are Google’s own showcase projects, picked by Google. Your messy internal repo is a different test.
If you have been following the agent tools I covered in Codex Cloud after DevDay, Argon in Antigravity is Google’s answer to the same “hand it a goal and let it run” style of work.
Safety and security: the parts worth reading
I spend a lot of time around security tooling, since my day job sits close to network detection and response. So I read this section twice.
Shortcuts and irreversible steps
Google says it trained Argon to respect system boundaries, report negative findings honestly, and pause for human confirmation before irreversible steps. On coding benchmarks, Google reports shortcut attempts (things like dodging a failing test) dropped to 2%. On top of that, an automated misalignment monitor reads the agent’s reasoning and actions to catch workarounds in real time.
Prompt injection
When an agent reads a webpage or an email, hidden instructions can try to hijack it. Google says Argon leads the Gray Swan Indirect Prompt Injection benchmark, with attack success across 15 adaptive attempts cut to 0.7%. That is low, but it is not zero. If your agent can read untrusted text and also take actions, you still need permission limits around it.
The Fairwind Program
Cyber defenders can apply to the Fairwind Program to get Argon without its cyber guardrails, on a rolling basis. Google says it is expanding the program to hundreds of additional partners starting this week. Wiz is named as an early user through its free “Scan for Good” work on critical infrastructure.
If you care about watching what these agents actually do in production, my LLM observability guide covers the tracing side.
Science, robotics and the free research offer
Google packs a lot of science into the post: help with quantum error correction, new counterexamples to Grad’s conjecture in fusion research with Professor Matt Landreman, rare-disease genome work with Gamow Labs, and enzyme design. These are hard to judge from a blog post, and I would wait for papers before repeating them as settled.
On robotics, Google reports an 87% score on the Embodied Reasoning (ERQA) benchmark and describes Argon watching a video of a person playing cornhole, building simulation code, and then running the skill on Apptronik’s Apollo humanoid, which sank seven shots in a row in a real-world trial.
The most practical item here is an offer. Google is giving 10,000 scientists the highest usage tier in Antigravity free for six months, subject to eligibility, capacity and selection. Applications are open now. If you work in research, it costs nothing to apply.
Google also launched a tool for checking long documents that week, and I wrote about Google Deep Review separately.
For an open weight comparison, my post on Meta Muse Spark Open looks at what builders can run themselves.
How I would test Argon this week
Launch posts are written to impress. Here is how I would make my own call instead.
- Pick three real tasks you already know the right answer to: one bug fix, one long document summary, one multi-step job.
- Run them on Argon and on the model you use today, with the same prompt and the same tools.
- Count what matters: wrong answers, broken builds, and how often you had to step in. Speed is nice, but rollbacks are what cost you a weekend.
- Price it at $4 and $20, the post-November rates, not the introductory ones.
- Keep a human gate on anything that deploys, deletes or pays. Google’s own safety notes assume one.
My next step is simple. Before November 30, I am running those three tasks on Argon and on my current setup, writing down the results, and only then deciding whether it earns a place in my workflow.
Leave a comment