Blog

Gemini 4 Argon for Coding: What I'm Testing Next to Grok 4.7

Google's new frontier model is cheap to try until November 30. Here's what it is, what it costs, and how I'm testing it next to my Cursor and Grok 4.7 setup.

Caleb Dorris 23 October 2026, 12:05 PM Updated 28 October 2026 7 min

Cover reading "Gemini 4 Argon" with an AI, What I am testing tag: a violet crystal and a clear crystal side by side, linked by streams of light.

Update, October 28, 2026

Gemini 4 Carbon is announced, but you can't use it yet

Two weeks after Argon, Google announced a preview of Gemini 4 Carbon, which it calls its strongest model, with state-of-the-art cybersecurity and long-horizon capabilities. Koray Kavukcuoglu, SVP at Google DeepMind and Google's Chief AI Architect, announced it on X too.

It isn't available to developers yet. Starting this week, Google is giving it to a subset of cyber defenders, then phasing in more trusted partners through what it calls its Fairwind Program. Developers, businesses and consumers come later, and Google hasn't given a date. When it does open up, Google says it will launch at Argon's introductory price, $2 per million input tokens and $10 per million output, then $4 and $20 after the intro period.

What Google claims. For coding, Google says its teams have seen Carbon run for days on multi-step engineering work, like root-causing and patching subtle bugs across large codebases. For security, it claims state-of-the-art results across detection, validation and remediation: 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation, and 71.0% on CWE-bench for patching. Those are Google's numbers, run without guardrails. The independent one I can check: Artificial Analysis lists Carbon first on its Intelligence Index at 59 out of 77 models.

What it changes for my testing: nothing yet. Argon's intro price still ends November 30, and Carbon isn't something I can put in front of my code. When it opens up, it gets the same tasks below on a low-stakes branch. The security review is the one I'll watch, since that's where Google says Carbon is strongest.

Google released Gemini 4 Argon on October 13, and Gemini 4 Argon for coding is the part I care about. My setup hasn't changed much: Cursor is my editor, Grok 4.7 does almost all of the work, and Gemini Pro is where I go for a second opinion on direction. Argon is the first Gemini model I'm seriously putting in front of my code.

There's a clock on it, too. The API is at an introductory price through November 30, and then it doubles. That's a good window to find out whether it earns a spot, so here's what I know and what I'm testing.

What is Gemini 4 Argon?

It's Google's newest frontier model. It's available in the Gemini API and AI Studio, in Google Antigravity for AI Ultra subscribers, and in the Gemini app for Pro and Ultra subscribers. The API model ID is gemini-4-argon.

The numbers that matter for code: a 1,048,576-token window shared between input and output, and three thinking levels (low, medium and high) in the API. Google's launch post leans on big engineering stories, like a Chrome refactor across hundreds of files and an overnight performance pass on its Agent Development Kit. Those are Google's results on Google's code. They tell me what it's aimed at, not how it'll handle mine.

How much does Gemini 4 Argon cost?

Through November 30, 2026, the Gemini API charges $2 per million input tokens and $10 per million output tokens, and output includes thinking tokens. Cached input is $0.10 per million. Starting December 1, those become $4, $20 and $0.20. It isn't available on the free tier.

Two details matter for coding. Caching is implicit only. The old explicit caching calls return an error, so code that creates cached content needs to change. And because thinking tokens bill as output, a long reasoning run on a hard task costs more than the prompt size suggests.

In Cursor, Argon is a third-party model, so it isn't part of a plan's included usage. On The X Subscription it's billed from prepaid credit, and legacy Cursor plans draw from the Other Models pool. Cursor lists the same $2 and $10 rates and says Argon scored 52.8% on CursorBench at default effort.

How does Gemini 4 Argon fit next to Cursor and Grok 4.7?

Grok 4.7 stays my default. It's fast, it lives in Cursor, and it handles the day-to-day work: describing a change, reading the diff, running the page. I'm not looking for a replacement. I'm looking for a second model I trust for specific jobs.

Argon's pitch lines up with three of them. The first is large refactors where logic is spread across includes, config files and a stored procedure nobody documented. Its window is twice the size of Grok 4.7's 500,000 tokens. The second is a second-opinion review of a diff before it ships. The third is planning a migration, where the hard part is reasoning about the whole app.

In Cursor, Argon has four effort levels: high, medium (the default), low and minimal. My plan is to start at medium and only go up when a task earns it.

How I'm testing Gemini 4 Argon for coding

I'm keeping it simple. I give the same task to Grok 4.7 and to Argon on a low-stakes branch, then compare the diffs myself. The tasks are ones I already know the right answer to:

  • A refactor on an inherited PHP app, moving repeated query code behind one function without changing behavior.
  • A security review of a diff, checking for missing CSRF tokens, permission checks and secrets in code.
  • A migration plan, like the ones in a Node-to-PHP conversion, where I want the risks listed before anyone writes code.

What I'm watching: whether it's right, whether it asks before doing something destructive, how much it changes that I didn't ask for, and what each task cost. Google says it trained Argon to pause for confirmation before irreversible steps. I'd like to see that hold up on a real repo.

If you're calling the API directly, read Google's migration notes before you swap the model ID. Sampling settings like temperature and top_p are ignored, you should set max_output_tokens explicitly, and Google recommends client timeouts of at least 20 minutes.

The caveats

A benchmark isn't my codebase. CursorBench and Google's launch stories are useful signals, but the only test that counts is my own repos.

Thinking costs money. High effort on every task will burn through credit fast.

The intro price ends November 30. Anything I learn after that is at double the rate, so the testing happens now.

None of this is a reason to skip it. It's a reason to test it on purpose instead of switching on a hunch.

What stays the same

Whichever model wins a task, the workflow doesn't change: agents do the typing and the first pass, and I decide what ships. New models get a trial on low-stakes work first, the same rule from my AI coding setup and the bigger picture in Cursor, Grok, and the other agents. I'll report back once I have results.

If you want an AI-assisted workflow like this behind your next project, get a free quote.

FAQ

What is Gemini 4 Argon good at?

Google positions it for complex, multi-step work like large codebase refactors, agentic coding and deep reasoning. It has a 1M-token window shared between input and output in the Gemini API.

How much does Gemini 4 Argon cost?

Through November 30, 2026, the Gemini API charges $2 per million input tokens and $10 per million output tokens, including thinking tokens, with cached input at $0.10. From December 1, it's $4, $20 and $0.20.

Is Gemini 4 Argon available in Cursor?

Yes. Cursor added it on October 13, 2026 as a third-party model with high, medium, low and minimal effort levels. It's billed separately from a plan's included usage.

Is Gemini 4 Argon free?

No. It isn't on the Gemini API free tier. Consumers can use it in the Gemini app with a Google AI Pro or Ultra subscription.

Can I use Gemini 4 Carbon yet?

Not as a developer. As of October 28, 2026, Google is rolling it out to a subset of cyber defenders and trusted partners through its Fairwind Program first. It says broader access is coming soon but hasn't given a date.

Sources

All posts