Gemini 4 Carbon Explained: What Google Is Testing

Last updated: October 11, 2026

Google is testing another version of Gemini 4 behind closed doors — and employee reactions are doing the marketing. Business Insider reports that an internal build codenamed Carbon has been deployed on the company’s private coding platform in recent days, and one engineer’s verdict — that its coding “feels like Opus 5.5” — has turned a routine internal checkpoint into the most talked-about unreleased model of the week. Here is what is actually known, what is one person’s claim, and what is still blank.

Quick answer: Google is testing an internal Gemini 4 checkpoint called Carbon. One engineer's 'feels like Opus 5.5' coding claim is unverified — no benchmark exists yet.

What is Gemini 4 Carbon?

Carbon is an internal checkpoint of Google’s Gemini 4 model family. According to Business Insider’s reporting on 9 October 2026 (relayed by MachineBrief, Storyboard18, AI Weekly and others on 10 October), the build — codenamed Carbon — was made available “in recent days” on Jetski, Google’s internal coding platform associated with Antigravity, its AI coding product.

The reporting rests on internal documents and screenshots seen by Business Insider. Two employee reactions leaked with it: one developer told the outlet that Carbon’s coding performance “feels like Opus 5.5” — referring to Anthropic’s flagship model for long-running agentic coding — while adding the caveat that they would need to test it more. Another wrote in an internal channel: “Carbon is really good!”

Google declined to comment, so everything below carries that uncertainty with it.

What is confirmed, what is reported, and what is still unknown

The story is a useful test of how to read unreleased-model news: three layers, kept separate.

Layer	Item	What backs it
Confirmed	Google announced Gemini 4 (public name: Argon) on 30 September 2026	Trade reporting on the launch and rollout
Confirmed	Google is internally testing a checkpoint line under the codenames Argon, Barium and Carbon	Internal documents seen by Business Insider
Confirmed	Google declined to comment on Carbon	BI’s reporting
Reported	Carbon was deployed on Jetski in recent days	BI, via documents and screenshots
Reported	One employee says Carbon “feels like Opus 5.5” for coding — but caveated it needs more testing	A single anonymous source speaking to BI; not independently confirmed
Reported	A second employee wrote “Carbon is really good!”	Internal channel, via BI
Reported	The Barium-B build was selected as the public Argon release	Internal document, via BI
Unknown	Whether Carbon ever ships publicly — as an Argon update or a separate model	BI: “could not be determined”; it may never be released
Unknown	Any public benchmark for Carbon	None exists

How Carbon fits into Google’s Argon–Barium–Carbon line

Google’s internal codenames are deliberately out of sync with public names. The internal build “Barium-B” was selected to become the model the public will know as Argon; Carbon is the next step along that same line. What nobody — including the reporting outlets — can determine yet is whether Carbon becomes another update under the Argon name or a separate model in the Gemini 4 family.

Context for the Argon side of the story: Google announced Gemini 4 Argon on 30 September 2026, giving first access to selected cybersecurity partners in its Fairwind Program, with paid API customers and Google AI Ultra subscribers expected next — no date given. Our sister site’s Argon explainer walks through its 1M-token output window and pricing in full; Carbon is not covered there.

Signals that Argon itself is approaching: TestingCatalog reported on 10 October that Antigravity’s model selector now hints at Argon with three context configurations — 256K by default, 512K at roughly 1.3x quota per turn, and 900K at around 1.8x quota — and that Gemini Web has added low, medium and high thinking-effort settings, possibly in preparation for the Gemini 4 generation.

Why Opus 5.5 is the comparison everyone reaches for

The Opus 5.5 comparison is the whole story — and the whole caveat. Business Insider describes Claude Opus 5.5 as Anthropic’s most powerful model for long-term agentic coding; it is the bar Google is chasing. The contrast is telling: earlier internal versions of Argon were generally well received by staff, but one employee compared them on some coding tasks to Claude Opus 5 — the older model. Carbon “feels like Opus 5.5” is shorthand for “Google may have closed a generation’s worth of gap.”

Keep the weight honest, though: that is one anonymous staffer, speaking to one reporter, about an internal build nobody outside the company has evaluated. As AI Weekly noted on 10 October, no second outlet has confirmed the comparison, and no public benchmark exists. A subjective “feel” quote is not a benchmark — and internal builds sometimes feel great on an internal coding platform and ordinary everywhere else. For context on why Opus 5.5 sits at the top of this particular ladder, see our Claude Opus 5.5 vs Sonnet 5.5 guide.

Waiting on Argon? What Carbon does and doesn’t change

Carbon changes nothing about what you can use today — but it changes how to read Google’s pace. Our read:

Carbon is not something you can wait for. It is an internal checkpoint with no release date, no public benchmark and no confirmed public name. Internal testing often ends in builds that never ship.
The Argon rollout path is unchanged: Fairwind partners first, then paid API customers and AI Ultra subscribers, with Google silent on dates.
If you are choosing a coding model this week, choose from what exists — not from a checkpoint quote. Single-employee impressions are the weakest signal in the model-news food chain.
The real signal is velocity: three internal codenames tested in roughly two weeks suggests Google is iterating fast — and Argon is unlikely to sit still once it reaches the public.

Frequently asked questions

Is Gemini 4 Carbon publicly available?

No. It exists only as an internal Google build; there is no release date, no public benchmark, and Google declined to comment.

Will Carbon become an Argon update or a separate model?

Unknown. Business Insider reported it could not be determined — Carbon could ship under the Argon name, become a separate Gemini 4 model, or never ship at all.

Who said Carbon feels like Opus 5.5?

One anonymous Google developer, speaking to Business Insider, with the explicit caveat that they would need to test it more. No other outlet has independently confirmed the comparison.

What is Jetski?

Google’s internal coding platform, associated with Antigravity — the environment where Carbon was deployed for staff testing.

How is Carbon different from Gemini 4 Argon?

Argon is the public model name announced on 30 September 2026 and rolling out to partners; Carbon is an internal checkpoint codename that may or may not ever become a public product.

Sources and methodology

Claims in this article rest on Business Insider’s 9 October 2026 original reporting as relayed by four secondary outlets on 10 October — MachineBrief, TestingCatalog, Storyboard18 and AI Weekly — since Business Insider’s article sits behind its own site. Single-source employee claims are labelled unverified throughout; there is no public benchmark for Carbon, and nothing here is hands-on tested since no public access exists. All claims re-verified against the cited pages on 11 October 2026.

MachineBrief’s re-report of the Business Insider story — 10 October 2026
TestingCatalog — Antigravity Argon selector hints and thinking-effort settings, 10 October 2026
Storyboard18 — Carbon summary, 10 October 2026
AI Weekly — analysis noting the comparison is unconfirmed by a second outlet, 10 October 2026
Gemini 4 Argon: Google’s New Frontier AI Model Explained — sibling coverage of the public model (Carbon not covered there)

Internal reading: Claude Opus 5.5 vs Sonnet 5.5: Which to Use · Gemini Skills Guide: How to Migrate from Gems · Gemini Free Limits Guide: What Changes Oct 9