Gemini 4 Carbon: What the Leaks Got Right, and Who Gets It First

Google just announced its top-scoring model, and you probably can't use it yet. Gemini 4 Carbon leaked for weeks before the launch, so I went back through every leak and checked it against what Google actually confirmed. Then I lay out how I'd get ready for the day it opens up.

The leak trail

On October 9, Business Insider reported that Google staff were testing a new Gemini 4 model, code named Carbon. REPORTED · BUSINESS INSIDER It ran inside Jetski, Google's internal coding tool, and Business Insider says it saw documents and screenshots.

One employee said Carbon feels like Opus 5.5 for coding, then added they would need to test it more. Google declined to comment.

The documents named three models: Argon, Barium and Carbon. An internal build called Barium-B was chosen to become the public Argon. So yes, Google is naming its models like a chemistry class. At the time, Business Insider couldn't tell whether Carbon was a new model or an update to Argon.

Then, on October 21, testers on X said Carbon had shown up on LMArena, the blind model comparison site. RUMOR · TESTER POSTS ON X One account said partners had early access and that "the release is very close."

Their read on the model: better than Argon at 3D, frontend still weak, and slower. One tester guessed two to four times slower, depending on the task.

Keep the source in mind here. As TestingCatalog pointed out, these were arena matchups and hand-picked prompts, and nobody outside Google could confirm which version was being served.

What Google confirmed

On October 28, Google made it official. Google calls Gemini 4 Carbon "our strongest model with state-of-the-art cybersecurity and long horizon capabilities," which in plain terms means security work and long, multi-step tasks. CONFIRMED · GOOGLE It builds on Argon, which launched two weeks earlier.

Carbon's number one spot comes from an independent scorer. On the Artificial Analysis Intelligence Index v5.0, Carbon scores 59, the highest on the board. INDEPENDENT · ARTIFICIAL ANALYSIS

Google also reports big security scores, measured with guardrails off. Google says Carbon hits 88.2% on SEC-bench Pro, a benchmark for finding real vulnerabilities. Google also says its own teams have watched Carbon keep working on complex engineering tasks for days, pushing through dead ends.

Now the catch. Most of us can't use it yet. Here's the order Google gave:

1. A subset of cyber defenders, starting this week.
2. Then more trusted partners, phased in through Google's Fairwind program.
3. Then developers, businesses and consumers, "as soon as possible."

There's still no date for step three. CONFIRMED · GOOGLE

Google did publish the price. Carbon launches at an introductory $2 per million input tokens and $10 per million output tokens, the same as Argon's intro price, with cached input at 95% off. After the introductory period, it doubles to $4 and $20. Google hasn't said when that intro period ends. UNKNOWN

Leaks versus launch

So which leaks held up? Here's my scorecard.

The leak	Who said it	What the launch shows	Verdict
Carbon is real, and Googlers were coding with it	Business Insider	Announced; Google says its teams used it internally	Held up
Maybe it's only an Argon checkpoint	Business Insider	Ships under its own name, Gemini 4 Carbon	Answered
"The release is very close"	Tester post on X, Oct 21	Announced Oct 28, about a week later	Held up
Coding "feels like Opus 5.5"	One employee, via Business Insider	59 vs 58 for Opus 5.5 on one independent index	Roughly
Better than Argon at 3D	Tester posts on X	No published 3D number	Unverified
Frontend still weak	Tester posts on X	Google says +89 points on WebDev Arena, to #4	Mixed
Two to four times slower	One tester on X	No official or independent speed data	Unknown

A few of these deserve more detail.

The Opus comparison. On the Artificial Analysis index, Carbon beats Claude Opus 5.5 by one point, 59 to 58. INDEPENDENT · ARTIFICIAL ANALYSIS One point, on one index. Google's own benchmark chart has Carbon on top in 15 of 17 rows. On one coding test, FrontierSWE 2, Opus and OpenAI's Astra both score higher: Google's chart shows Carbon at 61.2%, Opus at 62.3% and Astra at 65.5%.

3D. Google published no 3D number, and as of October 29 Carbon wasn't on the Artificial Analysis 3D leaderboard. That one stays with the testers for now.

Frontend. Google says Carbon jumped 89 points on WebDev Arena to fourth place, at 1768. Better, and still fourth.

Speed. Google hasn't published speed numbers, and with no public API, independent testers can't measure it yet. UNKNOWN

Argon, the baseline

While we wait, Argon is the model you can actually use. It's in the Gemini API today as gemini-4-argon, with a one million token window shared between input and output. CONFIRMED · GOOGLE

It has the same introductory price as Carbon, $2 and $10. For Argon, Google set an end date: the intro price runs through November 30, and on December 1 it goes to $4 and $20.

People are using it. Google's Logan Kilpatrick said Google processed about 25 trillion tokens in Argon's first 72 hours.

Here's the part I'd watch. Argon took the same path Carbon is on now. Google announced Argon on September 30, for cyber defenders first through Fairwind. Thirteen days later, on October 13, developers had it.

That's one precedent. Carbon could take longer. It does show how Google ships right now.

What changes for builders

If Google's numbers hold, I see three things changing. This is my opinion.

One: long-running agents. A model that keeps going through dead ends lets you hand it bigger chunks of work, with fewer check-ins and better logs.

Two: security work. Find, check and patch, all in one model. Remember, Google measured those security scores with guardrails off, so the version you get may behave differently.

Three: price. Carbon launches at the same per-token price as Argon. That makes the real question cost per finished task, and you'll only know that by testing it.

One caution from the independent side. Artificial Analysis found Carbon gets more questions right than Argon on its AA-Omniscience test, while its hallucination rate went up, from 4% to 11%. INDEPENDENT · ARTIFICIAL ANALYSIS Artificial Analysis defines that rate as wrong answers divided by wrong, partial and unanswered ones.

In its own internal testing, Google says Carbon makes about half as many unsupported claims as Argon. Those are different tests. Keep your own checks in the loop.

Ready for launch day

Here's how I'd get ready for the day Carbon opens up. This part is my opinion.

1. Build an eval harness now. Take 20 real tasks from your own work, or a client's. Run them on Argon today. Save the outputs and the cost of each finished task. On launch day, rerun the same 20 on Carbon, and you'll have real before and after numbers within a day.
2. Offer that as a service. Plenty of teams are on Argon right now. A model swap audit covers their tasks, both models, cost per finished task, and a clear recommendation.
3. Watch the price deadline. Argon's intro price ends November 30. Teams will want to know what their bill looks like in December, and that's a quick, concrete job.
4. Apply to Fairwind if it fits. If your organization does security work and qualifies, apply to the Fairwind program. That's where access starts.
5. Publish honest tests. If you make content, share launch day tests with real tasks and real costs, clearly labeled.

The verdict

Confirmed: Carbon is real, Google calls it its strongest model, and it's first on one independent index.

Reported, and it held up: Google staff were testing it internally. One tester rumor held up too: the release was close.

Still rumor: better 3D, and slower.

Unknown: the API date, the model ID, and when Carbon's intro price ends. UNKNOWN

My take: build your tests on Argon this week, so Carbon is a rerun the day it opens.