Gemini 4 Carbon tops the leaderboard. You can't use it yet

You have a project to ship this month and a model to choose for it. The model at the top of the leaderboard this week is one you almost certainly cannot use. Gemini 4 Carbon, which Google announced on 28 October, is open only to a small group of vetted security teams, with no date for anyone else. Here is what it changes, and what it does not, for a choice you have to make today.

What Google released, and to whom

Carbon is Google DeepMind's new top model, arriving two weeks after Gemini 4 Argon reached paying subscribers. Google held back a general release because of its security skills. A model that is good at finding and patching holes in software is also good at breaking in.

So for now it goes to a subset of cyber defenders in Fairwind, Google's invite-only programme, and to Google's own teams. Per NeuralTrust, applications go through a Google Cloud account team, with background checks. Everyone else has been promised it "as soon as possible".

There is no public model ID, no model card and no limits published by Google. On 1 November the Gemini API models and pricing pages list only Argon from the Gemini 4 family.

Top of the board, by a single point

On the Artificial Analysis Intelligence Index, an independent score built from many tests, Carbon scores 59. Claude Opus 5.5 scores 58 at its maximum effort setting, GPT-6 Astra 57, and Argon 50. Claude Fable 5.1, Anthropic's most expensive model, peaks at 50, and Claude Sonnet 5.5 at 54.

One point is a thin lead. The second independent tester, Vals, puts Carbon second on its own index at 70.36%, behind Sonnet 5.5 at 70.85% and just ahead of Opus 5.5 at 70.18%, all within its margin of error of about one point. Carbon does lead two specialist boards outright on Vals: finance (66.35%) and legal research (55.77%).

The fair summary is a three-way tie at the top, with Carbon a nose in front on one scoreboard.

Coding and agents: not the leader

Google's launch table, which is self-reported and carries no error margins, claims the top score on 15 of 17 rows. The two it concedes are the ones developers care about most. On FrontierSWE 2, a test of real software engineering, Carbon scores 61.2% against Astra's 65.5% and Opus's 62.3%. On OSWorld, where the model operates a computer by itself, it scores 81.5% against Opus's 87.2%.

The rivals' own self-reported figures point the same way. Anthropic puts Opus 5.5 at 66.4% on Terminal-Bench 4.0, a test of agents working in a command line, against 57.9% for Astra and 55.8% for Fable 5.1. Google put Argon at 57.4% on the same test.

Where Carbon stands out on Google's table is reading very long inputs: 96.2% on GraphWalks, against 71.8% for Astra and 66.8% for Opus. If your work means asking questions of a whole codebase or a data room at once, that is the number to watch.

Fewer invented answers, but more than Argon

Hallucination, a model stating something false with confidence, is where Carbon's lead is widest. On Artificial Analysis's Omniscience test it makes things up 11% of the time, per VentureBeat, against 50% for both Opus 5.5 and Astra.

That cuts against part of Google's own story. Google says Carbon makes about half as many unsupported claims as Argon, yet on the same test Argon's rate was 4%. Carbon answers more questions correctly, 58% against 48%, but it guesses more often when it does not know.

Price, and the bill per task

Google has announced Carbon at the same rates as Argon: an introductory $2 per million input tokens and $10 per million output, then $4 and $20. Argon's introductory rate ends on 30 November; Google has not said when Carbon's would.

At $4/$20, both Gemini models match Opus 5.5 per token. Fable 5.1 and Astra cost $10 and $50, two and a half times as much.

Per-token prices mislead, because models differ in how much they write. Carbon used 200 million output tokens to run the Artificial Analysis index, against a median of 70 million. At the later rate that works out at $7.76 per index task, against $7.68 for Argon, $10.04 for Opus 5.5 and $6.25 for Astra. Carbon is cheaper per task than Opus, but not by as much as the token count alone suggests.

Speed: a two-minute wait

Artificial Analysis measures 118.58 seconds before Carbon's first word appears, against a median of 3.71 seconds. Argon takes 49.13 seconds.

The rivals are slow too at their top settings, which is where their headline scores come from. On the same leaderboard Opus 5.5 at max effort waits 729.62 seconds and Astra at max 270.03. At medium effort they drop to 16.41 and 5.07 seconds, with lower scores. Any of these models at full strength suits a job you hand off and come back to, not a chat window.

The catch: every Carbon figure above comes from testers given early access and from Google's own table. Nobody outside Fairwind can check them on their own work, Google has not said whether the final public model will match what was tested, and its retention and training terms are not published. A leaderboard win for a model you cannot call is a forecast, not a product.

Which one for which job

• Long coding and agent jobs: Claude Opus 5.5. It leads Terminal-Bench 4.0 and OSWorld, sits a point behind Carbon on the index, and is on sale through five platforms today.
• Most production work on a budget: Claude Sonnet 5.5 or GPT-6.1 Sol, both $2/$10. Sonnet tops the Vals index; Sol scores 55 on Artificial Analysis at max effort.
• Video, audio and stacks of PDFs: Gemini 4 Argon, which reads all of them natively in a 1M-token window. Budget for the price doubling on 1 December.
• Research-grade maths and science: GPT-6 Astra, which OpenAI reports at 64.6% on Terminal-Bench Science, against 52.6% for Fable 5.1.
• The hardest reasoning, after Opus has fallen short: Claude Fable 5.1. Anthropic's own docs now say to start with Opus.
• Vetted security teams: Gemini 4 Carbon through Fairwind, or GPT-5.6-Cyber through OpenAI's Daybreak Red programme.
• Lower prices: Kimi K3.1 Pro at $2/$10, with open weights promised by 4 November.

Our take: Carbon matters most as a signal of where Argon's successor in the Gemini API is heading, and as a reason Gemini teams can hold their budgets steady. For everyone else the decision is the one it was last week. Opus 5.5 is the frontier model you can actually buy, a point behind the leader, and switching later costs less than waiting for a date Google has not set.