Google's Gemini 4 Carbon Tops an AI Index, but Cyber Defenders Get It First

Fifty nine. That is the score Google says Gemini 4 Carbon posted on version 5.0 of the Artificial Analysis Intelligence Index, enough to rank first on that leaderboard. Google announced the model on October 28, just two weeks after it shipped Gemini 4 Argon, and it describes Carbon as its strongest model yet, with a heavy focus on cybersecurity and long, multistep work.

There is a catch for most readers: you cannot use it yet. According to Google's announcement, Carbon goes first to a small group of cyber defenders, with access widening later through the company's Fairwind Program. Developers, businesses and regular Gemini users have to wait.

Here are the figures Google and early coverage have put on the table, followed by what they actually tell you.

Gemini 4 Carbon in eight figures

• 59: the Artificial Analysis Intelligence Index v5.0 score Google cites, ranked number one.
• 2 weeks: the gap between the Argon launch and the Carbon announcement, as both Google and 9to5Google note.
• 88.2%: Carbon's result on SEC-bench Pro, a test of finding vulnerabilities in real software.
• 54.8%: its score on ExploitGym, which measures how well a model can exploit flaws. Google calls it a new state of the art.
• 71.0%: its result on CWE-bench for patching security bugs.
• +89 Elo points: the jump on WebDev Arena, which moves Carbon to fourth place on that web development leaderboard.
• $2 and $10: the introductory price per million input and output tokens, the same as Argon.
• $4 and $20: the price per million input and output tokens Google lists for after the introductory period ends.

Three security scores that explain the slow rollout

The cyber numbers are the reason Carbon is not on your phone today. A model that is good at finding bugs (SEC-bench Pro), good at writing patches (CWE-bench) and also good at exploiting holes (ExploitGym) is useful to defenders and attackers alike. Google's answer is to hand it to defenders first.

The blog post says Google will make Carbon available to "a subset of cyber defenders" this week and phase in more trusted partners over time. It also says the company is taking part in the US government's voluntary process for testing models before release. That fits a pattern across the industry. Researchers have been warning about models that can probe networks on their own, a risk we covered in our report on Qwen3.8 trying to upload data to Alibaba Cloud, and outside auditors keep finding gaps in how labs watch their own systems, as in the METR review of OpenAI monitoring.

One caution about these scores. Google says its external cybersecurity benchmarks were run without guardrails, and every number in this article comes from Google itself. Independent testers have not yet published their own results.

What 0.0% means on the safety charts

Two more numbers in Google's post are zeros, and they are the ones the company seems proudest of. On Terminal-Bench 4.0, Google says successful cheating by Carbon dropped to 0.0%, with cheating attempts at 0.4%. On the Gray Swan indirect prompt injection test, Google reports a 0.0% attack success rate, down from 0.7% for Argon.

Prompt injection is the trick where hidden text on a web page or in a file tells an AI agent to do something its user never asked for. For a model meant to run long tasks with little supervision, that matters more than a few points on a quiz. Google also claims Carbon produces secure code more than twice as often as Argon and makes about half as many unsupported claims.

The price tag and the two week release cycle

The introductory price of $2 per million input tokens and $10 per million output tokens is identical to Argon's, Constellation Research reports. Google has not said how long that period lasts before the $4 and $20 rates apply.

The speed of the release says as much as the price. Larry Dignan of Constellation wrote that the takeaway is Google "stepping up its frontier model release pace." Google shipped Argon, then a stronger successor, inside half a month. Add the company's push into new model types such as Genie 4, its world model with audio, and its effort to sell its own chips to rivals, covered in our piece on OpenAI's plan to buy Google TPUs, and Google looks set on pressing every advantage it has right now.

Where Gemini 4 Carbon leaves regular Gemini users

For now, nothing changes in the Gemini app. Google says it is still gathering feedback and refining guardrails before bringing Carbon to developers, businesses and consumers "as soon as possible," without naming a date. If you use the free tier, keep in mind that Google has already been steering free users toward lighter models, as we explained when Gemini free users were moved to Flash Lite. There is no sign yet of which plans will get Carbon first.

When it does arrive, the number to check will not be 59. It will be how Carbon does on independent tests once people outside Google can try it. Until then, you can follow more model launches on our AI news page.