Google’s Gemini 4 Carbon came two weeks after Argon and beats it, but you can’t use it yet

Google went more than seven months without a new top-tier AI model, the most capable kind a company has and the one it puts up against its rivals’ best. Then it announced two in four weeks.

The first was Gemini 4 Argon. Gemini is Google’s AI assistant, and Argon runs on its top two paid plans. Google announced Argon on 30th September 2026 and opened it to paying customers on 13th October. The second is Gemini 4 Carbon, announced on 28th October, 15 days after that release. Google calls it its strongest model yet.

So if you’ve only just started paying for Argon, you’ll want to know whether Carbon is better and when you get it. It is better than Argon, but it doesn’t lead its rivals as widely as Google’s own table suggests. And you can’t have it yet, because Google is only releasing it to some cyber security teams for now.

Why it came so soon

Google’s explanation is one line: its “researchers and models are working hand in hand to advance capabilities faster”. We think that means it’s using its AI to help build its next AI, but it doesn’t say how much of Carbon came from that. It also hasn’t said when work on Carbon started, so we can’t tell how long it took to build.

What Carbon does better

Carbon is built for long jobs. An AI agent is a model that works through a multi-step task on its own, and the longer the task, the more chances it has to get stuck. Google says, without giving data, that Carbon has run for days inside the company and pushed through dead ends to find and fix subtle bugs in large codebases.

Its other strength is cyber security. That means finding flaws in software, working out which ones an attacker could use, and writing the fix. On the tests for those three jobs, Carbon scores 88.2%, 54.8% and 71%, ahead of every model Google lists a score for. But Carbon took those tests without guardrails, the filters that stop a model helping with attacks, so the scores describe a version the public wouldn’t get.

Is it the best AI model?

Carbon beats Argon on all 17 tests in Google’s comparison table and has the top score on 15 of them. But that table only includes three rival models, and Claude Sonnet 5.5, from rival AI company Anthropic, isn’t one of them. Google’s table puts Carbon first on the Vals Index, a test of finance, coding, legal and tax work. On Vals’s own leaderboard, which includes Sonnet 5.5, Sonnet was first as of 4th November.

Artificial Analysis, a firm that runs AI tests, also has Carbon in a tight race. Its Intelligence Index combines nine of them into one score.

On these three tests run by other firms, Carbon leads one, by a point.

Google also says Carbon makes about half as many unsupported claims as Argon, on its own internal testing. Artificial Analysis’s knowledge test measures something else: when a model doesn’t get an answer right, how often it gives a wrong answer when it could have said it didn’t know. On its high effort setting, Carbon does that 11% of the time, up from 4% for Argon. Carbon also gets more answers right than Argon, 58% to 48%. Claude Opus 5.5 on its top setting gets more right still, about 66%, but gives a wrong answer 50% of the time. So on its high setting, Carbon knows more than Argon, but gives more wrong answers when it doesn’t know.

When you can use it

Carbon is going first to “a subset of cyber defenders”, then to more partners in Google’s Fairwind Program, its group for security teams at governments, critical infrastructure and technology companies. The skill that helps a defender find a hole also helps an attacker find one, which we think is why it’s starting small. Google says it’s still working on Carbon’s guardrails, and it’s taking part in the US government’s voluntary testing of models before release.

Everyone else comes after that, “as soon as possible”. There’s no date. Argon took 13 days to go from announcement to paying customers, but Google hasn’t said Carbon will do the same. It also hasn’t said which Gemini plans will get Carbon, or whether Carbon replaces Argon.

Developers, who build these models into their own apps, at least know the price. Carbon will launch at Argon’s introductory rate of USD 2, or about KES 260, per million tokens you send it, and USD 10, or about KES 1,300, per million tokens it writes back. A token is a chunk of text, about three-quarters of an English word. Those prices later double, to USD 4 (about KES 520) and USD 20 (about KES 2,600). Argon’s introductory rate ends on 30th November, but Google hasn’t said when Carbon’s does. We converted at KES 129.87 to the US dollar, the rate on 5th November 2026.

What this means in Kenya

Nothing changes here yet. Google’s Kenyan subscriptions page lists Argon on AI Pro at KES 3,700 a month and on AI Ultra from KES 12,500, with no mention of Carbon. Don’t buy AI Pro expecting Carbon, because Google hasn’t said Pro will get it.