Gemini 4 Carbon Is Google’s New No. 1 AI Model. You Just Can’t Use It Yet

Google has a new best model, and almost nobody can touch it. Gemini 4 Carbon, announced on October 28, now sits at the top of Artificial Analysis’ Intelligence Index with a score of 59, edging out Anthropic’s Claude Opus 5.5 (58) and OpenAI’s GPT-6 Astra (57). It also jumped 89 Elo points on WebDev Arena. If you pay for Google AI Pro or Ultra and opened the Gemini app this morning looking for it, though, you came up empty.

That’s by design. Google is handing Carbon to a vetted group of security professionals first, and only later to developers, businesses and regular Gemini users. It’s an unusual launch for a consumer company, and it says a lot about what Google thinks this model can do.

What Google actually announced

Carbon arrives barely two weeks after Gemini 4 Argon opened up to paying subscribers. Koray Kavukcuoglu, Google DeepMind’s SVP and Google’s chief AI architect, wrote the announcement and pitched Carbon as a model shaped by feedback on Argon, with targeted gains in coding, agentic tasks, scientific work, full-stack web development and multimodal understanding.

The headline claims are about security. Google says Carbon can find vulnerabilities in real-world software, test whether they can be exploited and then fix the root cause. Its self-reported scores:

• SEC-bench Pro (finding vulnerabilities): 88.2%
• ExploitGym (exploitation): 54.8%, up from Argon’s 42.0%
• CWE-bench (patching, single attempt): 71.0%

Google also says Carbon beats Argon on all 17 benchmarks it published. For developers, the model takes text, image, video and speech input and has a 1 million token context window.

Why you can’t use Gemini 4 Carbon yet

Starting this week, we will make Carbon available to a subset of cyber defenders.

That line from Kavukcuoglu’s post is the whole story for now. Those defenders come through Google DeepMind’s Fairwind Program, which has more than 650 partners, including government cyber agencies, critical-infrastructure operators in healthcare, telecom and energy, and major tech platforms. According to WinBuzzer, participants need background checks and phishing-resistant multi-factor authentication, access is limited to internal security, incident-response or penetration-testing teams, and sharing or reselling access is banned.

The logic is simple. A model that is very good at finding holes in software is also very good at helping someone who wants to break in. Google wants defenders to get a head start on patching before that capability is one prompt away for everyone. The company is also taking part in the U.S. government’s voluntary pre-release testing.

There’s no date for anyone else. Google says it looks forward to bringing Carbon to developers, businesses and consumers soon, and Sundar Pichai told followers the model is rolling out “gradually + responsibly,” with “wider availability soon, so hang tight!”

When will it reach the Gemini app?

Argon is the best guide. Google announced it on September 30 for Fairwind partners only, then opened it on October 13 to paying API developers and eligible Gemini app subscribers. That’s a 13-day gap. If Carbon follows the same playbook, a mid-November arrival is plausible. Treat that as an educated guess based on Argon’s timeline, not a Google promise; a model with sharper offensive-security skills could easily get a slower rollout. When it does land, expect AI Pro and Ultra subscribers to see it first, as they did with Argon.

How Gemini 4 Carbon stacks up

Google’s own chart shows Carbon leading 15 of 17 benchmarks. Independent numbers are more mixed, which is the honest way to read any launch week.

So it’s not a clean sweep. GPT-6 Astra still leads the FrontierSWE 2 software-engineering test, and Claude Opus 5.5 wins the OSWorld computer-use subset by almost six points. On Artificial Analysis’ Coding Agent Index, Carbon running in Google’s Antigravity CLI scored 59, against 62 for Claude Code with Opus 5.5.

There’s a reliability wrinkle, too. Artificial Analysis measured Carbon’s hallucination rate at 11%, up from Argon’s 4%, though Carbon also answered more questions correctly and the figure remains among the lowest the group has recorded. And the security scores are Google-reported, from tests run without guardrails. The restricted version that eventually reaches the public may not score the same.

What it costs developers

On the API side, Google is matching Argon’s introductory pricing: $2 per million input tokens and $10 per million output tokens, with cached input 95% off. After an undated introductory period, that doubles to $4 and $20. Carbon is also wordier. Artificial Analysis found it generated about 200 million tokens on its index run versus roughly 110 million for Argon, which puts the cost per index task at $7.76 against Argon’s $7.68 at standard rates.

What it means for Android users

Most people reading this will meet Carbon through the Gemini app on their phone, not an API console. When it arrives, the bigger win probably won’t be trivia answers. It’ll be the agentic features Google keeps wiring into Android, the multi-step tasks and app actions where a model that keeps working through dead ends matters more than a few benchmark points. The Connected Apps platform Google previewed for developers this week, which lets apps plug into Gemini and Search, is exactly the kind of surface where a stronger agent model pays off.

There’s a quieter upside as well. If Carbon is as good at finding and fixing bugs as Google says, the first beneficiaries are the security teams behind services you already use. Google claims its Chrome team found Carbon more than ten times better than an unnamed rival’s frontier model at producing high-quality patches. That’s a vendor claim with no published data, but it’s the right job to point this model at first.

My take: defenders-first is the responsible call, even if it makes for an anticlimactic launch day. If you’re an AI Pro subscriber, there’s nothing to do yet. Keep using Argon, keep an eye on the model picker in the Gemini app, and don’t upgrade your plan expecting Carbon tomorrow.

Sources: Google, 9to5Google, WinBuzzer, Latent Space. Photo by Luca Bravo on Unsplash.