Two weeks ago, a small group of cybersecurity defenders got a phone call most engineers only dream about. Google was handing them a new AI model before anyone else — not to test a chatbot, but to hunt real vulnerabilities in real software. The defenders kept quiet, ran it through their hardest problems, and waited.
On November 10, the wait ended. Google made Gemini 4 Carbon available to everyone else: developers, enterprises, and AI Ultra subscribers.
It is, in Google’s words, its strongest model yet — built for the hardest jobs in software engineering, scientific research, and cybersecurity. And it arrives one day after Anthropic’s Claude Fable 5.2 stole the spotlight, which makes the timing feel less like a coincidence and more like a challenge.
What Is Gemini 4 Carbon?
Gemini 4 Carbon is the newest member of Google’s Gemini 4 family, and it sits above Argon, the model that has powered most of Google’s AI work since the Gemini 4 launch.
Where Argon was built as the everyday workhorse, Carbon is the specialist: a model designed for tasks that test the limits of what an AI can reason through.
According to Google, Carbon is built for long-horizon agentic workflows — jobs that take many steps, span many files, and require the model to keep a plan in mind for hours rather than seconds.
Think of a multi-package software migration, a cross-system debugging hunt, or an open-ended scientific research problem. These are the tasks where most models start strong and slowly lose the thread.
Carbon builds on Argon’s foundation, with extra post-training for deep reasoning, extended verification, and multi-step agent work. It accepts text, images, audio, and video, with a one-million-token context window and a March 2026 knowledge cutoff.
Google first introduced Carbon on October 28, but only to cybersecurity defenders admitted through its Fairwind Program. For two weeks they tested it on real security work. Now the general version is out, while the specialised cyber configuration stays gated behind Fairwind applications.
Gemini 4 Carbon Benchmarks: Where It Leads
Google’s headline claim is that Gemini 4 Carbon ranks first on three independent benchmarks designed to test real workflows: CursorBench 4.0, Perplexity’s WANDR research test, and Zapier’s AutomationBench.
The company adds that it does so at more than twice the lower cost per task than the next best outside model.
The numbers on the public leaderboards tell a more nuanced story. On the Artificial Analysis Intelligence Index, Carbon scores 60 — behind Claude Fable 5.2, which scores 62 on its high setting and 65 at maximum effort.
But Carbon costs $7.76 per task on that index, against $9.08 for Fable 5.2 on high and $26.54 at maximum. Google’s argument is not that Carbon wins every race; it is that it wins enough of them at a fraction of the price.
Elsewhere, Carbon’s scores are genuinely striking. It leads on GPQA Diamond with 96.5%, tops Arena Hard Prompts at 1546, and leads Arena Coding at 1567.
On FrontierSWE 2, a demanding software engineering test, it scores 61.2% — ahead of Argon’s 55.0%, though behind GPT-6 Astra at 65.5% and Claude Opus 5.5 at 62.3%.
Google also partnered with Artificial Analysis to launch a new test, the Multimodal Knowledge Agent benchmark, designed to measure how agents reason over dense files full of charts, images, blueprints, schematics, and tables. Carbon ranks second on it, but at more than six times lower average cost per task than the top-performing model.
Built for Cybersecurity First
The most distinctive thing about Gemini 4 Carbon is where it started: in the hands of security teams. Google says Carbon delivers its strongest cybersecurity capabilities to date, covering the full arc of defence — finding vulnerabilities, exploiting them to prove they matter, and patching them.
On the benchmarks Google cites, Carbon leads on SEC-bench Pro for finding flaws, ExploitGym for exploiting them, and CWE-bench v1 for patching.
It inspects software, traces how flaws can be chained together across systems, and generates verified root-cause patches. In simulated cyberattack tests, Google says Carbon blocked 99% of attack attempts, and it passed 92% of its honesty tasks, against 70% for Argon.
Gemini 4 Carbon Pricing and Availability
Gemini 4 Carbon is available now through the API in Google AI Studio and Gemini for business, for all paying developers and enterprises.
AI Ultra subscribers can also use it in the Gemini app, in AI Mode in Search, and in Google Antigravity. Argon stays available alongside it — Google says Argon remains the ideal choice for most work, while Carbon is built for the hardest jobs.
The pricing matches Argon’s introductory rate: $2 per million input tokens and $10 per million output tokens, until November 30. From December 1, those prices double to $4 and $20.
But there is a catch: Carbon thinks for longer than Argon, so the same job consumes more tokens. Running one benchmark index, it produced 205 million tokens — almost three times the median — at $7.76 per task at full price.
Carbon is also excluded from the Gemini API free tier, and Google’s pricing page says every processed token is billed, including thinking tokens. For teams budgeting an AI rollout, the headline price is only half the story; the thinking time is the other half.
/swarm: One Prompt, a Team of Agents
Alongside Carbon, Google is previewing a new tool in Antigravity called /swarm. The idea is simple to describe and hard to build: type one prompt, and a team of AI agents splits the work, coordinates, and delivers the result together.
Google positions /swarm as the showcase for Carbon’s long-horizon abilities. A single model answering a single question is one thing; a team of agents refactoring a codebase is another.
For web app development, Google says Carbon runs build-and-verify loops, checks each other’s work, recovers from errors, and checks both its code and the rendered interface to produce complete, working results.
Gemini 4 Carbon vs Claude Fable 5.2
No launch happens in a vacuum, and this one arrived a day after Claude Fable 5.2 knocked Carbon off the top of the leading independent ranking.
Anthropic’s model is faster and cheaper per token in some configurations, and it leads the Artificial Analysis index outright. Google’s counter is cost per task: on the benchmarks where Carbon leads, it does the job for less than half what the next best outside model charges.
The rivalry now has a clear shape. Anthropic is pushing Claude into the workplace — spreadsheets, whiteboards, meetings — while Google pushes Gemini deeper into hard technical work: cybersecurity, finance, and long-running engineering.
Carbon also lands in a wider Gemini moment: Gemini Robotics 2.5 just put humanoids to work on Mercedes-Benz lines, while Gemini’s privacy practices face scrutiny over human review of Gmail and Drive data.
What Happens Next
The first test for Gemini 4 Carbon is not a benchmark but a bill. Developers will find out quickly whether the longer thinking time is worth the extra tokens, and whether the introductory price holds their attention once it doubles on December 1.
If the cost per task stays low in real work, Carbon could become the default for heavy jobs; if not, Argon will keep the everyday work and Carbon will stay a specialist tool.
The second test is /swarm. Multi-agent coding has been promised for years and has mostly disappointed, because agents drift, duplicate work, or quietly contradict each other. If Google’s preview turns one prompt into a genuinely coordinated team, that changes the economics of software development in a way no benchmark can capture.
And the third test is the one Google started with: cybersecurity. Carbon spent its first two weeks with defenders, and the Fairwind Program continues to gate the specialised cyber configuration.
If defenders keep finding it catches flaws humans miss — and patches them with verified fixes — Carbon’s legacy may be less about chatbots and more about everyday software quietly getting safer.
For now, the model is out, the price is set, and the race with Claude Fable 5.2 is officially on.
According to 9to5Google, Carbon is rolling out to AI Ultra subscribers now, with Argon staying right beside it. The hardest jobs in AI just got a new applicant — and it is already on the clock.
