Google Gemini 4 Carbon Is Here: Google’s Strongest AI Model Yet, With a Price That Doubles December 1

Google has released its most powerful artificial intelligence model yet. Gemini 4 Carbon went live on November 10, 2026, for paying developers and businesses, built for the long, difficult jobs other AI systems still struggle with: multi-hour software engineering, deep research across dense documents, and cybersecurity defence.

The launch is also a direct challenge to Anthropic, whose Claude Fable 5.2 took the top spot on the Artificial Analysis Intelligence Index just one day earlier. Google’s answer is slower, but dramatically cheaper per task — and, for now, half its December price.

What Is Gemini 4 Carbon?

Gemini 4 Carbon is Google’s strongest model, sitting above Gemini 4 Argon in the company’s lineup. Google describes it as a model for long-horizon agentic workflows — jobs where an AI agent works on its own for a long time, using tools, reading files, and making decisions across many steps.

The model was first previewed on October 28, 2026, when Google gave it to vetted cybersecurity defenders through its Fairwind Program — a deliberate first step, since cybersecurity is Carbon’s headline strength. On November 10, it became generally available to all paying developers on the Gemini API.

Carbon accepts text, images, video, audio, and PDF files as input, and produces text as output. Its context window is 1,048,576 tokens — roughly a million shared between input, thinking, and output — so it can hold very large codebases and long reports in mind at once.

It also offers three thinking levels — low, medium, and high — with medium as the default. The higher the level, the longer the model reasons before answering, and the better it tends to do on hard tasks. The trade-off is time and cost.

Google still recommends Gemini 4 Argon for most coding and knowledge work. Carbon is for the hardest, longest jobs, where Argon would stall or give up — a specialist for work that needs stamina, not a replacement for everything.

Where You Can Use Gemini 4 Carbon?

Carbon is available now through the Gemini API in Google AI Studio and Android Studio, and through Gemini for business customers. Consumers can find it in the Gemini app, Google Antigravity, and AI Mode in Search — but only with a Google AI Ultra subscription. There is no free tier.

Carbon is aimed first at developers and enterprises. Free users will not see it; Ultra subscribers get it in the app, Antigravity, and AI Mode in Search.

The Benchmarks: Where Carbon Leads — and Where It Does Not

Google’s headline claim is that Carbon ranks first on three workflow benchmarks — CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench — at more than twice lower cost per task than the next best model.

Independent data backs up part of that story. On CursorBench 4.0, Carbon scored 60.4 percent on high thinking at 9.22 dollars per task. Claude Fable 5.2 at maximum effort scored 60.0 percent — but at 32.79 dollars per task, nearly four times the cost. On AutomationBench, Carbon scored 53.88 percent at about 2.12 dollars per task, ahead of Fable 5.2 at 51.45 percent and 4.64 dollars.

But the picture is not one-sided. On the broad Artificial Analysis Intelligence Index, Carbon reaches 60.0 on high thinking — behind Claude Fable 5.2 at about 65, and only just ahead of Claude Opus 5.5 at 59. On medium thinking, its default, it scores 58.3. On the widest measure, Google’s new model is second, not first; its advantage is in specific kinds of work.

The coding results split the same way. Carbon leads on Terminal-Bench 4.0, a test of real terminal-based software tasks, with 65.2 percent on high thinking, ahead of Claude Opus 5.5 at 59.6 percent. But on FrontierCode 2.0, which grades real changes to software repositories against the standards of an actual code review, Fable 5.2 scores 49.1 percent against Carbon’s 40.6 percent. Different tests, different winners.

Carbon also shines in knowledge work. Google partnered with Artificial Analysis to create a new test, the Multimodal Knowledge Agent benchmark, which checks how a model reasons over dense files full of charts, images, blueprints, and tables. Carbon ranks second on that test while costing more than six times less per task than the top model.

On cybersecurity, the numbers are striking. In Google’s own figures, Carbon scored 88.2 percent on SEC-bench Pro (finding vulnerabilities), 54.8 percent on ExploitGym, and 71.0 percent on CWE-bench (patching them).

One early independent review gave Carbon 4.5 out of 5 for hard reasoning and research — and only 2 out of 5 for anything a person has to wait for.

Code on a developer's screen. Carbon is built for long, multi-hour software engineering jobs. Photo by Lukas on Pexels.
Code on a developer’s screen. Carbon is built for long, multi-hour software engineering jobs. Photo by Lukas on Pexels.

The Price: Cheap Now, Double on December 1

Through November 30, 2026, Carbon costs 2 dollars per million input tokens and 10 dollars per million output tokens — exactly the same introductory price as Gemini 4 Argon. Cached input, for repeated context the model has already seen, costs just 10 cents per million tokens.

From December 1, the price doubles to 4 dollars per million input tokens and 20 dollars per million output tokens, with cached input rising to 20 cents. That is still far cheaper than Claude Fable 5.2, at 8 dollars input and 40 dollars output per million tokens.

But the sticker price is not the whole bill. Carbon thinks longer, and Google’s documentation says it spends more thinking tokens per turn than Argon — billed as output tokens, the expensive kind. Artificial Analysis estimates 7.76 dollars per task for Carbon on high thinking, against 26.54 dollars for Fable 5.2 at maximum effort.

If you want to try Carbon, try it before November 30, while the introductory price lasts. Route only the hard, open-ended jobs to it — planning, cross-system debugging, security review, and verification — and keep everyday coding and writing on cheaper models like Argon.

Carbon vs. Claude Fable 5.2: Which One Wins?

The honest answer is that it depends on the job. Fable 5.2, launched by Anthropic on November 9, is the top-scoring model on the broad Artificial Analysis index at about 65, and it famously coordinated more than 1,000 AI agents over eight days to build a working operating system from scratch. It is also expensive: 8 dollars per million input tokens and 40 dollars per million output tokens, with no introductory discount.

Carbon is cheaper per task on most workflow benchmarks that matter to working developers, and it leads on several of them. But it is slower, and on some tests — notably FrontierCode 2.0 — it trails Fable clearly. Fable is the stronger generalist; Carbon is the cheaper specialist for long, structured work.

For a business, the choice may come down to where you already live: Google Cloud and AI Studio point to Carbon, while Claude-based teams will stay with Fable 5.2. Many teams will end up using both — Fable for the hardest reasoning, Carbon for long-running automation where the bill matters.

The Catches: Slow Thinking and Security Questions

Carbon has three real weaknesses, and Google does not hide them.

The first is speed. On high thinking, the model can take about a minute before its first word, and one independent review measured 25.6 minutes per benchmark task on average. Fine for a job you leave running overnight; useless for a chatbot or anything where a person is waiting.

The second is cost control. Because Carbon thinks longer and bills that thinking as output, a vague request can run up a large bill. Google’s guidance is to use Carbon only for the steps where it finishes more tasks, and to keep everything else on cheaper models.

The third is security. A model this good at finding and exploiting vulnerabilities is a powerful tool for defenders — and a dangerous one in the wrong hands. That is why Google first gave Carbon to vetted defenders through the Fairwind Program, and why OpenAI’s own cybersecurity model, GPT-6 Cyber, remains limited to approved customers. The debate over who should get access to frontier cyber capabilities is only just beginning, and Carbon sits at the centre of it.

Who Should Try Gemini 4 Carbon?

If you are a software developer working on large, complex codebases, Carbon is worth testing now, while the price is low. It is built for jobs where you currently break work into small pieces because the model loses track — big refactors, debugging across several systems, and reviewing security-sensitive code.

A developer works on a laptop showing code. Carbon is available in Google AI Studio and the Gemini app for Ultra subscribers. Photo by Mike Ngo on Pexels.
A developer works on a laptop showing code. Carbon is available in Google AI Studio and the Gemini app for Ultra subscribers. Photo by Mike Ngo on Pexels.

If you are a researcher or analyst working with dense documents — financial reports, scientific papers, legal files full of tables and charts — Carbon is a strong candidate. The same goes for security teams.

If you are a casual user who mostly needs quick answers and short writing, Carbon is not for you. Argon, or the standard Gemini app, will be faster and cheaper. Free-tier users cannot use Carbon at all — it is a paid product.

What Happens Next

The next few weeks will tell us how good Carbon really is. Independent testers are already running it through real projects, and their results — not Google’s launch benchmarks — will decide its reputation. Watch whether developers keep using it after the December 1 price rise.

One thing is already clear. In the space of a week, Google shipped Carbon, Anthropic shipped Fable 5.2, Meta shipped its tiny and cheap Muse Ember model, and NVIDIA unveiled a new robot brain. The race is no longer about who has the smartest model in a lab, but who can do real work at a price real businesses can afford. Carbon is Google’s strongest answer so far — and until November 30, its cheapest.

Leave a Comment