← All writing
18 Nov 2026 · AI models

Google's Gemini 4 Carbon: what its strongest AI model can do

On 10 November 2026, Google DeepMind released a new AI model called Gemini 4 Carbon to developers, businesses, scientists and everyday users. Google describes it as its strongest model yet.

An AI model is the software brain behind tools such as the Gemini app. Carbon is built for the hardest jobs those tools are asked to do.

What is Gemini 4 Carbon?

Carbon is aimed at long, multi-step tasks, mainly writing and fixing software, cybersecurity, complex office work and scientific research.

Google calls much of this agentic work. An agent is an AI that does more than answer one question: it works through a task step by step and checks its progress as it goes.

Carbon follows Gemini 4 Argon, which Google released shortly before. Google says Argon is still the best choice for most everyday software and office work, and Carbon is for tasks that push a model to its limits, such as tricky coding problems or open-ended research.

The main difference is effort. AI models read and write text in small chunks called tokens, and Carbon uses more of them than Argon. That lets it explore a problem more thoroughly and double-check its work repeatedly.

If Argon is a quick, reliable check-up, Carbon is the full inspection you book when something really matters.

What can it do?

Google gives examples in four main areas:

  • Office work: reading large piles of dense documents in fields such as law and finance, including charts, tables and drawings, and turning them into tidy spreadsheets, presentations and reports.
  • Building software: creating complete, working web apps. It recovers from its own mistakes and checks both the code and how the finished page looks.
  • Security: finding weaknesses in software, working out how attackers could use them, and writing fixes.
  • Science: working through problems in physics, maths and life sciences, and making sense of experimental data.

The security company Wiz used Carbon to check two major online software providers, and it found previously unknown security holes in under two hours. According to Google, that turned weeks of research into a single afternoon. Wiz then worked with the affected organisations to fix the problems.

Axiom Bio studies whether drugs could be toxic. The company says Carbon found new explanations for toxicity that its team had not spotted before.

Google researchers also used Carbon to help prove important security features of the code behind Signal, the messaging app. That code scrambles messages so only the intended people can read them. The researchers did this in weeks; Google says such work has historically taken expert teams years.

There are playful demonstrations too. In one 3D game you explore Google's Mountain View campus as a pelican riding a bike. Another is a playable fortepiano from around 1780, rebuilt from technical drawings and X-ray images.

A team of AIs from one request

Google is also previewing a feature called /swarm in its Antigravity tool. It turns a single request into a coordinated team of AI agents, and you can watch them as they work.

Carbon takes the roles that need the deepest thinking, such as planning, coordinating and checking the work. Faster, lighter models handle other parts in parallel, and separate checkers test every change.

In one example, a team of Carbon agents turned historical photos and engineering drawings into an interactive 3D exhibition showing how the Eiffel Tower was built between 1887 and 1889. It all came from a single request.

How good is it, and what does it cost?

AI companies often measure models with benchmarks, which are standardised tests for particular kinds of task.

Google reports that Carbon came first on three of them: CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench. Google says it also cost less than half as much per task as the next best model from another company.

On a new test of how well AI handles documents full of images, charts and tables, Carbon came second. Google says it cost more than six times less per task than the winner.

For businesses that connect Carbon to their own software, input (what you send the model) costs $2 per million tokens, and output (what it writes back) costs $10 per million tokens. That is the same as Argon. Google calls this an introductory price.

How can you try it?

Paying developers and businesses can use Carbon through Google AI Studio and Gemini for business. They connect to it through an API, a way for one piece of software to plug into another.

Carbon is also rolling out to Google AI Ultra subscribers in the Gemini app, AI Mode in Search and Google Antigravity.

Scientists have another route. Google earlier opened applications to give 10,000 scientists free access to Argon for six months, and that programme now includes Carbon. Applications are still open, subject to eligibility, capacity and selection.

Security teams got early access two weeks before the public launch through Google's Fairwind Program, which is where Carbon's top security test results came from.

Some things are still unclear. The /swarm feature is only a preview. The rollout to subscribers is still under way, so not everyone will see Carbon straight away. Google has not said how long the introductory price will last.

Many of the results come from Google's own testing or early partners, and in some areas, such as protein research, Google says broader testing is still under way. For most everyday tasks, Google itself suggests Argon remains the better fit.


Source: Gemini 4 Carbon, Google DeepMind. Listed in AI News Week 47, where we track every AI release each week.