Summary: Google has launched Gemini 4 Carbon. It is the company’s newest and strongest AI model. It handles hard coding jobs, and it finds and fixes security flaws in software. Most people cannot use it yet — but its price and rollout plan are already public.
What happened
On October 28, 2026, Google announced Gemini 4 Carbon. The news came in ablog post from Google DeepMind. The post calls Carbon the strongest model Google has built so far.
And it arrived fast. Carbon comes just two weeks afterGemini 4 Argon, the first model in the Gemini 4 family. Google says its researchers now work side by side with its own AI models, and that is why new models arrive so fast.
Google says Carbon is its best model for two kinds of work. The first is cybersecurity — finding weak points in software. The second is long-horizon work, which means long jobs that take many steps to finish.
What Gemini 4 Carbon can do
Google shared test scores for Carbon. Keep in mind, these scores come from Google itself, so read them as the company’s own claims. Even so, they show where Google focused its work.
Carbon ranks first on the Artificial Analysis Intelligence Index, version 5.0. Its score is 59, while Argon scored 50 on the same index. That index ranks AI models by how smart they are across many tasks.
The biggest gains are in security tests:
- SEC-bench Pro: 88.2%. This test checks how well a model finds flaws in real software.
- ExploitGym: 54.8%. This test checks how well a model can prove a flaw is real.
- CWE-bench: 71.0%. This test checks how well a model can fix a flaw.
Google says Carbon holds a firm lead on all three. On its own internal tests, Google says Carbon writes secure code at more than twice the success rate of Argon. Its Chrome team also reports something striking: Carbon wrote high-quality fixes more than ten times better than another lab’s top model. Google did not name that model.
What about building websites? Carbon jumped 89 points on the WebDev Arena leaderboard and now sits in fourth place. Google says it also improved at science tasks and at understanding images and other media.
Then there are the long jobs. Google says its own teams ran Carbon for days on end. The model worked through dead ends on hard engineering tasks — finding the root cause of subtle bugs and fixing them across large codebases.
Google also worked on honesty. It says Carbon cuts unsupported claims by about half compared with Argon. When facts are missing, Carbon states that plainly instead of making up an answer. It also shows the user where sources disagree before it acts. And in multi-step jobs, Google says Carbon reduced overreach by 26% and under-service by 29%. In plain words? It stays in scope better, and it finishes more of the job.
Who can use it right now
Here is the catch: you cannot open the Gemini app today and use Carbon. It is not in the app, not in AI Studio, and not in the public API yet.
Google is rolling it out in stages. This week, a small group of cyber defenders gets access first. More trusted partners will follow through Google’s Fairwind Program, which works with vetted groups such as security teams and large firms.
Google has not set a date for full public access. It says developers, businesses, and everyday users will get Carbon “as soon as possible.”Google is also taking part in a voluntary pre-release testing process with the U.S. government, and that is one reason the rollout is slow. A model that can find and test software flaws is powerful. In the wrong hands, those same skills could be misused.
The price is already set, though. When Carbon reaches the API, it will launch at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens. Tokens are the small chunks of text that a model reads or writes. Cached input tokens cost 95% less than the input price. After the introductory period ends, the price rises to $4 per million input tokens and $20 per million output tokens. Google has not said when that change will happen.
Why this matters for freelancers and students
If you build websites or apps for clients, Carbon is worth watching. A model that finds flaws and writes fixes could save you hours on client projects. It could also help you check your own code for security holes before you hand it over — and that kind of check is a service you can sell.
Students who study coding can learn from this trend too. AI models now handle longer and harder coding jobs. So the skill that matters most is not typing code fast. It is knowing what to ask, how to check the output, and how to fix mistakes. You can practice those skills with the tools you have today, such as thefree AI tools for studentsin our guide.
There is also a warning in this launch. Tools like Carbon will not stay rare for long. Only two weeks passed between Argon and Carbon. The pace of new releases is now very fast, and if you wait a year to learn AI tools, you will fall far behind. Start small, use one tool well, and keep up with each new release.
What to watch next
Three things will tell us how big Carbon really is. First, when does it reach the public API and the Gemini app? Second, do independent testers match Google’s scores once they can try it? Third, does the price stay low enough for freelancers and small teams?
For now, Carbon is a promise, not a tool in your hands. But it is a strong sign of where AI is headed: models are getting better at long, hard, real-world jobs, not just short chats. The people who learn to work with them early will have the edge.