Google has opened access to its new flagship model Gemini 4 Carbon. It was created for complex tasks in programming, cybersecurity, data analysis, and scientific research, the company reported.
The model can perform multi-step tasks, verify its own results, and work together with other AI agents.
What Gemini 4 Carbon can do
Google describes Carbon as a model for tasks that require prolonged analysis and complex decisions. Unlike Gemini 4 Argon, which is oriented toward everyday programming and working with information, Carbon spends more computing resources on finding solutions and verifying results.
In particular, the model can work with large software projects, find errors in code, fix them, and check whether the changes have caused new problems.
For web development, Carbon is capable of creating applications, running them, analyzing the result, and independently fixing detected errors. The model also works with three-dimensional graphics and game development.
Separately, Google presented the /swarm feature in the Antigravity development environment. It allows launching several AI agents to jointly perform a single task. For example, a developer can assign Carbon to design the application architecture and verify the code, while delegating simpler parts of the work to faster models. The agents will work in parallel, and the user will be able to monitor their progress in real time.
In addition to programming, Carbon can analyze large volumes of documents, spreadsheets, charts, and images. Google notes that the model is capable of preparing financial reports, presentations, and spreadsheets according to specified requirements.
According to the testing results published by the company, Gemini 4 Carbon took first place in CursorBench 4.0, Perplexity WANDR, and Zapier AutomationBench. Google claims that in these tests, the cost of completing one task was more than twice lower than that of the closest competing model.
One of the main areas of development for Gemini 4 Carbon is cybersecurity. The model can analyze software, find vulnerabilities, verify the possibility of exploiting them, and propose fixes.
For example, during testing within the Wiz Scan for Good program, Carbon discovered previously unknown zero-day vulnerabilities in two major cloud software vendors in less than two hours.
Also, the Google Cloud Vulnerability Research team, with the help of Carbon, confirmed that two bugs in popular open-source software could be used for attacks.
In an internal test of the Project Fortify system, the model achieved 100% detection completeness and 95% accuracy when verifying the found problems. These figures relate to a specific Google test set, not to all possible vulnerabilities.
The company will provide Carbon's specialized cybersecurity capabilities to vetted professionals through the Fairwind program.
Google also tested Gemini 4 Carbon in scientific research. In particular, researchers used the model to search for related protein sequences in large databases. During an experiment with seven protein families, Carbon found on average 40 times more such sequences. This could help other models predict protein structure more accurately.
In addition, Carbon was used to analyze laboratory records and experimental data. The model helped find errors in laboratory procedures and improve research methods.
Where Gemini 4 Carbon is available and how much it costs
Developers can use Gemini 4 Carbon through the paid API in Google AI Studio and Android Studio, and companies through Gemini for business.
For regular users, the model is gradually becoming available in the Gemini app and AI Mode in Google Search with a Google AI Ultra subscription. It is also appearing in Google Antigravity.
Until November 30, 2026, an introductory API price will be in effect:
$2 per million input tokens;
$10 per million output tokens.
After the introductory period ends, the price will rise to $4 per million input tokens and $20 per million output tokens.
Google notes that Carbon is intended primarily for the most complex tasks, while Gemini 4 Argon remains the main model for most everyday work scenarios.