On November 11, Google DeepMind officially announced that its latest flagship model, Gemini 4 Carbon, is now available for developers, businesses, researchers, and general consumers. Carbon is currently the most capable publicly launched model in the Gemini family, but general consumers must subscribe to Google AI Ultra to experience it. Gemini Carbon excels in complex programming, cybersecurity, scientific reasoning, and long-duration multi-agent tasks. In the CursorBench 4.0, Perplexity WANDR, and Zapier AutomationBench benchmark tests, Carbon
ranked first, with single-task costs more than 50% lower than the next best external competitive model.
In terms of pricing, Carbon's API offers a discounted rate similar to Argon, with a cost of US$2 (approximately HK$16) per million input tokens and US$10 (approximately HK$78) per output token, which will be adjusted to US$4 (approximately HK$31) and US$20 (approximately HK$156) respectively starting December 1. In multi-agent collaboration, Carbon introduces the /swarm feature in the AI programming platform Google Antigravity, allowing developers to organize multiple AI agents for collaborative tasks through commands.
Carbon itself can handle architectural design and task coordination, having successfully delivered a complete open-world 3D game and generated an interactive 3D model containing approximately 906,000 triangles based on seven urban planning PDF documents (including 206 engineering drawings) during an overnight run.
Gemini 4 Carbon's Outstanding Performance in Cybersecurity and Research
In the field of cybersecurity, Carbon has demonstrated leading performance in benchmark tests such as SEC-bench Pro, achieving a 100% recall rate in internal system vulnerability scans and a classification accuracy of up to 95%. In practical tests, Carbon assisted Wiz in discovering zero-day vulnerabilities in two large SaaS providers in less than two hours and successfully blocked 99% of attack paths during simulated attack tests. In the research domain, Carbon has driven the development of the AlphaProof Code framework, completing mathematical proofs of key security properties of the Signal encryption protocol in weeks—traditional methods would take years.
Additionally, in the CASP17 blind test dataset across seven low-homology protein families, the number of homologous sequences identified by Carbon increased on average by 40 times.
At the enterprise operational level, Carbon is suitable for financial and business scenarios, capable of generating clearly structured Excel financial models and high-quality PPT presentations, and can extract key information from office data to directly produce analytical conclusions and action recommendations. In terms of access channels, developers can invoke Carbon through the paid tier of the Gemini API in Google AI Studio and Android Studio; enterprise users can access it via Gemini for Business; researchers can apply for a free research program to gain six months of high-access privileges; and cybersecurity researchers must apply through the Fairwind Program.
For general consumers, Carbon is only available to subscribers of Google AI Ultra, covering the Gemini App and Google Search AI Mode.

