Google Launches Gemini 4 Carbon for Developers and Enterprises
On November 10, 2026, Google announced the availability of Gemini 4 Carbon, its latest AI model aimed at developers, enterprises, scientists, and consumers. This model is touted as the strongest in the Gemini series, offering frontier performance in cybersecurity and advanced capabilities for complex knowledge work, software engineering, and scientific reasoning.
Two weeks prior, Gemini 4 Carbon was made accessible to select cybersecurity defenders through Google’s Fairwind Program. Following the introduction of Gemini 4 Argon, Carbon is designed for tasks that challenge the limits of AI intelligence, including intricate coding, debugging, and open-ended scientific research. It utilizes more tokens than Argon to facilitate deeper exploration and verification processes.
In independent benchmarks assessing model performance in real-world workflows, Carbon achieved top rankings on CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench, demonstrating over twice the cost efficiency per task compared to the next best external model.
To enhance its capabilities for larger and more complex projects, Google is previewing a feature called `/swarm` in Google Antigravity. This feature allows a single prompt to coordinate a team of agents in real time, enabling developers to assign Carbon to roles requiring extensive reasoning, such as architectural design and auditing.
Gemini 4 Carbon is accessible via the API in Google AI Studio and Gemini for business for all paying developers and enterprises, maintaining the same introductory pricing as Gemini 4 Argon at $2 per million input tokens and $10 per million output tokens. It is also being rolled out to Google AI Ultra subscribers across various platforms.
Advanced Knowledge Work and Data Analysis
In sectors like legal, finance, and enterprise operations, Gemini 4 Carbon integrates advanced visual understanding and reasoning to generate insights from large datasets. It adheres to strict schemas for producing polished spreadsheets, presentations, and reports.
To better evaluate multimodal performance, Google partnered with Artificial Analysis to create the Multimodal Knowledge Agent (MKA) benchmark. In this benchmark, Carbon ranked second while maintaining a cost per task more than six times lower than the leading model.
In pre-release evaluations, Carbon matched the performance of top models in deep financial research and excelled in various benchmarks assessing legal expertise and data engineering.
Enhanced Coding and Development Capabilities
Carbon excels in engineering tasks requiring collaboration among multiple agents, such as complex architectural redesigns and cross-subsystem investigations. It effectively manages build-and-verify loops for web app development, ensuring the production of complete and functional applications.
Cybersecurity Innovations
Gemini 4 Carbon delivers advanced cybersecurity capabilities, providing security teams with comprehensive intelligence for protecting critical software. In the Fairwind Program, Carbon demonstrated superior performance in identifying, exploiting, and patching vulnerabilities.
Early testers have reported significant improvements in cybersecurity tasks. For instance, in Project Fortify, Carbon achieved 100% recall in scanning and 95% precision in triage during internal benchmarks. Additionally, it has been instrumental in identifying previously undetected vulnerabilities in open-source software.
Scientific and Mathematical Reasoning Improvements
Gemini 4 Carbon has shown notable advancements in reasoning across various scientific domains, aiding in drug toxicity research and other complex scientific analyses. Researchers at Axiom Bio have reported that Carbon is now their leading model for translational science, discovering novel mechanisms of toxicity.
Google researchers have utilized Carbon to accelerate their work in Google Antigravity, applying it to tasks such as formal code verification and enzyme assay development, significantly improving efficiency and accuracy.
Safety and Security Enhancements
With the release of Gemini 4 Carbon, Google has reinforced its safety and security measures, focusing on misuse prevention and model alignment. The model’s safeguards have proven effective in simulated cyberattack scenarios, significantly reducing the success rate of potential breaches.
Furthermore, Carbon has achieved a 0% success rate for prompt injection attacks, marking a significant milestone in model security. Ongoing advancements in automated red teaming are expected to further enhance its robustness against such vulnerabilities.
Google continues to onboard participants for its research programs, expanding access to Gemini 4 Carbon for scientists and developers as they become available.
Source: Google DeepMind