Google launches Gemini 4 Carbon and Genie 4 AI models

New tools for security and worlds

Google announced two new artificial intelligence models on Wednesday, October 28, to expand its capabilities in cybersecurity and simulation. The new models, Gemini 4 Carbon and Genie 4, were introduced just two weeks after the release of Gemini 4 Argon. Each model serves a distinct purpose, ranging from deep code analysis to the creation of interactive digital spaces.

Gemini 4 Carbon is positioned as a specialized tool for complex engineering tasks that require long-running attention. It is built to identify software vulnerabilities and suggest specific code fixes to prevent exploits. By focusing on these high-level tasks, Google aims to improve AI-assisted security for software development and robotics research.

In contrast, Genie 4 focuses on the creation of explorable virtual environments. Unlike standard chatbots that simply react to text, Genie 4 mimics the behavior of objects and characters within a scene. It can generate visuals along with ambient noise, music, and interactive dialogue based on natural language prompts from the user.

How Gemini 4 Carbon handles code

Gemini 4 Carbon is designed to solve multi-step problems that often cause errors in large volumes of code. It can detect vulnerabilities, check for possible exploits, and recommend specific patches to fix issues. According to Google, the model is the most capable in its lineup for these types of complex, long-running engineering tasks.

The company provided specific performance metrics to demonstrate the model's accuracy. On the SEC-bench, the model achieved an 88.2% score for vulnerability detection. It also reached 54.8% for exploitation and 71% for vulnerability patching across the ExploitGym and CWE-bench testing environments.

Beyond just finding bugs, Google reported that the model has improved performance regarding secure code generation. It is also designed to better handle prompt-injection attacks, which are methods used to trick AI into performing unintended actions. These improvements are intended to provide a more robust layer of security for developers and businesses.

Interactive worlds with Genie 4

Genie 4 represents a move toward creating environments that users can control in real time. The model can change elements like the weather, camera movements, and specific objects through simple natural language prompts. This makes it a potential tool for various industries, including film pre-visualization and educational applications.

One of the standout features of Genie 4 is its ability to produce native audio alongside visuals. This includes the generation of music and interactive dialogue as scenes unfold. Google believes these capabilities could be particularly useful for simulation-based training and various video game applications.

The company is also looking into how it might integrate Genie and Gemini models together. The goal of this integration would be to enhance the planning and consistency of the generated environments. By combining these technologies, Google hopes to create more cohesive and predictable digital experiences for users.

Availability and limitations

Neither model is currently available to the general public for wide use. Gemini 4 Carbon is being rolled out to a group of cyber defenders as part of the Fairwind programme. It is also being used in a voluntary pre-release testing programme organized by the US government.

Genie 4 is currently in a research preview for trusted testers and selected partners. Users who want to try the project can join a waitlist through Project Genie. Google has noted that while the technology is impressive, Genie 4 is still considered an experimental product with some known limitations.

During longer sessions or complex physical interactions, the model can lose consistency in the environment. Additionally, spatial accuracy and the rendering of text are not yet perfect. To ensure transparency, Google will use SynthID to watermark all AI-generated video and audio, making it easier to identify content created by these models.

Pricing and future rollout

While the models are in limited release, Google has already established an API pricing structure for them. The cost is set at $2 per million input tokens and $10 per million output tokens. This pricing remains consistent with the previously released Gemini 4 Argon model.

Google plans to schedule wider access for developers, businesses, and general consumers at a later date. For now, the focus remains on specialized use cases in cybersecurity and research. The company will continue to offer access to Google AI Ultra subscribers through the API and Google AI Studio via Project Genie.

The rollout of these models highlights Google's strategy of moving toward specialized AI. By creating models that excel at specific tasks like vulnerability patching or world simulation, the company is providing tools for high-stakes engineering and creative work. The next steps involve refining these models based on feedback from the current trusted tester groups.