Google launches Gemini 4 Carbon and Genie 4 models

Technical focus of Gemini 4 Carbon

Google has introduced Gemini 4 Carbon as a new frontier model designed for demanding technical and security workloads. Unlike general-purpose models, Carbon is optimized for complex reasoning and long-running agentic tasks. It aims to assist with sophisticated software development and cybersecurity operations.

The model shows significant progress in identifying security weaknesses and patching software. As Brave New Coin reported, it achieved a score of 88.2% on SEC-bench Pro for vulnerability detection. It also recorded 54.8% on ExploitGym and 71.0% on CWE-bench for software patching, which are standard evaluations for AI security capabilities.

Beyond security, Carbon is being tested for its ability to handle intricate 3D modeling and computer-aided design (CAD). In a side-by-side comparison, the model managed to maintain structural consistency in an airplane design containing over 2,600 parts. This suggests it could be useful for engineering visualization where components must remain connected and coherent.

Interactive worlds with Genie 4

While Carbon handles technical data, Genie 4 is built to create interactive, photorealistic simulations. It differs from standard video generators because it allows users to influence the environment in real time. Users can provide text instructions to change scenes, camera angles, or specific actions as the simulation unfolds.

A key feature of Genie 4 is its native audio, which synchronizes with the visual movements. The model can generate ambient sounds, music, engine noises, and character dialogue that respond to events. For example, a demonstration showed a motorcycle journey where the engine revs and ocean waves change based on the vehicle's movement.

Google DeepMind integrated Gemini to help Genie 4 track the state of these simulated environments. This integration is necessary for maintaining continuity as objects move or users modify the scene. It ensures that multi-step events remain consistent rather than becoming disconnected as the simulation progresses.

Security and safety measures

Because these models can assist with vulnerability discovery, Google is taking a staged approach to their release. Gemini 4 Carbon is initially being shared with selected cybersecurity professionals through the Fairwind Program. This allows the company to gather feedback on safety and suitability before a wider public launch.

The model includes specific safeguards designed to reduce its susceptibility to indirect prompt injection attacks. Google reported that Carbon achieved perfect resistance to these attacks in a specific evaluation. However, the company noted that this does not guarantee immunity to every possible cybersecurity threat or injection technique.

The rollout also aims to address model misalignment, which refers to the AI behaving in ways that do not match human intent. By testing with trusted cyber defenders first, Google can refine these protections. The goal is to ensure that the tools help defenders find weaknesses without being easily manipulated by bad actors.

Future applications and access

Google plans to expand access to Gemini 4 Carbon and Genie 4 to developers and other users in the near future. For Carbon, the initial API pricing is set at $2 per million input tokens and $10 per million output tokens. These rates may change as the models move toward a broader public release.

Genie 4 is expected to reach Google AI Ultra subscribers through a program called Project Genie. Selected partners will also receive early access through the Gemini API and Google AI Studio. These platforms will allow developers to experiment with the model's interactive capabilities and world-building features.

The technology has potential uses in robotics, where it can create environments to test control systems safely. It could also be used for virtual prototyping in engineering or creating interactive experiences for entertainment. While these simulations are not perfect copies of reality, they offer a way to test scenarios that might be too expensive or dangerous to do physically.