Google unveils Gemini 4 Carbon and Genie 4 AI models

Specialized tools for security

Google recently announced the launch of Gemini 4 Carbon, a model specifically engineered to handle complex coding needs and software security. Unlike general-purpose models, Carbon is built to manage long-running engineering tasks that require deep reasoning across large amounts of code. It is designed to identify vulnerabilities and suggest specific fixes for software developers.

The company reported high performance on several industry benchmarks, including SEC-bench, ExploitGym, and CWE-bench. Specifically, the model achieved 88.2% for vulnerability detection, 54.8% for exploitation, and 71% for vulnerability patching. These metrics suggest that the model can effectively navigate multi-step problems that often cause errors in standard coding environments.

To ensure the model is effective in real-world scenarios, Google is providing access to a group of cyber defenders through its Fairwind programme. The company is also participating in a voluntary pre-release testing programme organized by the US government. While these experts get early access, the company plans to schedule wider availability for businesses and consumers at a later date.

Creating interactive worlds

In addition to security tools, Google unveiled Genie 4, a model capable of generating explorable virtual environments. Unlike traditional chatbots that simply react to text, Genie 4 mimics the behavior of objects, characters, and scenes in real time. This allows users to interact with a generated world as it unfolds, rather than just receiving a static response.

The model is notable for its ability to produce native audio, including ambient noise, music, and interactive dialogue alongside the visuals. Users can modify elements like the weather, camera movements, and specific objects using natural language prompts. This level of interactivity makes the model a potential tool for various creative and technical fields.

Google identifies several potential applications for this technology, including video game development and film pre-visualization. It could also be used for educational purposes or for robotics and simulation-based training. The company is currently exploring ways to integrate Genie and Gemini models to improve consistency and planning within these generated environments.

Technical limitations and safety

Despite its capabilities, Google has clarified that Genie 4 is currently an experimental product with several limitations. The model can lose consistency during longer sessions or when dealing with complex physical interactions. Additionally, the AI still struggles with perfect spatial accuracy and high-quality text rendering within the generated scenes.

To address concerns regarding the spread of synthetic media, Google will apply SynthID watermarks to all AI-generated video and audio. This technology helps people and systems identify content created by artificial intelligence. These safeguards are part of the broader rollout as the company tests the model's reliability with trusted partners.

The rollout of these models is being handled gradually to ensure they meet performance standards before reaching a wider audience. While Gemini 4 Carbon is being tested by cybersecurity experts, Genie 4 is entering a research preview for trusted testers. This staged approach allows Google to refine the models' behavior in specialized environments first.

Availability and pricing

For those interested in using these new models, Google has established a specific path for access. Users who want to try Genie 4 can join a waitlist through Project Genie. Currently, the model is only available for a trial by trusted partners and through the Project Genie research preview.

Existing Google AI Ultra subscribers will continue to have access to Google AI Ultra via the API and Google AI Studio. This access will also be available through Project Genie and selected partners. The company has not yet announced a specific date for when the general public will be able to use Gemini 4 Carbon or Genie 4.

Pricing for the new models remains consistent with the previous generation. Google's API pricing is set at $2 per million input tokens and $10 per million output tokens. This is the same pricing structure that was established for the Gemini 4 Argon model, providing a predictable cost for developers who transition to the new models.