SIAgentNews.com
← All news
DevelopingGemini

Google DeepMind releases Gemini 4 Carbon for complex software tasks

The new model can reverse engineer old games and is now available to developers for an introductory price.

· First reported by Inven Global

$2Introductory price per million input tokens

The short version

  • Gemini 4 Carbon can reverse engineer original game binaries and port them to the web in a single step.
  • The model ranks first on CursorBench 4.0, WANDR, and AutomationBench tests.
  • It is priced at $2 per million input tokens and $10 per million output tokens until November 30th.
  • Google also introduced /swarm, a feature that turns a single prompt into a team of real-time AI agents.

Why it matters

This release signals a shift toward AI models that can handle multi-step engineering workflows rather than just generating text. For developers, it offers a high-performing tool for complex coding and cybersecurity at a lower cost than competitors.

Reverse engineering the past

Google DeepMind CEO Demis Hassabis recently demonstrated the capabilities of the new Gemini 4 Carbon model by using it to reconstruct his own history. The model was tasked with reverse engineering the original binaries of Theme Park, a management game Hassabis programmed during the 1990s. This process involved taking the old code and porting the entire experience to the modern web.

The demonstration showcased the model's ability to handle deep logic and legacy code structures. In addition to recreating the original game, Carbon added a new DeepMind-themed ride to the experience. Hassabis described the ability of the model to perform these complex transformations in one shot as "kind of mindblowing."

This specific use case highlights a move toward functional software recreation. While previous versions of Gemini were used to create testbeds of the game in a few hours, Carbon handles the actual conversion of the original binaries. It shows a transition from simply mimicking a game's mechanics to actually understanding and rewriting its underlying architecture.

Engineering and science focus

Google describes Gemini 4 Carbon as its most powerful model for handling long and complex tasks. The development team specifically designed it for use in software engineering, cybersecurity, knowledge work, and scientific research. Unlike models optimized for quick chat responses, Carbon is built to manage projects that require sustained reasoning over large amounts of data or code.

To prove its effectiveness, Google released performance data from several industry benchmarks. According to Koray Kavukcuoglu, the Chief AI Architect at Google, the model ranks first on the CursorBench 4.0, WANDR, and AutomationBench tests. These results suggest that Carbon can outperform other models while maintaining a more efficient cost structure for the user.

The model is also being integrated into specialized platforms like Google Antigravity. This includes a new feature called /swarm, which allows users to turn a single prompt into a team of AI agents. These agents can be tracked in real time as they work together to solve a problem, providing a more collaborative approach to automated tasks.

Public release and pricing

After being tested by a group of cybersecurity defenders for two weeks, Gemini 4 Carbon is now moving into the public sphere. It is available to paying developers and enterprises through the API in Google AI Studio. This release marks a significant step in making high-end engineering tools accessible to a wider audience of creators.

Google is offering an introductory pricing structure to encourage adoption of the new model. Currently, the cost is set at $2 per million input tokens and $10 per million output tokens. These prices are scheduled to increase to $4 and $20, respectively, after the month of November ends.

Subscribers of Google AI Ultra can also access the model directly within the Gemini app. Furthermore, it is being integrated into other core services, including Antigravity and AI Mode in Search. This multi-channel rollout ensures that both professional developers and general AI users can interact with the model's capabilities.

Interactive demonstrations

To demonstrate the model's capabilities in a more visual and interactive way, Google showcased several game-based projects. One notable example is Pelican Quest, which is an open-world 3D browser game. The game is set on Google's Mountain View campus and allows players to navigate the environment by riding a bike as a pelican.

This project was built using a combination of Gemini 4 Carbon and the /swarm feature. By using these tools, the developers were able to create a 3D environment that functions within a web browser. It serves as a practical example of how the model can handle complex 3D assets and interactive logic simultaneously.

These examples serve to ground the technical specifications of the model in tangible results. By showing that Carbon can build a functional 3D game or port a 1990s management sim, Google is aiming to prove its utility for complex software development. It moves the conversation from what the AI can say to what the AI can actually build.

What comes next for Carbon

The release of Gemini 4 Carbon establishes a new baseline for how AI models interact with existing software infrastructure. Because it can handle long-form tasks, it is expected to become a primary tool for developers looking to migrate legacy systems to the cloud. Its ability to handle complex binaries suggests it could eventually be used for broader software auditing and modernization.

The integration of the /swarm feature also points toward a future of multi-agent workflows. Instead of a human directing every single step of a coding project, they may soon only need to provide the high-level goal. The AI agents will then coordinate the sub-tasks, such as debugging, testing, and deployment, in a real-time environment.

As the model moves past its introductory pricing period, the industry will watch how enterprises adopt it for high-stakes work. Its performance on the AutomationBench and CursorBench will be key indicators of whether it can truly replace manual engineering steps. For now, it stands as Google's strongest offering for users who need more than just a conversational partner.

Sources and links

First reported by Inven Global. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news