Google adds Gemini 4 Carbon to Search AI Mode for paid subscribers
Google is rolling out its most powerful model to AI Ultra users to improve reasoning and intent understanding in search results.
The short version
- Gemini 4 Carbon is Google's strongest model for cybersecurity and long-horizon agentic workflows.
- The model is available to all Google AI Ultra subscribers in Search AI Mode, the Gemini app, and Google Antigravity.
- Google claims the model offers better intent understanding for deeper back-and-forth conversations.
- Early testers reported the model can complete complex tasks, like rebuilding a game, from a single prompt.
Why it matters
This update moves Google Search toward a more conversational and reasoning-heavy experience. For power users, it means the AI can handle multi-step instructions and complex technical tasks more reliably.
A new frontier model
Google has officially released Gemini 4 Carbon, which the company describes as its strongest model to date. This new iteration is designed to provide frontier performance across several high-level categories. It specifically targets complex knowledge work, software engineering, and scientific reasoning.
The model is built to handle long-horizon agentic workflows. In the world of artificial intelligence, these are sequences of actions where the AI must plan and execute multiple steps to reach a final goal. By focusing on these workflows, Google intends for the model to be more capable of managing multi-stage projects rather than just answering single questions.
Beyond general knowledge, the model is highlighted for its cybersecurity capabilities. This suggests that the underlying architecture is robust enough to handle sensitive data security and complex threat analysis. It represents a significant step up in the capabilities available to Google's high-end subscription tiers.
Availability for subscribers
The rollout of Gemini 4 Carbon is currently targeted at Google AI Ultra subscribers. Users who pay for this premium tier can now find the model within several different platforms. This includes the Gemini app, Google Antigravity, and the AI Mode within Google Search.
To access the new model, paid subscribers can look for it in the Google Search AI Mode selector. This integration means that the core search experience will now be powered by a more sophisticated engine. It is designed to provide more helpful responses for users who ask the toughest questions.
Google employees have shared their enthusiasm for the release on social media. Robby Stein from Google noted that the model offers deeper reasoning and better instruction following. He expressed his excitement for more people to try the tool after using it internally to solve difficult problems.
Improved intent and reasoning
One of the primary goals of Gemini 4 Carbon is a better understanding of user intent. Rajan Patel from Google explained that this improvement allows for deeper back-and-forth conversations in AI Mode. Instead of the AI simply providing a static answer, it can better grasp the underlying goal of a user's query.
This focus on intent is crucial for complex interactions where a user might need to refine their request over several turns. By understanding what the user actually wants to achieve, the model can provide more relevant and accurate information. It reduces the need for the user to perfectly phrase every single prompt to get a usable result.
The model's ability to follow instructions more accurately is another key feature. This means that when a user gives specific constraints or multi-part directions, the AI is less likely to skip over details. This reliability is what makes it suitable for professional environments like software engineering and scientific research.
Real-world performance
Early testing of Gemini 4 Carbon has shown it to be quite powerful in practical applications. One tester reported using the model to build a new version of a classic Atari 2600 game from a single prompt. The model completed the task in just a few minutes and produced a working result.
This example illustrates the model's ability to handle complex coding and logic tasks. While the user only provided one prompt, the model had to interpret the requirements, write the necessary code, and ensure the final product functioned correctly. This demonstrates the 'long-horizon' capabilities mentioned by Google during the launch.
The model was released to cyber defenders two weeks before becoming available to the wider paid API and AI Ultra user base. This phased rollout suggests that Google is prioritizing high-stakes environments where security and precision are paramount. Now, those same capabilities are being opened up to the general subscriber audience.
What comes next for Search
The integration of Gemini 4 Carbon into Google Search marks a shift toward a more agentic search experience. As the model becomes more proficient at understanding intent, search results will likely become more conversational. Users may find themselves interacting with the search bar as a collaborator rather than just a directory.
Google is also previewing a swarm feature in Google Antigravity. A swarm feature typically involves multiple AI agents working together to solve a problem. This could eventually influence how Gemini 4 Carbon handles even larger and more complex tasks within the Search AI Mode.
For the average subscriber, this means the search tool is becoming more capable of handling 'tough questions.' As the model continues to be refined, the boundary between simple information retrieval and complex problem-solving will continue to blur in the Google ecosystem.
What to watch
- How the 'swarm' feature will eventually integrate into the standard Search AI Mode.
- Whether the model's cybersecurity capabilities will be released as a standalone tool for non-subscribers.
Try it
If you are a Google AI Ultra subscriber, check the selector in Google Search AI Mode to see if Gemini 4 Carbon is available in your region.
Sources and links
First reported by Search Engine Roundtable. 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.