Google is rolling out its new Gemini 4 Carbon model to select users of the Gemini app. Designed to work alongside the existing Argon model, Carbon is aimed at those tackling AI Ultra tasks—think complex coding, scientific research, and cross-file reasoning—while Argon will remain the go-to for standard knowledge work and software engineering.
What Carbon Brings to Gemini
Carbon is built for higher-order intelligence. It shines when dealing with multi-step coding and debugging, processing dense materials like blueprints or schematics, and formats that require high fidelity—such as precise spreadsheets and formatting-heavy reports. For visual and spatial tasks, Carbon handles architectural design, UI validation, and even 3D game development with improved attention to detail. To back up those claims, Google teamed with Artificial Analysis to run a 200-task benchmark called Multimodal Knowledge Agents (MKA), where Carbon led the field while using just a third of the tokens of its nearest rival.
Where & How You Can Use Carbon
Carbon is available to AI Ultra subscribers in the Gemini app, including in AI Mode and Google Antigravity, the company’s experimental space for agent-driven experiences. At the same time, Argon will continue to be available under AI Pro and higher plans. Enterprises and developers can access Carbon through the Gemini API or model selections in AI Studio and Android Studio, while researchers who meet eligibility criteria may apply for trial periods of free access through Antigravity.
Cost & Early Access Details
During the introductory period through November 30, Carbon is priced at $2 per million input tokens, and $10 per million output tokens. After that, rates double to $4 for inputs and $20 for outputs per million tokens. Users in legal, finance, research, and cybersecurity—among other fields—are specifically being targeted for access, with special programs and application-based pathways set up to ensure they can try Carbon also via the Antigravity platform or enterprise agreements.
Launching a new AI model isn’t just about raw performance—it’s also about fitting real-world needs. Carbon’s strengths with multimodal tasks and stricter formatting suggest Google is aiming to make Gemini not just smarter, but more usable in professional workflows.
Why this matters: As tasks get more complex—like handling multiple documents, visual data, or strict output specs—the limitations of general-purpose AI models become clearer. Carbon’s benchmark-leading results on the MKA tasks signal that Google is leaning into models that excel under specificity and structure. What to watch next: how Carbon fares in actual enterprise use, especially in legal, finance, and scientific teams, under the real pressures of deadlines and precision.
