Gemini 4 Carbon Launches as Google Shares 3D Generation Examples

Google DeepMind announced that Gemini 4 Carbon became available to developers, businesses, scientists and consumers on 10 November 2026. According to Google, the model is available to paid developers and businesses through the Google AI Studio API and Gemini for business; it is also rolling out gradually to Google AI Ultra subscribers through Google Antigravity, the Gemini app and AI Mode in Search. The announcement also included examples of the model being used for complex tasks and 3D generation.

How do access and pricing work?

Google describes Gemini 4 Carbon as its most capable model, with a particular focus on multi-step coding, debugging, scientific research and long-running agent workflows. The company says the model can use more tokens than Gemini 4 Argon to perform more extensive reasoning and verification.

The introductory pricing for the API and Gemini for business is the same as Argon: $2 per 1 million input tokens and $10 per 1 million output tokens. These are introductory prices stated in the announcement. The source does not specify separate pricing for app and service access for Google AI Ultra subscribers. It also provides no list of hardware requirements or compatibility with 3D software.

3D generation examples shared by Google

One example with a direct connection to architecture and visualization is an office model created using the /swarm feature in Google Antigravity. According to Google, an agent team used 206 engineering drawing views from seven municipal planning PDFs to create an interactive Three.js reconstruction of the Platform 37 office. The model contains 906,000 triangles, and the work was carried out in cycles over 10 nights.

Google also says it generated an interactive 3D exhibit showing the construction of the Eiffel Tower between 1887 and 1889, based on historical photographs and engineering drawings. Another example is a textile visualizer featuring interactive networks of yarn structures and stress-strain curves, created from an academic paper. These are examples presented in the company’s own product announcement; they are not presented as independent performance benchmarks.

What does this mean for architecture and archviz workflows?

The example of generating geometry from plans and drawings is notable for architecture firms and visualization teams because it points to the possibility of turning project documents into interactive 3D representations. Google says Carbon can assign agents different roles in Antigravity and take on tasks requiring deeper reasoning, such as architectural design, coordination and review. However, the announcement does not explain at what stages the example projects required human oversight or how accurate the output is for production projects.

For that reason, the model should be treated as a production tool to test, not as a replacement for existing modeling and verification steps. Geometry generated from plans should be checked for dimensional and drawing consistency; compatibility with the applications in use and required hardware should also be verified separately. Google’s announcement gives no technical requirements for these last two points.

Agent workflows and evaluation

Google describes /swarm as a shift from a single prompt to a team of agents that can be monitored in real time. According to the company, Carbon can work across different workspaces alongside faster, lighter models; verifier agents can also be used to test changes. This approach is interesting for dividing work, particularly in multi-step 3D generation and coding experiments.

Still, the examples shared do not show that the same results can be achieved across different project types. For firms, a practical evaluation should consider how readable the plans are, whether the resulting model is editable, how errors are detected and how API usage costs affect the workflow. Other results Google shared about the model’s performance and safety are also based on the company’s own evaluations.