Following last month's preview, Google is now introducing its latest Gemini 4 Carbon model into the Gemini app. AI Ultra subscribers can now access Carbon within the Gemini app, AI mode, and Antigravity. Notably, Argon (AI Pro and above) will continue to be available alongside the new model. Google officials state that Argon remains the ideal choice for most software engineering and knowledge work, while Carbon is designed for tasks that test the limits of model intelligence.
Examples of such tasks include complex multi-step coding and debugging, as well as open scientific research and discovery.
To showcase performance on larger, more complex projects, Google Antigravity is previewing /swarm. This will transform a "single prompt into a coordinated team of agents that can be tracked in real-time." Developers can assign Carbon to roles requiring deep reasoning, such as architectural design, coordination, and review, while pairing it with faster, lighter models in parallel independent workspaces, using adversarial verifiers to test each change. For legal, financial, and enterprise work, Carbon is believed to deliver "higher precision in knowledge work, artifact generation, and data analysis."
This includes how the model adheres to "strict patterns and formats to generate refined and accurate spreadsheets, presentations, and reports." Existing benchmarks fail to fully capture the reasoning capabilities of agents when handling dense documents containing charts, images, blueprints, schematics, and tables, prompting us to collaborate with an analytics firm to launch a new multimodal knowledge agent (MKA) benchmark.
Google Carbon model excels in multimodal performance testing
Carbon ranks first in the number of tokens used and, among 200 tasks testing the model's multimodal performance, uses three times fewer tokens than the next best model. Highlights in other areas include Carbon's outstanding performance in engineering tasks requiring multi-agent collaboration, such as multi-package migration, root cause investigation across subsystems, and complex architectural redesign. For web application development, it can execute build and validation cycles, recover from errors, and check its code and rendered user interfaces to generate fully functional applications.
Stronger visual design, spatial reasoning, and attention to environmental details also extend to 3D game development and creative coding.
Carbon's introductory pricing is US$2 per million input tokens (approximately HK$15) and US$10 per million output tokens (approximately HK$78), valid until November 30. After that, the price for input tokens will be US$4 per million (approximately HK$31), and output tokens will be US$20 per million (approximately HK$156). Carbon is also available for the following users: for developers: building with the paid tier of the Gemini API in AI Studio and Android Studio; if you are a Google AI Ultra subscriber, you can try it in Google Antigravity
and AI Studio. For enterprises: try it in Gemini for Business through API or model dropdown menus, Google Antigravity, or dedicated agents like CodeMender and AlphaEvolve. For scientific researchers: apply for a chance to gain free access to Argon and Carbon for six months (subject to eligibility, capacity, and rolling selection).
For cybersecurity defenders: apply to join the Fairwind program for cybersecurity access to Carbon.

