Google is opening Gemini 4 Carbon to all users, extending beyond its original cybersecurity-defense use case. Its launch price is $2 per million input tokens and $10 per million output tokens, but both rates will double on December 1 to $4 and $20, respectively. The model has already entered real development tasks; pricing it low before raising rates shows that vendors are still using short-term discounts to win usage.
What this is
Gemini 4 Carbon is Google's next-generation model, offering low, medium, and high reasoning levels, with medium selected by default. A token is the unit into which text is divided for model processing, and billing largely depends on the length of the input and output.
Simon Willison used Carbon to create an animated New York city block based on Three.js. He then used Google Antigravity's swarm feature to have multiple subtasks collaborate on a 3D game: generating model assets, building the web scene, configuring graphics shaders, and adding procedural sound effects and Lyria music. Subtasks run in parallel, allowing one complex request to be broken into smaller tasks that are completed separately.
Industry view
Our judgment is that model competition is shifting from “who performs better on the leaderboard” to “who can enter a complete workflow.” Carbon no longer merely answers questions. It can work with a programming environment, map data, music tools, and browser technologies to deliver runnable results. For enterprises, systems like this are closer to usable software than one-off chat responses.
The risks, however, are equally clear. The Antigravity demo is a showcase, and whether complex projects can be reproduced reliably across different teams and codebases still requires more evidence from production environments. The increase from $10 to $20 per million output tokens also reminds users that low prices do not necessarily mean manageable long-term costs. Another issue is that model-generated applications may still contain missing dependencies, unclear copyright provenance, and runtime errors. Demo videos cannot be treated as proof of reliability.
Impact on regular people
For enterprise IT: procurement priorities will shift from a single model's benchmark score toward the development toolchain, stability in task execution, and total cost of ownership. Carbon's price-increase schedule is well suited to stress-testing budgets.
For individual professionals: the ability to describe requirements in natural language, inspect results, and break work into tasks is becoming a new collaboration skill. For designers and developers, the model increasingly resembles a workbench that can handle text, images, code, and sound simultaneously.
For the consumer market: model-generated websites, mini-games, and interactive content will continue to proliferate. Yet the quality gap will likely come more from process control and human acceptance testing than from which model was used.