Google Rolls Out Gemini 4 Carbon to the Public, Priced Per Million Tokens

Google is rolling out Gemini 4 Carbon to the general public, two weeks after the model first reached a select group of cybersecurity defenders through its Fairwind Program, according to The Verge.

Access is staged by product. Per the same report, Carbon is rolling out through the Gemini API in Google AI Studio and Gemini for business to all paying developers and enterprises. All Google AI Ultra subscribers will get it in the Gemini app, in AI Mode in Search, and on Antigravity, Google's agentic development platform, where AI agents carry out multi-step tasks with limited supervision.

Pricing is set at an introductory rate of $2 per million input tokens and $10 per million output tokens. A token is a small chunk of text, often a word or part of one, and billing counts both what you send the model (input) and what it returns (output). The Verge reports those figures for the launch.

Google positions Carbon as its strongest model to date. In its announcement, the company says Carbon delivers frontier performance in cybersecurity and in long-horizon agentic workflows, meaning tasks that run through many steps over a long stretch. Those span knowledge work, software engineering, and scientific reasoning. Google also says Gemini 4 Argon remains the better fit for most software engineering and knowledge work. Carbon, by Google's account, is built for complex multi-step coding and debugging and for open-ended scientific research.

The same announcement says Carbon consumes more tokens than Argon. It explores more options, reasons more deeply, and runs extended verification loops, in which it checks its own intermediate work repeatedly.

On benchmarks, Google states that Carbon ranks first on CursorBench 4.0, Perplexity's WANDR, and Zapier's AutomationBench, at more than 2X lower cost per task than the next best external model. Google also reports that it worked with Artificial Analysis to launch the Multimodal Knowledge Agent (MKA) benchmark, which measures multimodal performance in agentic workflows. On that benchmark, Google says Carbon ranks second, at more than 6X lower average cost per task than the top-performing model.

Separately, Google is previewing a swarm feature on Antigravity that turns a single prompt into a coordinated team of agents. The Verge describes it as a preview, so it is not yet a general release.

The broader context here is that the cost comparisons are expressed per task, not per token. Because Carbon generates more tokens per task than Argon, the per-task savings and the per-token list price will not map neatly onto one another. Teams building on the API will need to measure token consumption on their own workloads before extrapolating from either figure. The introductory pricing also carries a date horizon.

In my view, the more interesting question is what the tiered rollout reveals about where Google expects these models to be used. A model that spends more tokens per task makes most economic sense where one long chain of reasoning replaces several human review cycles, such as multi-step debugging or a literature-driven research workflow. The swarm preview points the same way, since a coordinated team of agents only pays off if each agent's output can be verified cheaply. If the verification loops Google describes can be made reliable at this cost, long, multi-step engineering tasks become something a developer can delegate and check, rather than something they must carry through every step.