Technology

Google's Newest AI Model, Gemini 4 Carbon, Is Now Open to the Public

Martin HollowayPublished 6d ago3 min readBased on 2 sources
Reading level
Google's Newest AI Model, Gemini 4 Carbon, Is Now Open to the Public
source:blog.google

Google is releasing its Gemini 4 Carbon AI model to the general public. Two weeks earlier, it reached a small group of cybersecurity defenders through the company's Fairwind Program, according to The Verge.

Who gets access depends on the product they use. Carbon is rolling out through the Gemini API in Google AI Studio, and through Gemini for business, to all paying developers and enterprises. The API is a way for software to send requests to an AI model.

All Google AI Ultra subscribers will get Carbon in the Gemini app, in AI Mode in Search, and on Antigravity, Google's platform for AI agents. An agent is a program that works through a series of steps toward a goal.

Pricing starts at an introductory rate of $2 per million input tokens and $10 per million output tokens. A token is a small piece of text, often part of a word. Input tokens are what you send to the model, and output tokens are what it writes back. The Verge reports those figures for the launch.

Google calls Carbon its strongest model so far. In its announcement, the company says Carbon performs at the top level in cybersecurity. It also says Carbon handles long, multi-step tasks in knowledge work, software engineering, and scientific reasoning.

Google says Gemini 4 Argon is still the better choice for most software engineering and knowledge work. Carbon, according to Google, is built for complex coding and bug fixing, and for open-ended scientific research.

The same announcement says Carbon uses more tokens than Argon. Google says this is because Carbon explores more options, reasons more deeply, and runs longer checking loops.

Google reports benchmark results as well. A benchmark is a standard test that lets different models be compared. Google says Carbon ranks first on CursorBench 4.0, Perplexity's WANDR, and Zapier's AutomationBench. On those tests, Google says Carbon costs less than half as much per task as the next best outside model.

Google also worked with Artificial Analysis to launch a new test called the Multimodal Knowledge Agent (MKA) benchmark. It measures how well AI handles mixed media, such as text and images, during multi-step work. On MKA, Google says Carbon ranks second. It costs less than one-sixth of the average per-task cost of the top-ranked model.

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

The broader context here is that the cost comparisons are made per task, not per token. Because Carbon uses more tokens per task than Argon, the savings per task and the price per token will not line up neatly. Anyone building with the API will need to measure their own usage before relying on either number. The introductory pricing is also temporary, so the rate is set to change later.

In my view, the more interesting question is what the staged rollout suggests about where Google expects these models to be useful. A model that spends more tokens on each task makes the most sense when one long chain of reasoning can replace several rounds of human review. Tracking down a bug that spans many files is one example. The swarm preview points the same way, because a team of agents is only worthwhile if each agent's work can be checked cheaply. If the checking loops Google describes can be made reliable at this price, long engineering projects could become tasks a developer hands off and verifies, rather than tasks they must guide through every step.