On 10 November Google made its new top model, Gemini 4 Carbon, available to all paying developers and businesses through the Gemini API, Google AI Studio and Gemini for business. For an introductory period that expires on 30 November it costs the same as its sibling Gemini 4 Argon: 2 dollars per million input tokens and 10 dollars per million output tokens. After that the price doubles to 4 and 20 dollars.
For a small company thinking about building a service on top of the model, that date is the most important detail in the launch. Anyone who tests now and budgets at today's price ends up with numbers that only hold until the end of the month.
What Google says Carbon is for
Carbon was announced on 28 October and was at first available only to a small group of security teams in Google's Fairwind Program. Two weeks later it opened up more widely. In the announcement Google describes the model as built for tasks that "test the limits of a model's intelligence", such as complex multi-step coding, debugging and open-ended scientific research.
Just as interesting is what Google says about the older model. According to the company, Argon is still the best choice for the majority of software engineering and knowledge work. Carbon works more thoroughly and therefore uses more tokens than Argon to reason and check its own work. The same price per token does not mean the same price per task. A question Argon answers briefly can become a longer and more expensive run in Carbon.
Google also points out that Carbon follows strict schemas and formatting when it produces spreadsheets, presentations and reports, and that it ranks first on, among others, Zapier's AutomationBench at less than half the cost per task of the next best external model. These are Google's own figures on tests Google chose. They indicate a direction, but say nothing about how the model handles your quotes or your accounting export.
Who gets access, and where?
The way in depends on who you are. Developers reach Carbon through the paid tier of the Gemini API in AI Studio. Businesses can pick it in Gemini for business, either through the API or in the model drop-down. In the Gemini app and AI Mode in Google Search it is available to Google AI Ultra subscribers. Other subscriptions are not mentioned in the announcement.
At the same time Google is previewing a feature called /swarm in its development environment Google Antigravity. It turns a single prompt into a coordinated team of agents that you can follow while they work. The feature is aimed at larger programming projects.
More honest, but not infallible
Two figures in the launch deserve a closer look for anyone who lets a model handle customer requests. On Google's own 375-task benchmark for honesty in agentic work, Carbon passed 92 percent of tasks against 70 percent for Argon. According to Google, the model declined to fill in missing experimental data 98 percent of the time. That is exactly the behaviour you want when a request lacks a date or a floor area: a follow-up question rather than an invented detail. We showed what such a flow can look like in September with a simple prototype of our own.
The second figure concerns prompt injection, when text in an email or a document tries to give the model new instructions. In Gray Swan's test no such attack against Carbon succeeded. Yet in the same paragraph Google writes that robustness against prompt injection remains an area for continued improvement. That is an unusually frank caveat for a launch post, and reason enough to let a person approve everything that goes out to a customer.
What a small business can do before 30 November
If you already have an idea of what a model should do in your business, the introductory price makes for a cheap test. Take twenty or so real, anonymised examples of the task, run them through both Argon and Carbon and note how many tokens each run uses. Then work out the cost at the price that applies from December, not today's. If Argon handles the task just as well, the cheaper model is the one to choose, as Google itself suggests.
The test material does not have to be made up. If you want to see what ordinary requests look like when private individuals write them themselves, you can read plain-language job requests on a Swedish services marketplace, with location and description. That makes better test material than examples you invent yourself.
Think about repeated text as well. According to Google's October announcement, cached input tokens cost 95 percent less than regular ones. A flow that sends the same long instruction and price list before every request can therefore become considerably cheaper if it is built to reuse that part.
What automating customer replies costs in practice, with and without a language model, is covered in our article What does it cost to automate customer replies?. The conclusion there was that the model cost is rarely the problem. The item people tend to forget is maintenance, and that does not change because a new model becomes cheaper or more expensive.
Sources: Google, Gemini 4 Carbon is now available (10 November 2026) and Gemini 4 Carbon: our next frontier model (28 October 2026).