If you have been waiting to see whether Google’s newest model is something you can actually open and use, the answer changed on 10 November 2026: Gemini 4 Carbon moved from a small cybersecurity preview to a broad release. Whether it is the right pick for your work is a different question. Start with the job you need done.
What Gemini 4 Carbon is, and how it relates to Argon
Google announced Carbon on 28 October 2026 and calls it its strongest model, with state-of-the-art cybersecurity and long-horizon capabilities. “Long horizon” means work that takes many steps over a long stretch of time, such as tracking down a bug in a large codebase and then patching it. Google says internal teams have seen Carbon run for days on tasks like that. Carbon builds on Gemini 4 Argon, which launched about two weeks earlier, and it uses the same multi-layered safety and security stack.
Who can use it today
The 10 November announcement opened access in several tiers. Google AI Ultra subscribers get Carbon in the Gemini app, in AI Mode in Search and in Google Antigravity, where Google says the rollout is still ongoing. Developers on the paid Gemini API tier can use it through Google AI Studio and Android Studio, and businesses can reach it through Gemini for business. Cybersecurity defenders get access to Carbon’s cyber capabilities through Google’s Fairwind Program on a rolling basis, and scientists can apply for six months of free access to Argon and Carbon through Antigravity’s highest usage tier, subject to eligibility and capacity. The practical consequence: if you are on a free or standard consumer plan, the announcement does not list you, so check your plan before planning around it.
What it costs, and when the price changes
For API users, the introductory price is $2 per million input tokens and $10 per million output tokens, the same as Argon’s introductory price. A token is a small chunk of text, roughly a short word or part of one, and providers bill by the million. That introductory price runs until 30 November 2026. After that the public Carbon model costs $4 per million input tokens and $20 per million output tokens, double the introductory rate. Google’s October post also lists a 95 percent discount on cached input, meaning text the model has already been given and can reuse.
My own reading, which is an inference and not something Google states: because the price doubles at the end of the month, any evaluation you plan should be finished and costed at the higher rate, not the introductory one. A project that looks affordable at $10 per million output tokens may not look the same at $20.
When to choose Carbon over Argon
Google is direct about this. It says Argon remains the better choice for most software engineering and knowledge work, while Carbon is meant for tasks that push a model’s reasoning to its limits, such as complex multi-step coding and debugging, and open-ended scientific research. Carbon also uses more tokens than Argon because it explores and verifies more thoroughly. Choose the fit, not just the size: a bigger model that spends more tokens on an easy task is simply a slower, pricier way to get the same answer. The sensible test is to run one hard task you already know well on both models and compare the result and the token bill side by side.
What the benchmark claims do and do not show
Google reports that Carbon ranks first on the Artificial Analysis Intelligence Index v5.0 with a score of 59, and first on CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench, at more than twice the cost efficiency per task of the next best external model. On cybersecurity, it reports 88.2 percent on SEC-bench Pro for finding vulnerabilities, 54.8 percent on ExploitGym for exploiting them and 71.0 percent on CWE-bench for patching them; Google notes these were evaluated without guardrails. It also says Carbon blocked 99 percent of attack trajectories in simulated cyberattack testing and passed 92 percent of tasks on a 375-task agentic honesty benchmark, versus 70 percent for Argon.
These numbers come from Google’s own announcements and have not been independently confirmed here. Rankings also depend on which tasks a benchmark contains, so treat them as a reason to test Carbon on your work, not as a verdict. Google says it is taking part in the US government’s voluntary pre-release model testing, which is useful context, but it is not an audit of these specific claims.
