A model can top a leaderboard and still be out of reach for almost everyone. That is the situation with Gemini 4 Carbon, which Google announced on 28 October 2026 as its new frontier model. If you are wondering whether to switch tools because of it, the honest first step is to look at what Google actually published and who can use the model today.
What is Gemini 4 Carbon?
Google describes Carbon as its strongest model for cybersecurity and for long-horizon work. “Long-horizon” here means agentic tasks: the model plans and carries out many steps on its own, such as a multi-stage engineering job, instead of answering a single prompt. Carbon builds on Gemini 4 Argon, which Google released two weeks earlier. The announcement was written by Koray Kavukcuoglu, Chief AI Architect at Google.
What do Google’s numbers say?
Google reports first place on the Artificial Analysis Intelligence Index v5.0, an aggregate score built from many tests, with 59 points. On security tests it lists 88.2 percent on SEC-bench Pro (finding vulnerabilities), 54.8 percent on ExploitGym (exploiting them) and 71.0 percent on CWE-bench (patching them). On WebDev Arena, a web development ranking, Carbon sits fourth with 1768 Elo. Google also says the Chrome team found Carbon more than ten times better at writing high-quality patches than “another lab’s” large model, without naming that lab.
One point matters for comparisons you may have heard: the post does not mention Claude Opus 5.5 or any other named rival with numbers. Claims that Carbon “beats Opus 5.5” are therefore not something Google itself states.
Who can use it, and when?
Right now only a subset of cyber defenders has access. Google plans to widen this step by step to trusted partners through its Fairwind Program. For developers, businesses and consumers, Google only says access is coming “as soon as possible” after more work on guardrails. The post gives no date and does not confirm availability in the Gemini app or the public API.
What will it cost?
The introductory price is 2 dollars per million input tokens and 10 dollars per million output tokens, with cached input 95 percent cheaper. After the introductory period, Google lists 4 and 20 dollars.
What does Google say about safety?
According to Google’s internal tests, Carbon’s cheating rate on Terminal-Bench 4.0 fell to 0.0 percent, and the success rate of indirect prompt injection attacks (hidden instructions planted in content the model reads) was 0.0 percent on the Gray Swan benchmark, against 0.7 percent for Argon.
Our assessment
The facts: Carbon is announced, priced and ranked by Google, but not open to most users. Our reading: for everyday work nothing changes yet, because you cannot choose a model you cannot access. Start with the job you need done. If that job is security research or long agent runs, Carbon is worth watching once access opens; until then, the self-reported scores are a signal to test later, not a reason to switch today.
