Gemini 4 Carbon: Google's new flagship model built for cybersecurity

What's new

Google unveiled Gemini 4 Carbon on October 28, 2026. It comes about two weeks after Gemini 4 Argon. Google calls Carbon its new frontier model, meaning one of the most advanced AI models on the market, with special strengths in cybersecurity and long-running work.

According to Google, the model can run for days, work its way out of dead ends and finish tasks that require many steps. For now, only a select group of cybersecurity defenders gets access through Google's Fairwind program while Google tests its safeguards. Access will then expand to more trusted partners. Developers, businesses and regular users will get access later. No date has been given, and the sources do not say whether the model will come to Denmark or the EU. AI Weekly reports that broader access is expected in mid-November, but Google has not confirmed this.

What's smart

Carbon is built to find and patch holes in software. According to Google, its success rate for writing secure code is more than double that of Argon, measured on Google's own internal tests. The model also produces fixes (patches) at more than ten times the rate of another lab's large frontier model. The competitor is not named.

Google's own figures: 88.2 percent on SEC-bench Pro (find vulnerabilities), 54.8 percent on ExploitGym (exploit them) and 71.0 percent on CWE-bench (fix them). Google also reports 0.0 percent successful cheating on Terminal-Bench 4.0 and 0.0 percent successful attacks in prompt injection tests, where hidden instructions try to trick an AI. According to Google, serious alerts that the model is not behaving as expected have dropped by almost 80 percent. The figures are Google's own, and independent verification is limited.

Cheaper or better?

The introductory price is $2 per million input tokens and $10 per million output tokens. Tokens are the chunks of text the model reads and writes. That is the same introductory price as Argon. Cached input, meaning text the model has seen before, gets a 95 percent discount. After the introductory period, the price rises to $4 for input and $20 for output per million tokens, according to AI Weekly. A monthly price for regular users has not been announced.

On the Artificial Analysis Intelligence Index v5, Carbon ranks first with 59 points, one point ahead of Anthropic's Claude Opus 5.5. OfficeChai has calculated that one task in the index costs $7.76 with Carbon versus $10.04 with Claude Opus 5.5, about 23 percent less. Carbon also uses about 123,000 output tokens per task versus 173,000, about 29 percent fewer.

It is not a clean win, though. On Artificial Analysis' Coding Agent Index, Carbon scored 59 points, three points behind Claude Code, and AI Weekly reports that independent tests show worse results than Claude Opus 5.5 on coding where the AI carries out tasks on its own. The hallucination rate, where the model makes up answers, is 11 percent on AA-Omniscience versus 4 percent for Argon. Accuracy is 58 percent.

What it's good for

Carbon is aimed first and foremost at cybersecurity professionals who need to find vulnerabilities, understand how they can be abused and fix them. It is also meant for long, complex engineering tasks that require the AI to work for hours or days.

For regular users and most businesses, the model is not yet available. The higher hallucination rate suggests that answers should be checked where factual accuracy matters. Competitors in the comparisons include OpenAI's GPT-6 Astra, Grok 4.7 and Kimi K3.1 Pro.