Google's Gemini 4 Carbon arrives two weeks after Argon — and cyber defenders get it first

Google DeepMind announced Gemini 4 Carbon on Wednesday, Oct. 28, 2026, two weeks after launching Gemini 4 Argon, with state-of-the-art claims on cybersecurity and long-horizon coding and a staged rollout that starts with cyber defenders.

Two frontier models in two weeks is the headline, but the rollout is the real story: Google is treating a model that is very good at finding and exploiting bugs as something you hand to defenders first and the public later. That's a sensible order of operations — and also a convenient way to ship benchmark wins without letting outsiders poke at the model yet. Every number here is Google's own, on Google's chosen tests; the independent check comes when developers get access. Meanwhile the introductory price quietly doubles later, so the cheap part is temporary too.

Google announced Gemini 4 Carbon on Oct. 28, 2026, and called it its strongest model, with state-of-the-art cybersecurity and long-horizon capabilities. It comes two weeks after Gemini 4 Argon. Those lines are Google's, from the DeepMind blog post the same day. The post is by Koray Kavukcuoglu, senior vice president of Google DeepMind and Google's chief AI architect.

Carbon ranks number 1 on version 5.0 of the Artificial Analysis Intelligence Index, with a score of 59. Google's table puts Argon at 50, GPT-6 Astra at 57, and Claude Opus 5.5 at 58. The Artificial Analysis Intelligence Index is an outside scoreboard that blends several tests into one number. Those scores are Google's, from the table in the post.

On security tests, Google's numbers, run without guardrails, are 88.2 percent on SEC-bench Pro for finding vulnerabilities, 54.8 percent on ExploitGym for exploitation, and 71.0 percent on CWE-bench for patching. A vulnerability is a hole in software someone can attack. Exploitation means actually using that hole. Patching is writing the fix. Guardrails are the limits a lab usually puts on a model so it refuses dangerous requests. Google's chart puts GPT-6 Astra at 85.4 percent and Argon at 83.4 percent on SEC-bench Pro. Those figures are Google's.

Google says Carbon has more than twice the success rate of Argon at generating secure code on internal benchmarks. The Chrome team reported it performed more than 10 times better than "another lab's large frontier model" at generating high-quality patches. A frontier model is one of the most capable models a lab has shipped. Google does not name the other lab. Those comparisons are Google's.

Google built Carbon for long agentic engineering work. Agentic, here, means the model keeps going on a multi-step job instead of answering once and stopping. Google says internal teams have seen Carbon run for days on tasks such as root-causing and patching subtle bugs across large codebases. Root-causing is finding the change that actually broke. It jumped 89 Elo points to number 4, a rating of 1768, on WebDev Arena, a web-development leaderboard. Elo is a score that rises when an entry beats others. Those lines are Google's.

On safety, Google says successful cheating on Terminal-Bench 4.0 fell to 0.0 percent, and cheat attempts fell to 0.4 percent. High-severity misalignment alerts in live internal coding dropped almost 80 percent. Misalignment, here, means the model doing something harmful or something the user did not ask for. Unsupported claims were cut by about twice versus Argon. Attack success was 0.0 percent on Gray Swan's indirect prompt-injection benchmark. Argon scored 0.7 percent on that test. An indirect prompt injection is hidden text, in a page or a file, that tries to hijack the model. Those figures are Google's.

Google says it is taking part in the U.S. government's voluntary pre-release testing process. Starting this week, Carbon goes to a subset of cyber defenders, then gradually to more trusted partners through Google's Fairwind program, before developers, businesses, and consumers "as soon as possible." Those lines are Google's.

Introductory pricing is $2 per million input tokens and $10 per million output tokens, the same as Argon. Cached input is 95 percent off. A token is a small chunk of text the model reads or writes. Cached input is text the service has already stored, so sending it again costs less. After the introductory period, the price is $4 per million input tokens and $20 per million output tokens. Those prices are Google's.

The picture is Google's SEC-bench Pro chart. Gemini 4 Carbon is at 88.2 percent, ahead of GPT-6 Astra at 85.4 percent and Gemini 4 Argon at 83.4 percent. The credit is Google.

In plain terms, Google on Wednesday announced Gemini 4 Carbon, two weeks after Argon, and called it the strongest model it has shipped. Its own table puts Carbon first on an outside intelligence index, and ahead on tests for finding bugs, exploiting them, and writing patches. A vetted set of cyber defenders gets it first. Developers and the public come later. The opening price matches Argon, then doubles.