Gemini 4 Carbon: Google’s new model for coding and security

Announced two weeks after Argon, it starts with cyber defenders and reaches everyone later.

Google has announced Gemini 4 Carbon, its newest frontier model, only about two weeks after Gemini 4 Argon. Google says it is much stronger at coding and at finding and fixing security flaws. For now only selected cyber defenders can use it, and wider access comes later.

Key facts

Google announced Gemini 4 Carbon on October 28, 2026, about two weeks after Gemini 4 Argon.
Google reports significant gains in coding and agentic tasks, plus better cyber defense work such as finding and patching vulnerabilities.
Access starts with selected cybersecurity defense companies for pre release testing. Google says it is coming soon to developers, businesses and consumers.
Most benchmark results are reported by Google itself and were not independently checked in the coverage we reviewed.

What happened

Google introduced Gemini 4 Carbon as its new frontier model. It follows Gemini 4 Argon, which arrived on October 13, and Google says the short gap reflects how closely its researchers and models now work together.

According to 9to5Google, Google reports significant gains across coding and agentic tasks, and also on scientific tasks, full stack web development and multimodal understanding, based on feedback about Argon. The company says Carbon ranks first on the Artificial Analysis Intelligence Index v5.0 with a score of 59. Engadget adds that Claude Opus 5.5 and GPT-6 Astra scored just behind it.

Cybersecurity is where Google puts the most emphasis. The model is described as able to find and prioritize vulnerabilities and to patch them at the root cause. Google reports 88.2 percent on SEC-bench Pro, which measures finding vulnerabilities, 54.8 percent on ExploitGym and 71.0 percent on CWE-bench, which measures patching. Engadget notes that Google says Carbon beat rival models by fairly wide margins in this area, without giving exact figures.

Google also describes behavior changes. In its own testing, unsupported claims were cut by about half compared with Argon, and the model is said to resist prompt injection and to take shortcuts less often. Engadget reports that it is twice as likely as Argon to generate secure code.

On access, Carbon first goes to selected cyber defenders. Google plans to extend it to more trusted partners through its Fairwind Program, followed by a broad release. Argon followed the same path and is now available in the Gemini app and in AI Mode in Search. 9to5Google reports introductory pricing of $2 per million input tokens and $10 per million output tokens, rising later to $4 and $20.

Why it matters

Most people will not use Carbon this week. The model is not in the Gemini app yet, and Google has not given a date for broad release. Still, the announcement shows where the race between the big labs is heading: coding, long running agent work and security.

For everyday users the likely effect is indirect. Better coding models make the apps and tools you use easier and faster to build, and security focused models could help companies find flaws before attackers do. At the same time, a model that is good at finding vulnerabilities is also sensitive, which is probably why Google is starting with vetted defenders.

A careful reading helps here. The headline ranking and the security scores come from Google, and a model that leads one index can fall behind on another. Independent tests over the coming weeks will say more than the launch post.

How to try it

You cannot use Carbon yet, but you can practice the skill this story calls for: turning a dense technical announcement into something you can act on. The Try it prompt, Explain simply, rewrites complicated text in plain language.

1. Open the prompt and paste it into ChatGPT, Claude or Gemini. 2. Where it asks for the text, paste the availability section or the benchmark paragraph from the announcement. 3. Where it asks for your level, write something like “I use AI tools at work but I am not a developer”. 4. Ask for three things: a plain summary, the one term you should look up, and one question to ask your IT team.

For example, paste the sentence about SEC-bench Pro and ask what a benchmark like that measures and why a company would care. Check the answer against the original. If the reply states a number that is not in the text you pasted, ask the assistant to point to the sentence it came from.

Try it

Explain a topic simply

Frequently asked questions

Can I use Gemini 4 Carbon now?

Not in general. Google says access begins with selected cyber defenders, and that Carbon is coming soon to developers, businesses and consumers. No date for broad release has been given in the coverage we reviewed.

How is Carbon different from Gemini 4 Argon?

Google presents Carbon as a step up from Argon, with gains in coding, agentic tasks and security work, and with fewer unsupported claims in its own tests. Argon is already available in the Gemini app and in AI Mode in Search.

Are the benchmark scores reliable?

They are reported by Google, and the coverage we reviewed did not include independent confirmation of every figure. Treat them as a starting point and look for outside tests before choosing a model for important work.

Sources