Gemini 4 Carbon: Google's new frontier model goes to cyber defenders first
Google announced Gemini 4 Carbon on October 28, two weeks after Gemini 4 Argon reached the Gemini app. Google calls it its strongest model, with the lead in cybersecurity and long tasks. For now only a group of cyber defenders can use it; developers, businesses and Gemini app users get no date.
What happened
Koray Kavukcuoglu, Google's chief AI architect, announced Gemini 4 Carbon on October 28, 2026. It sits above Gemini 4 Argon, which opened to the public on October 13, and Google pitches it at two jobs: finding and fixing security holes, and long tasks that run for many steps.
Access starts small. This week it goes to a subset of cyber defenders, and Google will widen it to more trusted partners through its Fairwind Program. Developers, businesses and people on Gemini plans get it "as soon as possible", with no date. Argon took the same route: announced for defenders on September 30, then opened to the Gemini app for AI Pro and Ultra and to the API on October 13.
The numbers Google gives
| Test | Gemini 4 Carbon | What it says |
|---|---|---|
| Artificial Analysis Intelligence Index v5.0 | 59, #1 | Top of the independent overall ranking |
| SEC-bench Pro | 88.2% | Fixing real security bugs; Google calls it state of the art |
| ExploitGym | 54.8% | Writing working exploits; Google calls it a new record |
| WebDev Arena | #4, 1768 Elo | Not first for web front ends |
| Gray Swan prompt injection, attack success | 0.0% | Argon was 0.7% |
Against Argon, Google reports more than twice Argon's success rate on writing secure code and about half as many unsupported claims. On Terminal-Bench 4.0, Google says Carbon gamed the test on 0.0% of tasks and tried to on 0.4%.
What it will cost
| Per 1M tokens | Input | Output |
|---|---|---|
| Introductory price | $2.00 | $10.00 |
| After the introductory period | $4.00 | $20.00 |
Cached input is 95% off, and Google doesn't say how long the introductory period lasts.
What it means for you
- Gemini app users: keep working with Argon.
- API builders: nothing to change yet. There's no public model name to call and no date.
Plans and the current lineup are in our Gemini review; prompts to try sit in the Gemini prompt collection.
Try it
Review the code change below for security problems only. Ignore style. For each problem: - the line or function, - what an attacker could do with it, in one sentence, - how sure you are (high / medium / low), - the smallest fix, as a code snippet. If you find nothing, say so and list what you checked. Don't invent a vulnerability to have something to report. Language and framework: [e.g. Python, Django] The change: [paste the diff]
I'm deciding whether to switch models. Here are three real tasks from my work and the answer my current model gave to each. For each task: 1) Do the task yourself. 2) Compare your answer with the one I pasted: what's better, what's worse, what's different but equal. 3) Rate both from 1 to 5 on correctness and on how much editing they'd need. Be fair to the old answer. If it's better, say so. Task 1: [task] / Current answer: [paste] Task 2: [task] / Current answer: [paste] Task 3: [task] / Current answer: [paste]
Sources
- Google, Gemini 4 Carbon: our next frontier model, October 28, 2026
- Google, Gemini 4 Argon announcement, September 30, 2026
- Google DeepMind, Fairwind Program, checked October 28, 2026