Gemini 4 Carbon is Google’s newest and most capable AI model, and it arrived only two weeks after Gemini 4 Argon. Google announced it on 28 October 2026 and aims it at hard coding work and cybersecurity. There is a catch: for now, only a small group of security defenders can use it. This guide explains what the model is, what its benchmarks show, how much it costs and when ordinary users may get it.
⚡ Quick facts
- Announced: 28 October 2026
- Maker: Google DeepMind
- Focus: hard coding tasks and cybersecurity
- Access: limited to the Fairwind Program for now
- Intro price: $2 input / $10 output per million tokens
What is Gemini 4 Carbon?
This frontier model from Google DeepMind succeeds Gemini 4 Argon, which Google announced on 30 September and opened more widely on 13 October. Google has not shared the number of parameters, and it has not published a technical report. It also has not said whether the model uses a new architecture.
Who can use Gemini 4 Carbon right now?
Google limits access to Gemini 4 Carbon at launch. Google first gives it to selected cybersecurity defenders through its Fairwind Program, where organisations use it to find and fix security flaws. A broader release will follow, Google says, as soon as possible, after it collects feedback and improves the guardrails. Artificial Analysis lists no public API yet, and Google has given no date for wider availability through Antigravity.
Gemini 4 Carbon benchmarks in simple words
A benchmark is a standard test that scores how well an AI model does a task. Google reports that Carbon posts the top result in 15 of 17 rows in its own comparison. On the Artificial Analysis Intelligence Index, it scores 59, ahead of Claude Opus 5.5 at 58, GPT-6 Astra at 57 and Gemini 4 Argon at 50. Remember that the Gemini 4 Carbon numbers are mostly company-reported, so independent tests will matter.
| Benchmark | Argon | Carbon ✅ |
|---|---|---|
| Terminal-Bench 4.0 | 57.1% | 65.2% |
| FrontierSWE 2 | 55.0% | 61.2% |
| Agents’ Last Exam v1.1 | 69.0% | 77.4% |
| OSWorld-2.1 (offline subset) | 74.7% | 81.5% |
| GraphWalks | 84.2% | 96.2% |
| ExploitGym | 42.0% | 54.8% |
| SEC-bench Pro | 83.4% | 88.2% |
Gemini 4 Carbon does not win everywhere. GPT-6 Astra and Claude Opus 5.5 lead on FrontierSWE 2, and Opus leads the OSWorld offline subset at 87.2%. On the Artificial Analysis Coding Agent Index, Claude Code with Opus 5.5 scores 62, while Carbon scores 59.
Gemini 4 Carbon cybersecurity upgrades
Google built the model around three security jobs: finding vulnerabilities, checking that an exploit really works, and writing a patch. The biggest gains over Argon came on ExploitGym (up 12.8 points), SEC-bench Pro (up 4.8) and CWE-bench (up 3.0). The company also says Carbon generates secure code more than twice as well as before, and the Chrome team reports that its patches beat a rival model’s by more than ten times. Nobody outside Google has seen the data behind those claims, and it ran the external security tests without guardrails, so real results may differ.
Safety and reliability claims
- Gemini 4 Carbon cut unsupported claims; they fell by roughly half in Google’s internal tests.
- Severe alignment alerts dropped almost 80% in internal coding environments.
- On Terminal-Bench 4.0, Carbon attempted to cheat on 0.4% of tasks, and none of those attempts worked.
- In the Gray Swan prompt-injection test, attacks succeeded 0.0% of the time, compared with 0.7% for Argon.
⚠️ Watch out: one number points the other way. On the AA-Omniscience test, Gemini 4 Carbon answers 58% of questions correctly, but its hallucination rate rose from 4% to 11%. In other words, it answers more often, and it also invents more when it does not know.
How much does Gemini 4 Carbon cost?
Google lists introductory prices of $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached input. Later, the price rises to $4 for input and $20 for output. Google has not said when the introductory period ends. Since only Fairwind members can use the model today, most people cannot buy it yet.
What Gemini 4 Carbon means for you
If you are a student, a developer or a website owner, you will not use Gemini 4 Carbon today. Still, it shows where AI is heading: models that find security bugs and write fixes faster than people can. That is good for defenders, and it also raises the pressure to patch quickly, as with the recent Arista VeloCloud zero-day.
Key takeaways
- Google announced Gemini 4 Carbon on 28 October 2026, two weeks after Argon.
- It leads most of Google’s own benchmarks, but rivals win a few.
- Only cybersecurity defenders in the Fairwind Program can use it for now.
- Introductory pricing is $2 input and $10 output per million tokens.
- The hallucination rate on one test rose to 11%, so verify its answers.
FAQ
Can I use Gemini 4 Carbon today? Not yet, unless you belong to Google’s Fairwind Program. Google plans a wider release but has given no date.
Is it better than Claude Opus 5.5 and GPT-6 Astra? It leads the Artificial Analysis Intelligence Index, but each rival wins some tests, so the best model depends on your task.
What changed from Argon? Carbon scores higher on coding, agent and security tests, and Google says it has stronger guardrails.
Related reading: our earlier guide to Google’s Gemini 4 Argon and our look at AI incidents and safety.
