TL;DR
What is Gemini 4 Carbon, and can I use it?
Google announced Gemini 4 Carbon on Wednesday, its strongest AI model, built for coding and security. Only a small group of cyber defenders can use it now. Artificial Analysis says it costs less per task than Claude's top model. Nothing to do today. Keep your current tools and watch for a public release.
What did Google announce?
On October 28, 2026, Google announced Gemini 4 Carbon, two weeks after it launched Gemini 4 Argon. Google calls Carbon its strongest model, aimed at "long horizon" work. That means big, multi-step jobs like tracking down a subtle bug across a whole codebase. Google says its own teams have seen Carbon run for days, pushing through dead ends on complex engineering tasks.
It also ranks #1 on the Artificial Analysis Intelligence Index v5.0, an outside scoreboard that compares AI models, with a score of 59.
Who can use Gemini 4 Carbon?
Almost nobody yet. Starting this week, Google is giving it to a subset of cyber defenders (people whose job is finding and fixing security holes) through its Fairwind Program. More trusted partners come next. Developers, businesses and regular users get it "as soon as possible," Google says, with no date.
How much will Gemini 4 Carbon cost?
| What Google says | |
|---|---|
| Intro API price | US$2 per million input tokens, US$10 per million output tokens |
| Cached input | 95% off the input price |
| Price after the intro period | US$4 input, US$20 output |
| Intro period ends | Not stated |
Tokens are the small chunks of text AI models read and write. Artificial Analysis says Carbon costs about 23% less per task than Claude Opus 5.5 at its top setting.
How good is it?
Google's own numbers lean hard on security: 88.2% on SEC-bench Pro (finding flaws), 54.8% on ExploitGym (testing them) and 71.0% on CWE-bench (patching them). Google says it writes secure code at more than twice Argon's success rate in internal tests, and blocked every attack in Gray Swan's prompt-injection test (tricks hidden in web pages or files that try to hijack an AI).
Two caveats from Artificial Analysis: Carbon ranks second to Claude Opus 5.5 on its coding-agent index (59 vs. 62), and it makes things up more often than Argon on its knowledge test (11% vs. 4%).
Our take
The model you can't use is still useful news, because of the price tag. Top-tier Google AI at US$2/US$10 puts pressure on everyone else, right after OpenAI made GPT-6.1 Terra cheaper. Whatever you pay for AI this year, the same quality will likely cost less next year.
The prompt-injection result matters most for small teams letting AI agents browse the web or read their email. Just don't plan around the intro price. Google says it doubles later and won't say when, and the higher error rate on facts means you still check its work.
— Artificially Nerdy
What should I do now?
- Don't switch tools. Carbon isn't available to you yet.
- Watch for Carbon in the Gemini API or the Gemini app. Google hasn't given a date.
- If you budget for API costs, plan with the US$4/US$20 price, not the intro one.
- Keep AI agents on a short leash, even ones that resist prompt injection well.
FAQ
Is Gemini 4 Carbon free?
No. Google has only announced API prices so far. It hasn't said which Gemini app plans will include it.
When will Gemini 4 Carbon be public?
Google hasn't said. It says developers, businesses and consumers will get it "as soon as possible" after it gathers feedback from cyber defenders.
Is Gemini 4 Carbon better than Claude?
It depends on the test. Artificial Analysis ranks it #1 overall, but second to Claude Opus 5.5 for coding agents.
Is Gemini 4 Carbon more accurate than Gemini 4 Argon?
Not on facts. On Artificial Analysis's knowledge test, Carbon made things up more often than Argon (11% vs. 4%).
Sources
- Google: Gemini 4 Carbon, our next frontier model: the announcement, rollout plan, prices, benchmarks and prompt-injection result.
- Artificial Analysis on X: the Intelligence Index and coding-agent rankings, cost per task and hallucination rates.
