Google on Wednesday released Gemini 4 Carbon, which now sits at No. 1 on a widely followed independent AI leaderboard, but almost nobody can use it: the first users are a hand-picked group of cyber defenders.

Google DeepMind announced Carbon with a score of 59 on the Artificial Analysis Intelligence Index, one point ahead of Anthropic's Claude Opus 5.5. The model is especially strong at finding and fixing software security holes, and that is exactly why Google is holding it back.

The short answer: Gemini 4 Carbon is Google's most capable model, priced at $2 per million input tokens and $10 per million output tokens for now. Only vetted security teams in Google's Fairwind Program get it this week. Developers, businesses and consumers come later, with no date given.

What is Gemini 4 Carbon?

Carbon is built on Gemini 4 Argon, the model Google made generally available on October 13. Google says that inside the company, Carbon has run "for days" on large engineering tasks.

The headline benchmarks are all about security. Google reports 88.2% on SEC-bench Pro (a test of finding vulnerabilities in real code), 54.8% on ExploitGym (a test of actually exploiting them) and 71.0% on Collinear's CWE-bench (a test of writing the patch). Google's Chrome team says Carbon's patches were "more than 10X better than another lab's large frontier model."

It is not only a security model. Carbon gained 89 Elo points (a rating system borrowed from chess) on WebDev Arena to reach No. 4 with 1768, and Artificial Analysis ranks it second only to Claude Opus 5.5 on its Coding Agent Index.

Why is Gemini 4 Carbon gated?

A model that can find a bug, exploit it and patch it is useful to both sides. Google's bet is to hand it to defenders first so they can fix holes before attackers get the same capability.

Google says it is actively engaged in the US government's voluntary process for pre-release model testing. Access starts with a subset of members of the Fairwind Program, Google DeepMind's cyber defense partnership, which has more than 650 partners. Governments, critical infrastructure operators and core tech platforms get priority.

The bar to join is high. Fairwind requires background checks and phishing-resistant multifactor authentication, limits access to internal security, incident response and penetration testing teams, and bars partners from reselling access.

Google also published safety numbers meant to show the model behaves. On Terminal-Bench 4.0, Carbon cheated successfully 0.0% of the time and tried to cheat 0.4% of the time. Its attack success rate on Gray Swan's indirect prompt injection test (hidden instructions planted in content the model reads) was 0.0%, versus 0.7% for Argon. In live coding work with its own teams, Google says high-severity misalignment alerts fell by almost 80%.

When will Gemini 4 Carbon be available to developers?

Google's only commitment is "as soon as possible," after a gradual expansion through Fairwind. There is no date.

The closest guide is what happened last time. Google introduced Argon on September 30 as a Fairwind-only model. Thirteen days later, on October 13, Argon reached the paid Gemini API, AI Studio and the Gemini app for Ultra and Pro subscribers, according to Google's launch post and the Gemini API changelog.

That is a precedent, not a promise. Carbon is a stronger model at offensive security work, and Google may keep it with defenders longer. If you build on Gemini, plan around Argon today.

How much does Gemini 4 Carbon cost?

Google says Carbon will launch at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens, with cached input (text the model has already processed and can reuse) 95% cheaper. The standard price after that is $4 and $20. Google's post does not say when the introductory period ends.

Argon gives a hint. Its introductory price runs until November 30, and the $4 and $20 rate starts December 1, per Google's API pricing page. Artificial Analysis also notes there is no confirmed end date for Carbon's 50% launch discount, so its charts use the standard price.

Per-token prices only tell part of the story, because smarter models often use fewer tokens per job. On the Artificial Analysis leaderboard as of Thursday morning, at standard prices:

  • Gemini 4 Carbon (high): index score 59, $7.76 to run the index
  • Claude Opus 5.5 (max): score 58, $10.04
  • GPT-6 Astra (max): score 57, $6.24
  • Gemini 4 Argon (high): score 50, $7.68

So Carbon costs about the same per task as Argon while scoring nine points higher, and it is cheaper than Opus 5.5 at its top setting. For contrast, OpenAI's GPT-6.1 Terra lists at $1 and $5 per million tokens. You can run your own numbers in our AI API cost calculator or line up models in the AI model comparison tool.

What does Gemini 4 Carbon mean for the AI race?

Google now holds the top overall score, Anthropic still leads on coding agents, and OpenAI's cheaper Terra competes on price. The top of the index is tightly bunched, with only two points separating Carbon from GPT-6 Astra.

The bigger shift is the release pattern. Google has now shipped its two newest models to security teams first and the public second. For anyone tracking the labs as businesses, including Anthropic ahead of a possible IPO, being No. 1 matters less than when customers can actually pay for it.

Frequently asked questions

Can I use Gemini 4 Carbon right now?

Probably not. This week it is limited to a subset of Fairwind Program partners, with priority for governments, critical infrastructure operators and core tech platforms. Developers, businesses and consumers get it later.

What is the Gemini 4 Carbon release date for developers?

Google has not given one. Argon went from Fairwind-only to the public API in 13 days, but Google has not said Carbon will follow the same timeline.

How much does Gemini 4 Carbon cost?

$2 per million input tokens and $10 per million output tokens during the introductory period, rising to $4 and $20 after. Google has not said when the introductory price ends.

Is Gemini 4 Carbon better than Claude Opus 5.5?

On the Artificial Analysis Intelligence Index, yes, by one point (59 to 58). On the Coding Agent Index, Opus 5.5 still ranks first and Carbon second.

How do I get access to Gemini 4 Carbon?

For now, only through Google DeepMind's Fairwind Program, which requires background checks, phishing-resistant multifactor authentication and use by internal security teams only.