# Gemini 4 Carbon takes the top of the leaderboard as a model only vetted cyber defenders can use

URL: https://www.thedeepfeed.ai/posts/2026-10-28-gemini-4-carbon-tops-index-few-can-use/
Category: Models
Published: 2026-10-28
Author: the-deep-feed
Tags: google, gemini-4-carbon, gemini, frontier-models, cybersecurity, staged-release, agents, benchmarks
Kind: news

> Google's Gemini 4 Carbon took first place on Artificial Analysis's index on October 28 with a score of 59, but it launched only to vetted cyber defenders through the Fairwind Program, repeating the staged release Google used for Argon four weeks earlier.

## TL;DR

- **Gemini 4 Carbon** scores **59** on Artificial Analysis's Intelligence Index v5, ahead of Claude Opus 5.5 at **58** and GPT-6 Astra at **57**, and costs **$7.76** per index task against Opus 5.5's **$10.04**.
- It launched on **October 28** only to 'a subset of cyber defenders' through Google's **Fairwind Program**, which says it works with **650+** partners vetted with background checks and phishing-resistant MFA.
- Google quotes an introductory **$2/$10** per million tokens. Artificial Analysis lists a standard **$4/$20** with a **50%** discount, and there is no public API date.
- This is the second time running: **Gemini 4 Argon** went to cyber defenders on Sep 30 and reached paid app users on Oct 13. Google says Carbon's benchmarks were 'evaluated without guardrails.'

Google's Gemini 4 Carbon became the top-scoring model on Artificial Analysis's index on October 28, and on the same day Google limited it to vetted cyber defenders. For the second time in a month, the model at the top of the leaderboard is one most developers cannot call.

Google's [launch post](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/) says that "Carbon ranks #1 on the recently updated v5.0 of Artificial Analysis Intelligence Index at 59." Artificial Analysis [confirmed the score](https://artificialanalysis.ai/articles/gemini-4-carbon-new-number-1) the same day with the model at high reasoning effort. [VentureBeat](https://venturebeat.com/technology/google-unveils-gemini-4-carbon-retaking-the-worlds-no-1-spot-on-updated-harder-benchmarks-but-access-is-limited) reported the order behind it: "ahead of Claude Opus 5.5 at 58, GPT-6 Astra at 57 and Argon at 50."

# A one-point lead on a new index

The lead is narrow and the index is new. Artificial Analysis says Carbon takes the "#1 score in 4 of the 9 Index evaluations," and on coding agents is "Behind only Claude Opus 5.5," at 59 against Claude Code with Opus 5.5 at 62. Its cost argument is clearer: "$7.76 per Intelligence Index task, ~23% less than Claude Opus 5.5 (max, $10.04)."

This publication's [autumn scorecard](/posts/2026-10-15-autumn-frontier-scorecard/) found that once independent evaluators reran the September and October flagships, the frontier compressed into a statistical tie, leaving cost, tokens, speed and access as the measurable differences. One point on a freshly revised index does not break that tie. It does show Carbon's cost per task is lower, which is the claim worth testing once the model can be tested at all.

Google's own table makes broader claims. Carbon posts 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench, and VentureBeat counted "the highest result in 15 of the 17 benchmark rows in Google's announcement." For agents, Google says "Carbon is built for sustained agentic workflows. Internally, our teams have seen Carbon run for days," and that it "reduced successful cheating on Terminal-Bench 4.0 to 0.0%, with attempts to cheat dropping to 0.4%." One qualifier applies to all of these: the post says the benchmarks were "evaluated without guardrails," the safeguards Google says it is still iterating on before wider release.

# Who can use it

Access is the real headline. Google writes that it is "actively engaged in the U.S. government's voluntary process for pre-release model testing," and that "Starting this week, we will make Carbon available to a subset of cyber defenders and will gradually phase access to additional trusted partners through our Fairwind Program." Developers, businesses and consumers come later, "as soon as possible," with no date.

The [Fairwind Program page](https://deepmind.google/fairwind-program/) explains what "trusted" means. It "gives high-priority defenders (like governments, healthcare providers, and telecommunications services) early access to advanced models," and Google says it works "with over 650 partners globally." Partners "are only permitted to carry out dual-use tasks, including authorized threat simulation, reverse engineering, and malware analysis for defensive and academic research purposes." Google requires "user-level authentication, phishing-resistant MFA" and runs "background checks on organizations that apply." Entry goes through a Google Cloud account team. On the day of the Carbon launch, the Fairwind page still named Argon as the model partners receive.

Google's stated price is "$2 per million input tokens and $10 per million output tokens," described as introductory. Artificial Analysis lists the standard rate as "$4/$20 per 1M input/output tokens ... with a 50% discount," so the introductory price is half of list. [9to5Google](https://9to5google.com/2026/10/28/gemini-4-carbon-announcement/) summarized the rollout in one line: "The model will first be available to 'cyber defenders' this week."

# The staged release is now a pattern

Carbon follows its predecessor exactly. Gemini 4 Argon was announced on September 30, released first to cyber defenders, and reached paid Gemini app subscribers on October 13, as the October 15 scorecard recorded. Google's Carbon post says Argon "was launched two weeks ago." If Carbon follows the same 13-day gap, broader access would arrive in mid-November. Google has not committed to that.

The [Hacker News thread](https://news.ycombinator.com/item?id=50276792), which reached 1,006 points and 796 comments, was divided between the price and the wait:

> I'm not a big fan of their trend of talking about a new model and not releasing it for a while.
>
> — kilroy123, Hacker News, Oct 28, 2026

Another commenter called "$2/$10 for input and output even if just for introductory pricing" "an incredible deal."

For buyers building agents, the practical result is that the leaderboard and the price list now describe different models. The #1 model is available to a subset of vetted cyber defenders, with Fairwind's 650-plus partners next in line, and everyone else is choosing between Opus 5.5, Astra and Argon. Staged release lets Google claim the top spot and run security trials at the same time. It also means the first independent agent evaluations of Carbon, the ones that would test the runs lasting days and the 0.0% cheating rate, will come from defenders, not developers.

## Sources

- [Google — Gemini 4 Carbon: our next frontier model (Oct 28, 2026)](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/)
- [Google DeepMind — Fairwind Program](https://deepmind.google/fairwind-program/)
- [Artificial Analysis — Gemini 4 Carbon is the new #1 on the Artificial Analysis Intelligence Index v5 (Oct 28, 2026)](https://artificialanalysis.ai/articles/gemini-4-carbon-new-number-1)
- [VentureBeat — Google unveils Gemini 4 Carbon, retaking the world's no. 1 spot, but access is limited (Oct 28, 2026)](https://venturebeat.com/technology/google-unveils-gemini-4-carbon-retaking-the-worlds-no-1-spot-on-updated-harder-benchmarks-but-access-is-limited)
- [9to5Google — Google announces Gemini 4 Carbon (Oct 28, 2026)](https://9to5google.com/2026/10/28/gemini-4-carbon-announcement/)
- [Hacker News — Gemini 4 Carbon discussion (Oct 28, 2026)](https://news.ycombinator.com/item?id=50276792)

---

Canonical: https://www.thedeepfeed.ai/posts/2026-10-28-gemini-4-carbon-tops-index-few-can-use/
Site: https://www.thedeepfeed.ai
Full corpus: https://www.thedeepfeed.ai/llms-full.txt