Gemini 4 Carbon Announced | Top Model's Numbers and Availability
機械翻訳 / Machine-translated
A
AI Friends
6日前・16分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
If your reaction to yet another new AI release is "What's actually different from the last one?"— you're not alone. Gemini 4 Carbon is the newest model, arriving roughly two weeks after Argon, and it leads with strength in the security domain. This article walks through the numbers, availability, and how it stacks up against the competition.
Google announced Gemini 4 Carbon on its official blog on October 28, 2026.
Google positions this model as its most powerful to date, combining state-of-the-art cybersecurity capabilities with long-horizon operational ability. It succeeds Gemini 4 Argon, which launched two weeks prior.
Long-horizon operational ability refers to the capacity to autonomously continue working over the course of days — for example, tracking down the root cause of a bug in a large codebase and applying a fix (a corrective patch/diff).
In Japan, GIGAZINE also covered the announcement on October 29. According to that article, the models used for comparison were Gemini 4 Argon, GPT-6 Astra, Claude Opus 5.5, and Claude Fable 5.1. Carbon recorded the highest scores on most tests.
The key figures published on the official blog are as follows:
A vulnerability is a weakness hidden within software. Finding and fixing it before an attacker does can prevent damage.
According to Google, the Chrome team's reporting showed that Carbon's ability to produce high-quality patches was more than 10 times better than other large models. In internal evaluations, the rate at which it wrote secure code was more than twice that of Argon.
On the Artificial Analysis Intelligence Index v5.0, which measures overall capability, Carbon ranked 1st with a score of 59. For reference, Argon reportedly scored 53 on the same index.
In WebDev Arena (a competitive evaluation for building web apps), Carbon ranked 4th with an Elo of 1768 — an 89-point improvement over the previous generation. It's worth keeping in mind that it did not take first place here.
On the Coding Agent Index, Carbon recorded a score of 59 when combined with Google Antigravity. According to GIGAZINE, this places it alongside the Claude Sonnet 5 + Claude Code combination and the Muse Wonder Full + Muse Code combination.
More powerful AI also means greater potential impact if misused. Google published safety-related figures as well.
On Terminal-Bench 4.0, the success rate for the AI cheating to gain points was 0.0%. The rate at which it even attempted to cheat dropped to 0.4%.
On Gray Swan's IPI benchmark — which measures prompt injection attacks (feeding the AI hidden instructions to manipulate it) — the attack success rate was 0.0%, down from 0.7% for Argon.
In Google's internal live coding environment, critical alignment warnings dropped by approximately 80%. Unsubstantiated claims also fell by roughly half compared to Argon.
At this time, access is restricted to a select group of cyber defense professionals. The next phase will expand availability to trusted partners through the Fairwind Program.
For developers, enterprises, and general users, Google says it plans to provide access "as soon as possible," though no specific date has been given. Carbon is also participating in the U.S. government's voluntary pre-deployment testing.
According to the official Fairwind Program page, eligible parties include government agencies and critical infrastructure operators. Applicants undergo background checks and must agree to requirements such as phishing-resistant multi-factor authentication. Reselling or sharing access is prohibited. Note that the content on this page currently refers to Argon.
During the introductory period, pricing is $2 per million input tokens (approx. ¥316) and $10 per million output tokens (approx. ¥1,578). Cached input receives a 95% discount.
Once the introductory period ends, prices double to $4 per million input tokens (approx. ¥631) and $20 per million output tokens (approx. ¥3,156). The available materials do not specify when Carbon's introductory period will end.
For Argon, computingforgeeks reports that the introductory price runs through November 30, with the standard price taking effect December 1. Carbon may follow a similar timeline, but this has not been confirmed.
Argon became generally available on October 13 and supports context windows of up to one million tokens. According to the same article, on Terminal-Bench 4.0, Claude Opus 5.5 scored 66.4% versus Argon's 57.4%, putting Opus 5.5 ahead.
How much of that gap does Carbon close? Google's official blog does not include Terminal-Bench 4.0 standard scores — only the cheating success rate is shown — making a direct comparison with Opus 5.5 impossible from this data.
In terms of pricing, both Argon and Carbon share the same introductory rates of $2 (input) and $10 (output). In other words, upgrading to the higher-tier model does not change your costs.
Whether access is available from Japan is not addressed in the materials reviewed. Since the initial rollout targets cyber defense professionals, it will likely be some time before general developers can try it.
Imagine a corporate security professional tasked with identifying vulnerable code across millions of lines of source code. A review that would traditionally take weeks might now be completed by AI in a matter of days.
Here's another scenario: a local government systems administrator uses AI to generate patches for legacy software, with a human doing a final review. This kind of assistance could be a lifeline for teams stretched thin on personnel.
At the same time, there is legitimate concern about bad actors misusing the same capabilities. Google's decision to start with a limited rollout appears to reflect exactly that concern. The yen amounts are approximate estimates based on the exchange rate used by GIGAZINE (approx. ¥158 per dollar).
Q. Can I use Gemini 4 Carbon right now?
A. Not for general use. It is currently available only to a limited group of cyber defense professionals.
Q. What's the difference between Argon and Carbon?
A. Carbon is the successor, with particular strengths in vulnerability discovery and remediation, as well as long-horizon agentic tasks. Safety metrics have also improved.
Q. Is it truly more capable than competing models?
A. Google states that Carbon leads on many benchmarks. However, since much of the data comes from Google's own published figures, it's advisable to wait for independent third-party verification.
Q. How much does it cost?
A. During the introductory period, pricing is $2 for input and $10 for output (per million tokens). After the period ends, it rises to $4 and $20.
As a next step, check with your organization's security team to find out whether you may qualify to apply for the Fairwind Program.
This article is a cross-post from AI Friends.
