What Is Gemini 4 Carbon? Cybersecurity, Price, Access

Google announced Gemini 4 Carbon on October 28, 2026. In a post signed by Koray Kavukcuoglu of Google DeepMind, the model is presented as the company's "strongest model with state-of-the-art cybersecurity and long horizon capabilities." Carbon arrived about two weeks after Gemini 4 Argon and, according to Google, builds on Argon.

This post is a news roundup and compilation; I have not tested the model, and for now nobody outside a small group can. Carbon is open only to a selected group of cyber defenders. The numbers below come from Google's announcement and news sites; there is no independent reproduction. Sources are at the end. Access and the comparison are in separate posts: When Is Gemini 4 Carbon Coming? Who Can Use It? and the Carbon, Argon, Opus 5.5 and Astra comparison.

The news at a glance

Feature	Value
Model name	Gemini 4 Carbon
Announced	October 28, 2026
Developer	Google DeepMind
Google's description	Strongest model for cybersecurity and long-horizon work
Intro price	$2 / million input tokens, $10 / million output tokens (same as Argon)
Standard price	$4 input, $20 output
Cached input	95% off the input price
Access (now)	A selected group of cyber defenders
Access (next)	Additional trusted partners through the Fairwind Program
General availability	No date; Google says "as soon as possible"
Context window, output limit, model card	Not in the announcement

Google does not say how long the intro period lasts. For Argon's intro price, WinBuzzer reported November 30; no such date has been given for Carbon.

Relationship to Argon

Google says Carbon builds on Argon and uses the same multi-layered safety and security stack. It writes that the improvements draw on user feedback on Argon and target coding, agentic tasks, scientific tasks, full-stack web development and multimodal understanding.

There are no technical details: no parameter count and no technical report. Dealroom relays that Ars Technica wrote that Carbon could be a post-training upgrade of Argon rather than a new architecture. That is an interpretation; Google only says "builds on," and that wording is open to either reading.

Benchmark results

The numbers from Google's announcement and the news coverage based on it:

Benchmark	Gemini 4 Carbon	Note
Artificial Analysis Intelligence Index v5.0	59, first	Per Dealroom: Claude Opus 5.5 58, GPT-6 Astra 57, Argon 50
SEC-bench Pro (finding vulnerabilities)	88.2%	Google calls it state-of-the-art
ExploitGym (exploitation)	54.8%	Google calls it a new state-of-the-art
CWE-bench (patching, Collinear)	71.0%
WebDev Arena	1768 Elo, fourth	Up 89 points
FrontierSWE 2	61.2%	GPT-6 Astra ahead at 65.5%
OSWorld-2.1 (offline subset)	81.5%	Claude Opus 5.5 ahead at 87.2%

Three things to keep in mind when reading the table:

• The index was recently updated. Google says "the recently updated v5.0," so do not compare directly with Argon's ranking on older versions. Dealroom also relays that on Artificial Analysis's coding agent index Carbon scores 59, three points behind Claude Code running Claude Opus 5.5.
• The sources are mixed. The table combines scores Google computed itself, third-party leaderboards and rivals' self-reported values.
• The cyber benchmarks ran without guardrails. Google says these benchmarks were evaluated without guardrails. This is my interpretation: the version that eventually ships with guardrails need not produce the same results.

For Terminal-Bench 4.0, Google gives no score, but it says accuracy improved and that successful cheating on the benchmark dropped to 0.0% and cheating attempts to 0.4%. Terminal-Bench was one of the benchmarks where Argon trailed. In the Google table as relayed by Eden AI, Carbon scores 65.2% (63.1% in Vals's own run), clearly above Argon's 57.1%.

The claims Google highlights

All of the following come from Google's internal tests or its partners; none is independently verified.

Cybersecurity. Google's Chrome team reportedly found that Carbon was more than 10 times better at producing high-quality patches than another lab's large frontier model, which is not named. The security firm Wiz also says the model leads on web penetration-testing benchmarks.

Long-horizon work. According to Google, internal teams have watched Carbon run for days, working through dead ends to root-cause and patch subtle bugs in large codebases and low-level systems.

Behavior and safety. Internal measurements compared with Argon:

• Success rate for secure code generation more than doubled.
• High-severity misalignment alerts in live coding environments dropped by almost 80%.
• Unsupported claims roughly halved.
• Overreach fell 26% and under-service 29%.
• Attack success on the Gray Swan indirect prompt-injection benchmark is 0.0% (0.7% for Argon).

Google also says that on an ambiguous request Carbon retrieves sources and surfaces discrepancies instead of guessing, and that when blocked it lists what remains instead of making things up. That, too, is Google's own measurement, and it does not fully match the independent one, as I explain below.

What the independent measurements show

Third-party pages such as Artificial Analysis and Vals show the less-discussed sides of Carbon. I give the detail in the comparison post; in short:

• The hallucination rate went up. In Artificial Analysis's AA-Omniscience v1.1 measurement, Carbon has the highest net score (54), but its rate of giving a wrong answer to a question it does not know is 11%; for Argon it is 4%. Google's "unsupported claims roughly halved" is about internal measurement in coding environments; the two measure different things. (The roughly 15% in my Argon post belonged to the older version of the benchmark.)
• It was slow and wordy. 84.6 tokens per second, and 118.6 seconds to the first token on Google's API. It spent 200 million output tokens to complete the index; the median is 70 million.
• Not first in agentic coding. On Artificial Analysis's coding agent index, Claude Code with Opus 5.5 scores 62 and Carbon 59 (tested in a single harness).

Price

The intro price is the same as Argon's: $2 per million input tokens and $10 for output, with 95% off cached input. After the intro period it is $4 and $20. These are announced prices; because the model is not in the general API yet, you cannot use it at this price today. There is no price information for a consumer subscription.

Who can use it, and when?

For now only a group of cyber defenders selected in announcement week; then additional trusted partners through the Fairwind Program. For developers, businesses and consumers Google says "as soon as possible"; there is no date for the Gemini app, Antigravity or the API. Google also says it is gathering feedback and iterating on guardrails, and that it takes part in the U.S. government's voluntary pre-release model testing process.

Which routes are open right now, what Argon's 13-day timeline could mean for Carbon, and how to apply to Fairwind are in When Is Gemini 4 Carbon Coming? Who Can Use It?. In short: you cannot reach Carbon today; you can reach Argon with a Google AI Pro or Ultra subscription.

Frequently Asked Questions

What is Gemini 4 Carbon?

Google DeepMind's new frontier model, announced on October 28, 2026. Google says it is its strongest model for cybersecurity and long-horizon agentic work, and that it builds on Argon.

Is Carbon a different model from Argon?

Yes, a separate model; but Google describes the relationship as "builds on Argon." Whether it is a new architecture or an improved version of Argon has not been disclosed.

Can I use Gemini 4 Carbon?

Not for now. It is open to selected cyber defenders and will then extend to additional partners through Fairwind. There is no date for general access.

Is Gemini 4 Carbon free, and what does it cost?

There is no announcement that it is free. The API intro price was announced as $2 per million input tokens and $10 for output, but the model is not in the general API yet.

Is Carbon really the best model?

It is first on Artificial Analysis's v5.0 index, but the gaps to rivals are small, and in Google's own table GPT-6 Astra leads on FrontierSWE 2 and Claude Opus 5.5 on OSWorld-2.1. Remember too that the cyber benchmarks were evaluated without guardrails and that its hallucination rate looks higher than Argon's. All the numbers are in the comparison post.

Can I trust the benchmark results?

The numbers are values reported by Google and the measuring organizations. Because the model is not generally accessible, there is no independent reproduction; do not decide until you have tried it on your own work.

Sources

• Google, Gemini 4 Carbon: our next frontier model (October 28, 2026): relationship to Argon, benchmark results, pricing, safety claims, access plan.
• 9to5Google, Google announces Gemini 4 Carbon with 'significant' coding gains (October 28, 2026): coding and agentic gains, behavior improvements.
• Dealroom, Google DeepMind's Gemini 4 Carbon takes the #1 spot on Artificial Analysis, two weeks after Argon: Artificial Analysis ranking, relationship to Argon as reported from Ars Technica.
• Tech My Money, Gemini 4 Carbon Is Google's New Frontier Model, Starting With Cyber Defenders (November 1, 2026): rivals' values in Google's table and missing information (context window, model card).
• Google DeepMind, Fairwind Program: the program through which access is staged.