> ## Content Index
> Fetch the complete content index at: https://engineerpal.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Google's Gemini 4 Carbon edges Opus 5.5, but cyber defenders get it first
- URL: https://engineerpal.com/gemini-4-carbon-cyber-defenders/
- Published: 2026-10-29T11:00:00.000Z
- Updated: 2026-10-29T11:00:00.000Z
- Description: Google says Carbon leads three security benchmarks. Developers will pay $2 per million input tokens when it opens up, but there's no date.
- Author: Nina Park
- Tags: AI, Google, Gemini, Cybersecurity

Google announced Gemini 4 Carbon on Wednesday (Oct. 28), a new frontier model it calls its strongest yet, and said the first people outside the company to use it will be a vetted group of cybersecurity defenders.

Carbon arrives two weeks after [Gemini 4 Argon](https://engineerpal.com/gemini-4-argon-vs-opus-gpt/). It scores 59 on version 5.0 of the Artificial Analysis Intelligence Index, an independent benchmark roundup, putting it one point ahead of Anthropic's Claude Opus 5.5 at 58 and two ahead of OpenAI's GPT-6 Astra at 57, according to the [Artificial Analysis leaderboard](https://artificialanalysis.ai/leaderboards/models?ref=engineerpal.com).

"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," Google DeepMind SVP Koray Kavukcuoglu said in Google's [announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/?ref=engineerpal.com).

Everyone else has to wait. Google said developers, businesses and consumers will get Carbon "as soon as possible," after it gathers more feedback on guardrails, but it didn't give a date. The company also said it's taking part in the US government's voluntary process for pre-release model testing.

When Carbon does open up, it will carry Argon's introductory price of $2 per million input tokens and $10 per million output tokens, with cached input 95 percent cheaper. After the introductory period, the price doubles to $4 and $20, Google said. Artificial Analysis noted the discount has no confirmed end date.

## Built to find, exploit and fix bugs

The security pitch rests on three outside benchmarks, which Google says it ran without guardrails. Carbon scored 88.2 percent on SEC-bench Pro, which tests whether a model can find real vulnerabilities, 54.8 percent on the exploitation benchmark ExploitGym and 71.0 percent on Collinear's CWE-bench for patching.

Google's SEC-bench Pro chart puts GPT-6 Astra at 85.4 percent and Argon at 83.4 percent, [WebProNews reported](https://www.webpronews.com/google-gemini-4-carbon-cyber-defenders-fairwind/?ref=engineerpal.com). Those are Google-reported results, not independent tests.

Google also cited internal results. It says Carbon more than doubled Argon's success rate at generating secure code, and that the Chrome team found it did more than 10 times better than "another lab's large frontier model" at producing high-quality patches. Google didn't name the lab.

The [Fairwind Program](https://deepmind.google/fairwind-program/?ref=engineerpal.com) is Google's gated channel for that work. It has more than 650 partners, with priority for governments, critical infrastructure operators and core technology platforms. Partners must use phishing-resistant multifactor authentication, can give access only to internal security, incident response or penetration testing teams, and can't resell access.

Carbon is also aimed at long jobs. Google says its internal teams have seen it "run for days to persist through dead ends" on tasks like tracking down and patching subtle bugs in large codebases.

## Not the best coder on the board

The independent numbers are more mixed than the headline rank. In its [analysis](https://artificialanalysis.ai/articles/gemini-4-carbon-new-number-1?ref=engineerpal.com), Artificial Analysis said Carbon took first place on 4 of the index's 9 evaluations, but "is not leading in coding and knowledge work."

On its Coding Agent Index, Carbon running in the Antigravity CLI scored 59, behind Claude Code with Opus 5.5 at 62\. Artificial Analysis still called the gain notable, since Google "has historically underperformed in this area."

The tester also found Carbon's hallucination rate rose to 11 percent from Argon's 4 percent, though that is still the second lowest it has measured. Opus 5.5 and GPT-6 Astra both came in at 50 percent.

On cost, Artificial Analysis put Carbon at $7.76 per index task at standard pricing, about 23 percent less than Opus 5.5 but roughly 1.2 times GPT-6 Astra.

Google's safety numbers are also its own. It says Carbon cut successful cheating on Terminal-Bench 4.0 to 0 percent, reduced high-severity misalignment alerts in live internal coding by almost 80 percent and resisted every evaluated attack in Gray Swan's indirect prompt injection benchmark, compared with a 0.7 percent success rate against Argon.

Argon took a similar path. Google announced it on Sept. 30 with access limited to Fairwind, then made it generally available on Oct. 13\. Google hasn't said whether Carbon will reach the Gemini API on a similar timeline.