Google Unveils Gemini 4 Carbon, a Cybersecurity-Focused AI Model Most People Can’t Use Yet

Google has announced Gemini 4 Carbon, a new frontier AI model that the company calls its strongest yet, with a focus on cybersecurity and long, multi-step coding tasks. But unlike most model launches, almost nobody can use it yet. Google said on Wednesday, Oct. 28, 2026, that Carbon will go first to a small group of cyber defenders, then to trusted partners, before any public release to developers, businesses and consumers.

The announcement came just two weeks after Google launched Gemini 4 Argon, and it lands in a crowded week for AI: OpenAI also released a cheaper computer-operating model, GPT-6.1 Terra, this week, and Anthropic filed for its IPO.

What Google says Gemini 4 Carbon can do

In a blog post by Koray Kavukcuoglu, senior vice president at Google DeepMind and Google’s chief AI architect, the company said Carbon ranks first on version 5.0 of the Artificial Analysis Intelligence Index, an independent composite score, at 59. Google said the model was designed for “much harder, long-horizon engineering problems,” and that internally its teams have seen Carbon “run for days” on complex tasks such as finding and patching subtle bugs across large codebases.

On web development, Google said Carbon gained 89 Elo points to rank fourth on the WebDev Arena leaderboard, with a score of 1768. The company also reported gains in agentic, scientific and multimodal tasks.

The cybersecurity pitch, by the numbers

The headline claims are about security. Google describes the “cybersecurity defense lifecycle” as three steps: finding a vulnerability, confirming whether it can be exploited, and fixing it. It reported a score for each, measured without guardrails, using outside benchmarks.

According to Google, Carbon scored 88.2% on SEC-bench Pro, a test of finding real vulnerabilities in software; 54.8% on ExploitGym, which measures whether a model can validate that a flaw is exploitable; and 71.0% on Collinear’s CWE-bench, which tests whether it can write a working fix at the root cause. Google called all three state-of-the-art results.

The company also cited internal results. It said Carbon achieved more than twice the success rate of Argon in generating secure code, and that Google’s Chrome team reported Carbon performed “more than 10X better than another lab’s large frontier model” at producing high-quality patches. Google did not name the other lab. Wiz, the cloud security company Google owns, said Carbon leads on its web penetration-testing benchmarks, according to the post.

These figures come from Google and have not been independently reproduced. Ars Technica, reviewing Google’s comparison charts, reported that Carbon beat OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5 and Claude Fable 5.1 in almost every test, but noted that the gap between Carbon and its rivals is “just a few points.” SecurityWeek also reported the comparison against those three models.

Safety claims and the slow rollout

Google said Carbon builds on Argon’s safety stack, which combines model training, real-time misuse defenses, prompt-injection resilience and chain-of-thought monitoring. Among the specific claims in the blog post:

• On Gray Swan’s indirect prompt injection benchmark, Carbon’s attack success rate was 0.0%, compared with 0.7% for Argon. Ars Technica reported Google’s chart listed Claude Opus 5.5 at 1% and GPT-6 Astra at 8.5%.
• On Terminal-Bench 4.0, successful cheating fell to 0.0%, and attempts to cheat dropped to 0.4%.
• In live coding with internal teams, high-severity misalignment alerts dropped by almost 80%.
• Unsupported claims fell by about half compared with Argon, overreach fell 26% and under-service fell 29%, by Google’s internal measures.

Google said it is “actively engaged in the U.S. government’s voluntary process for pre-release model testing.” Starting this week, Carbon is going to a subset of cyber defenders, with access phased in for more partners through the Fairwind Program. SecurityWeek reported the program includes more than 650 organizations. Google said it is gathering feedback and iterating on guardrails before making the model widely available “as soon as possible,” but gave no date.

Price and what it means for developers

When it reaches developers, Carbon will carry the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens, with cached input tokens 95% cheaper. After the introductory period, the price rises to $4 and $20. Ars Technica reported that Google declined to give an end date for Carbon’s discount; Argon’s discounted rate runs through the end of November.

The staged release mirrors the approach Google took with Argon, putting a model with strong security skills in the hands of defenders before the general public, a plan Ars Technica described as a release that could take a while. That matters for ordinary website owners too: AI-found flaws are already turning up in widely used software, as in the PHP2Shell vulnerabilities we covered last week. For more on the current model race, see our reports on OpenAI’s GPT-6.1 Terra and Google DeepMind’s Genie 4 world model.

Sources

• Google: Gemini 4 Carbon, our next frontier model
• SecurityWeek: Google Touts Powerful Cybersecurity Capabilities in Gemini 4 Carbon Model
• Ars Technica: Google announces Gemini 4 Carbon, its latest frontier AI model that you can’t use
• The Verge: Google launches Gemini 4 Carbon with a ‘significant leap’ in cybersecurity