Gemini 4 Carbon Explained: Google's Cyber-Defense AI and What It Means for Your IT Security

 What is Gemini 4 Carbon? 

Google DeepMind's new frontier AI model, announced on 28 October 2026, called Gemini 4 Carbon, is designed to help with cybersecurity defense and long-horizon agentic coding. It drives best practices in finding, validating, and patching software vulnerabilities. The release is currently limited to a select group of cyber defenders, vetted by Google's Fairwind Program, and has not been announced. 

What is the name of the Carbon in the Gemstone 4? 

On 28th October 2026, Google DeepMind unveiled Gemini 4 Carbon as their most capable frontier model, where they are focusing their efforts on two areas: cybersecurity and long-horizon tasks. 

Google released Gemini 4 Argon two weeks ago, which Carbon is based on. Argon works on broad, general-purpose, agentic tasks, whereas Carbon is optimized for what it means to be a full-time security professional: Know where to look, understand it's exploitable, and put it to the right fix. 

The key aspect of the model is that there are three steps to framing. Most security tooling is limited to detection and presents a long list of findings to a human.  Carbon is built to carry a vulnerability through validation and remediation inside one sustained agentic workflow, which is where the "long-horizon" label comes from. 

Google also reports that Carbon is more than twice as likely as Argon to generate secure code on its internal benchmarks — a signal that the security tuning shows up in everyday development output, not just in red-team scenarios. For rival approaches, see how Claude Opus 4.6 handles long coding tasks. 

Who can access Gemini 4 Carbon today? 

Gemini 4 Carbon is not publicly available. At launch, Google made the model available only to a subset of cyber defenders, with staged expansion to approved partners through its Fairwind Program, which counts more than 650 participating organizations worldwide, primarily governments, critical infrastructure operators, and large technology platforms. 

Access carries conditions. Google ran its offensive-security benchmarks without standard guardrails to measure raw capability, and in exchange restricts the model to vetted defenders under terms that require strong authentication, tracked usage, and an explicit prohibition on generating malware. Google has also said it is participating in a voluntary U.S. government effort to test frontier models before release, and that it is continuing to gather feedback and iterate on guardrails before opening Carbon to developers, businesses, and consumers. 

The practical takeaway: You cannot procure Gemini 4 Carbon today, and no date has been announced. Plan for a staged, partner-first rollout. 

What do the benchmarks actually show? 

Two numbers deserve a closer read

ExploitGym at 54.8% is the lowest score in the set, and that is expected — proving exploitability is harder than spotting a suspicious pattern. It is also the number that matters most for alert fatigue: validation is what separates a real finding from noise. 

0.0% on the Gray Swan indirect prompt injection benchmark is the figure security teams should care about operationally. An agentic model that reads tickets, logs, and repositories is constantly ingesting untrusted text. Google additionally reports an almost 80% reduction in high-severity misalignment alerts in internal testing and a 0.0% success rate for reward-hacking behaviour on Terminal-Bench 4.0. 

Benchmarks are vendor-reported. Treat them as a capability ceiling under favourable conditions, not as a service-level guarantee for your environment. 

Gemini 4 Carbon vs Gemini 4 Argon vs Gemini 3.8 Live 

If you are choosing a model for production work this quarter, Carbon is not the decision in front of you — the choice is between Argon and Gemini 3.8 Live for real-time multimodal tasks. Carbon tells you where the capability frontier is going, which is relevant for your roadmap planning, and the security assumptions in your current stack. 

Why is “long-horizon” important for coding teams?

"Long-horizon" refers to the model being trained to remember what it needs to know about the context and intent of a task, which may last hours and be composed of many steps, instead of just creating one useful response and then forgetting. 

This has implications for engineering teams as to what they should be delegating. The short-horizon models work well for a function, test, regex. A long-horizon model is used for tasks that require a process to be long or complex, such as migrating a service off a deprecated dependency, tracing a vulnerability class through a monorepo, or a refactor that affects forty files, but still needs to build at the end.  Teams that fix those foundations now benefit the moment Carbon-class capability ships. 

What does Gemini 4 Carbon cost? 

Google published introductory pricing of $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 after the introductory period, with cached input tokens discounted by 95%. 

That cached-input discount is the line item to design around. Long-horizon agentic workloads re-send large, stable context — codebases, policies, prior findings — on every step. Structuring prompts so that the stable portion is cacheable is the difference between a predictable bill and a surprising one. 

What should mid-market companies do now? 

• Check Fairwind eligibility: If you operate critical infrastructure or run a security operations function — access is partner-gated, not purchase-gated. 

• Inventory where AI already touches your code: Agent-generated code is entering most pipelines faster than review policy is adapting. 

• Harden against prompt injection now: Carbon's 0.0% benchmark result is Carbon's; the agents you run today likely score far worse. Our managed cybersecurity services from TechNow cover exactly this gap. 

• Establish a model-agnostic AI policy: This is the policy for data classification, logging, and human sign-off, avoiding the need to rewrite governance as you go through model launches. 

• Budget for secure-code review, not model licences: In 2027, the issue will be verification capacity, not budget for model licences, and secure coding should be a part of that process. 

We deliver day-to-day IT support for mid-market companies. To determine the resilience of your pipelines to agentic AI, schedule an AI security readiness assessment. 

FAQ 

Is Gemini 4 Carbon publicly available? 

No. As of 30 October 2026, Gemini 4 Carbon is limited to a vetted subset of cyber defenders, with staged expansion through Google's Fairwind Program. Google has not announced a general availability date for developers, businesses, or consumers. 

How is Gemini 4 Carbon different from Gemini 4 Argon? 

Carbon is built on Argon and is tuned for cybersecurity and long-horizon agentic work. Google reports Carbon is more than twice as likely as Argon to generate secure code on internal benchmarks.

What are Gemini 4 Carbon's cybersecurity benchmark scores? 

Google reports 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation validation, and 71.0% on CWE-bench for patching. 

Can Gemini 4 Carbon be used to create malware? 

Google's access terms for vetted defenders explicitly prohibit generating malware and require strong authentication plus tracked usage. Google reports a 0.0% attack success rate on the Gray Swan indirect prompt injection benchmark. 

What does Gemini 4 Carbon cost per token? 

Introductory pricing is $2 per million input tokens and $10 per million output tokens, moving to $4 and $20 after the introductory period, with a 95% discount on cached input tokens.

Sources 

• Google: Gemini 4 Carbon: our next frontier model (official announcement, 28 Oct 2026) 

• SecurityWeek: Google's cybersecurity capability claims 

• The Hacker News: benchmark coverage