Gemini 4 Carbon Tops AI Index With Cyber Gains, Rolls Out to Defenders First

Gemini 4 Carbon is Google's next frontier model following Gemini 4 Argon, announced Oct. 28, 2026, with state-of-the-art cybersecurity and long-horizon capabilities and a top score of 59 on the v5.0 Artificial Analysis Intelligence Index.

Google unveiled Gemini 4 Carbon on Oct. 28, 2026, as the successor to Gemini 4 Argon. The release combines top benchmark results with a deliberately narrow launch: the first users will be a subset of cyber defenders admitted through the Fairwind Program, with broader availability for developers, businesses and consumers to follow. Google described the model as its strongest yet, with state-of-the-art cybersecurity and long-horizon capabilities, and the rollout order signals a more cautious approach to frontier-model deployment.

The announcement lands in a crowded frontier-model cycle. Artificial Analysis, an independent evaluation firm, updated its Intelligence Index to v5.0 and placed Gemini 4 Carbon first with a score of 59. The same analysis ranks Google's model second only to Anthropic's Claude Opus 5.5 on the Artificial Analysis Coding Agent Index, an indication that Carbon's strengths extend beyond security-specific evaluations. Google's own benchmark reporting adds cybersecurity results that the company says are state of the art.

What makes Gemini 4 Carbon different from Argon?

Google says Carbon's largest gains over Argon are in cybersecurity and long-horizon tasks. On internal benchmarks, Carbon achieves more than 2x the success rate of Argon in generating secure code, according to the company. The Chrome team separately found that Carbon performed more than 10x better than another lab's large frontier model at generating high-quality patches. Those results are the clearest public evidence that Google optimized Carbon for defensive engineering work rather than only for conversational performance.

Long-horizon capability is a measure of how well a model sustains a complex objective over many steps, retaining context and adjusting course as it works. That skill is central to agentic systems, which are expected to plan, execute and verify multi-stage workflows in areas such as software engineering, security operations and data analysis. By coupling long-horizon gains with cybersecurity benchmark leadership, Google is positioning Carbon for use cases that require both extended reasoning and domain-specific competence.

The model also arrives with a stricter release protocol than Argon. Google is not opening general API access immediately; instead, a vetted set of cyber defenders gets the model first through the Fairwind Program. The company has not detailed the full criteria for admission, but the program is designed to put Carbon's capabilities in the hands of security professionals whose primary mission is defense.

How does Gemini 4 Carbon perform on cybersecurity benchmarks?

Google published scores on three cybersecurity evaluation suites. On SEC-bench Pro, which tests a model's ability to find vulnerabilities in software, Carbon scored 88.2%. On ExploitGym, an interactive benchmark that requires a model to generate working exploits, Carbon scored 54.8%. On CWE-bench, which tests vulnerability identification and remediation against MITRE's Common Weakness Enumeration taxonomy, Carbon scored 71.0%.

Benchmark	Gemini 4 Carbon	What it measures
Artificial Analysis Intelligence Index v5.0	59 (high reasoning)	Aggregate intelligence across reasoning, knowledge and instruction-following tests
SEC-bench Pro	88.2%	Finding vulnerabilities in software source code
ExploitGym	54.8%	Generating working exploits in simulated environments
CWE-bench	71.0%	Identifying and remediating Common Weakness Enumeration vulnerabilities
Secure-code generation (Google internal)	More than 2x vs. Argon	Success-rate uplift in generating secure code
Chrome patch-generation test	More than 10x vs. another lab's frontier model	High-quality patch generation

Artificial Analysis reached a broader verdict on the model's overall intelligence. In its v5.0 index update, Gemini 4 Carbon scored 59, placing it first among evaluated frontier models. The same evaluation put Google's model second only to Claude Opus 5.5 on the Coding Agent Index, a benchmark focused on agentic coding tasks. Together, the results give Google a leadership claim on aggregate intelligence while leaving the agentic-coding crown with Anthropic.

Why is Google limiting initial access to cyber defenders?

Google's rollout plan for Carbon is unusual for a frontier-model launch. Instead of opening API access to all developers at once, the company is starting with a subset of cyber defenders admitted through the Fairwind Program. The program is meant to give vetted security professionals early access to Carbon for defensive work while Google evaluates the model's behavior in real-world, high-stakes settings.

Google's chief AI architect framed the decision as a product and safety milestone. In the announcement, Koray Kavukcuoglu, senior vice president at Google DeepMind and chief AI architect at Google, described Carbon as the company's strongest model, with state-of-the-art cybersecurity and long-horizon capabilities. The company has not specified when general availability will begin, saying only that access for developers, businesses and consumers will come as soon as possible.

Today, we're announcing the next frontier model: Gemini 4 Carbon, our strongest model with state-of-the-art cybersecurity and long horizon capabilities.Koray Kavukcuoglu, SVP, Google DeepMind and Chief AI Architect, Google

The defender-first rollout creates an unusual dynamic for enterprise buyers. Security teams inside the Fairwind Program can begin evaluating Carbon for vulnerability discovery, patch generation and secure-code review immediately, while other developers must wait. That sequencing may give defensive security products built on Carbon a head start in the market, and it gives Google a controlled environment to observe how the model performs when used against real-world systems.

What do the new benchmark numbers actually measure?

SEC-bench Pro is designed to measure how well a model can find and patch vulnerabilities in source code. A score of 88.2% means Carbon identified the vulnerability in nearly nine in 10 test cases, according to Google. Benchmarks in this category are increasingly used by security teams to compare models for code-review and triage workloads, where a missed vulnerability can be more costly than a false positive.

ExploitGym tests a different skill: generating a working exploit against a vulnerable program in an interactive environment. Carbon's 54.8% score is lower than its SEC-bench Pro result, which is consistent with the general pattern that writing a functional exploit is harder than identifying a flaw. CWE-bench, where Carbon scored 71.0%, ties the evaluation to MITRE's Common Weakness Enumeration list, giving security teams a common taxonomy for comparing model performance on known vulnerability classes.

Artificial Analysis' Intelligence Index v5.0 takes a broader view, aggregating results across multiple reasoning, knowledge and instruction-following evaluations. The index's 59 score for Carbon makes it the top model in that ranking. The separate Coding Agent Index, where Carbon trails only Claude Opus 5.5, measures how well models handle multi-step coding tasks in agentic settings, a category Google says is central to Carbon's design.

How is Carbon priced and when will it reach developers?

Google has set a clear pricing structure for Carbon and a staged availability plan. The company says broader developer access will follow the initial defender cohort, though it has not committed to a date.

1. Google set introductory pricing at $2 per million input tokens and $10 per million output tokens.
2. Cached input tokens are priced at 95% off the standard input rate.
3. Initial access begins this week for a subset of cyber defenders through the Fairwind Program.
4. Broader availability for developers, businesses and consumers is planned as soon as possible, according to Google.

The pricing puts Carbon in reach of high-volume developers, particularly those running agentic workloads where models must process long context windows repeatedly. The 95% discount on cached inputs is aimed at applications that reuse the same instruction sets or codebases across many calls, a common pattern in security scanning and automated software engineering. Google has not said whether the introductory prices are temporary.

What are the market and stakeholder implications?

Carbon's benchmark position intensifies competition across frontier AI. Artificial Analysis' index places Google first on aggregate intelligence, while Claude Opus 5.5 retains the top spot on agentic coding. That split gives enterprise buyers a clear set of trade-offs: organizations that prioritize general reasoning may favor Carbon, while teams focused on autonomous coding may still lean on Anthropic's model. No single lab currently leads every major axis.

For the security software market, Carbon's scores matter beyond marketing. Vulnerability-finding tools, automated patch generation and secure-code review are becoming core agentic use cases, and a model with 88.2% on SEC-bench Pro and a more than 10x patch-generation improvement over another lab's frontier model could change the economics of those products. Google's Chrome team reported the patch-generation result, giving the company an internal validation that the capability works in a large, real-world codebase.

The defender-first rollout also carries policy significance. By limiting early access to vetted defenders, Google can demonstrate that it is taking the dual-use risks of cybersecurity models seriously. That posture may ease conversations with regulators and enterprise security officers, but it also creates a window in which only a select group of organizations can use the model's most advanced capabilities. Google says the window will be short, though it has not committed to a date for general availability.

What happens next for Gemini 4 Carbon?

Google's next milestone is expanding access beyond the Fairwind Program. The company has not disclosed how many defenders are in the initial cohort or how long the restricted phase will last. If early defensive deployments go well, the path to general API access is likely to accelerate; if problems emerge, Google may extend the controlled rollout while it adjusts safeguards.

Longer term, Carbon's long-horizon and cybersecurity capabilities position it for agentic workflows that run for extended periods: automated security operations, multi-stage code review, incident response and large-scale patch management. These are applications where a model must maintain context over many steps and act on the results of its own work. Carbon's internal secure-code and patch-generation benchmarks suggest Google built the model with those workflows in mind.

For now, the public record is limited to benchmark scores and controlled access. Google has published results for SEC-bench Pro, ExploitGym, CWE-bench, the Artificial Analysis Intelligence Index and its own internal tests, but independent verification of Carbon's real-world security performance will take time. The Fairwind Program may provide the first evidence of how the model behaves outside benchmark settings, and that evidence will shape whether Google's defender-first strategy becomes a template for future frontier-model releases.