AI Models

Gemini 4 Carbon: Google's Cybersecurity-First Frontier at 88.2% SEC-bench

By Yogesh · November 2, 2026 · 9 min read

Sources checked and dated on publication. See our editorial policy.

Two weeks after launching Gemini 4 Argon, Google announced Gemini 4 Carbon on October 28, 2026—the company's strongest model with state-of-the-art cybersecurity capabilities. Carbon ranks #1 on the Artificial Analysis Intelligence Index v5 at 59, leads on three major security benchmarks (88.2% SEC-bench Pro, 54.8% ExploitGym, 71.0% CWE-bench), and is being rolled out first to a subset of cyber defenders through Google's Fairwind Program.

The short answer

Cybersecurity-first rollout — Google is making Carbon available initially only to selected cybersecurity defenders, expanding via Fairwind Program, then to developers—a phased approach reflecting dual-use capabilities.
State-of-the-art security — Carbon achieves 88.2% on SEC-bench Pro (finding vulnerabilities), 54.8% on ExploitGym (exploitability), 71.0% on CWE-bench (patching)—leading all published results.
#1 on Intelligence Index — Carbon scores 59 on Artificial Analysis Intelligence Index v5, ahead of Claude Opus 5.5 (58) and GPT-6 Astra (57).

What Carbon is and who gets it

Carbon is Google's newest frontier model, built on Gemini 4 Argon with additional training for cybersecurity defense and long-running engineering. Phased rollout by use case:

Week of October 28: Subset of cybersecurity defenders
Following weeks: Trusted partners via Fairwind Program
Later (no date): Developers, businesses, consumers via API and product surfaces

Google is "actively engaged in U.S. government voluntary pre-release testing," gathering feedback on guardrails. This controlled rollout reflects Carbon's dual-use nature—capabilities helping defenders could assist attackers if misused.

Cybersecurity benchmark sweep

Carbon achieves state-of-the-art across the defense lifecycle—discovery, validation, remediation:

SEC-bench Pro: 88.2%

SEC-bench Pro tests autonomous discovery of complex security flaws in real software. Carbon's 88.2% is the highest published result.

ExploitGym: 54.8%

ExploitGym tests whether vulnerabilities are exploitable—key triage for prioritization. Carbon's 54.8% sets a new benchmark (measured without guardrails).

CWE-bench: 71.0%

CWE-bench tests turning findings into working root-cause fixes. Carbon's 71.0% demonstrates the final step: actually fixing what's found.

Google reports Carbon achieves more than 2x uplift in secure code generation vs Argon. Chrome team observed more than 10x better patch quality than another lab's frontier model.

Tops Intelligence Index

Model	Index v5	Organization
Gemini 4 Carbon	59	Google DeepMind
Claude Opus 5.5	58	Anthropic
GPT-6 Astra	57	OpenAI

Carbon also jumps +89 elo to #4 (1768) on WebDev Arena, reflecting improved code generation and artifact quality.

Long-horizon engineering

Beyond cybersecurity, Carbon handles sustained agentic workflows—tasks running hours or days. Google says internal teams observed Carbon running days to persist through dead ends, root-causing and patching bugs across large codebases and low-level systems.

Alignment and safety

Carbon builds on Argon's safety stack with alignment for high-pressure scenarios:

Reduced shortcuts

On Terminal-Bench 4.0, Carbon reduced successful cheating to 0.0%, attempts to 0.4%. High-severity misalignment alerts dropped almost 80% in live coding with internal teams.

Greater transparency

Carbon cuts unsupported claims by ~2x vs Argon. Puts caveats directly into deliverables, withholds blocked actions with clear required lists.

Perfect prompt injection resistance

On Gray Swan's IPI Benchmark, Argon led with 0.7% attack success. Carbon achieves 0.0%—resisting every evaluated attack.

Better scope adherence

Reduced overreach 26%, under-service 29%—more reliable multi-step execution while respecting constraints.

Pricing and context

Carbon launches at same introductory pricing as Argon:

$2 per million input tokens
$10 per million output tokens
$0.10 per million cached input (95% discount)

After introductory period: $4 input / $20 output. No end date announced. Context window unpublished; "long horizon capabilities" suggests substantial capacity.

Where Carbon fits

Model	Announced	Focus	Initial access
Gemini 4 Argon	Oct 13	General frontier	API + consumers
GPT-6.1 Terra	Oct 26	Computer use, efficiency	Plus/Pro/Enterprise
Gemini 4 Carbon	Oct 28	Cybersecurity + long-horizon	Cyber defenders only

Carbon is the first frontier model explicitly positioned for defensive cybersecurity as primary use case, with access gated accordingly. The phased rollout—defenders first, developers later—is unprecedented and signals how labs grapple with dual-use capabilities.

Verdict For cybersecurity teams: If in the initial Fairwind cohort, Carbon is the model to test for vulnerability discovery, exploitation validation, patch generation. The 88.2% SEC-bench Pro and 2x secure code uplift are strong signals—verify on your workflows before production.

For developers waiting: Establish a baseline with Argon now, keep integration flexible. When Carbon's API opens, compare quality, time saved, and cost on your actual tasks—not just benchmarks.

For the frontier: Google's controlled rollout sets precedent. Carbon demonstrates frontier labs can ship state-of-the-art with phased access based on use-case risk rather than immediate broad availability.

Bottom line: Gemini 4 Carbon tops Intelligence Index at 59, sweeps cybersecurity benchmarks (88.2% SEC-bench Pro, 71.0% CWE-bench), introduces defenders-first rollout reflecting dual-use nature. For those with early access, strongest published model for defensive security. For everyone else, signal that frontier deployment is evolving beyond "launch day = general availability."

Frequently asked questions

Cybersecurity-first rollout

Google is making Carbon available initially only to selected cybersecurity defenders, expanding via Fairwind Program, then to developers—a phased approach reflecting dual-use capabilities.

State-of-the-art security

Carbon achieves 88.2% on SEC-bench Pro (finding vulnerabilities), 54.8% on ExploitGym (exploitability), 71.0% on CWE-bench (patching)—leading all published results.

#1 on Intelligence Index

Carbon scores 59 on Artificial Analysis Intelligence Index v5, ahead of Claude Opus 5.5 (58) and GPT-6 Astra (57).

Sources

Prices and benchmark figures are standard published rates as of November 2, 2026 and can change without notice. Some outbound links may be affiliate links — see our disclosure. Found an error? Tell us and we'll correct it in public.