Google has introduced its latest artificial intelligence model, Gemini 4 Carbon, which is designed to enhance cybersecurity by identifying software vulnerabilities, assessing their exploitability, and generating solutions. This innovative model was revealed on October 28, 2026, and builds upon the capabilities of its predecessor, Gemini 4 Argon. However, access to Carbon will initially be limited to selected cybersecurity professionals, with a broader release planned later.
Enhanced Cybersecurity Capabilities
Gemini 4 Carbon stands out for its enhanced cybersecurity features, which include improved capabilities for handling complex, time-consuming security tasks. Google’s internal teams have observed Carbon’s ability to navigate through complicated codebases, detecting subtle bugs in low-level systems and creating effective patches that address these issues at their core.
In terms of performance, Carbon achieved impressive scores across several benchmarks: 88.2% on SEC-bench Pro for vulnerability detection, 54.8% on ExploitGym for exploitation testing, and 71.0% on CWE-bench for patching vulnerabilities. These evaluations highlight Carbon’s proficiency in identifying flaws, assessing their impact, and rectifying the underlying code issues, all without protective measures usually in place in real-world deployments.
Benchmark Success and Comparisons
Carbon also leads the Artificial Analysis Intelligence Index v5.0 with a score of 59, reflecting its capabilities across various domains including coding, scientific analysis, and general knowledge. While this index offers a broad performance overview, it does not replace specialized security assessments. In Google’s internal testing, Carbon’s ability to generate secure code significantly surpassed that of Argon, more than doubling its success rate. The Chrome team reported a tenfold improvement in producing quality patches compared to an unnamed competitor, though this claim could not be independently verified.
Safety is a key focus in Carbon’s development, incorporating multiple layers such as misuse detection, prompt injection defense, and continuous monitoring of its processes. On the Gray Swan Indirect Prompt Injection benchmark, Carbon demonstrated a 0.0% attack success rate, a significant improvement over Argon’s 0.7%.
Deployment and Pricing
Carbon’s rollout will commence with select cyber defenders, followed by additional trusted partners through Google’s Fairwind Program. Furthermore, Google is engaging in voluntary testing with the U.S. government before a broader deployment. The introductory pricing for Carbon is set at $2 per million input tokens and $10 per million output tokens, with substantial discounts for cached inputs. These rates are expected to increase to $4 and $20 in the future, although a specific date for general availability has not been disclosed.
Gemini 4 Carbon represents a significant advancement in Google’s AI-driven security solutions, offering robust protection against emerging digital threats. By initially partnering with a select group of cybersecurity experts, Google aims to refine and enhance Carbon’s capabilities before its wider release.
