Google’s Gemini 4 Carbon Targets Cybersecurity Gains While Remaining Limited to Selected Defense Partners

Google has launched Gemini 4 Carbon, a frontier model positioned as an upgrade to Gemini 4 Argon with stronger coding, agentic reasoning, scientific analysis, and cybersecurity performance. The model is available only to a selected group of cybersecurity defense companies during its initial release, with broader access planned through Google’s Fairwind Program. The Verge reported the launch on October 28, 2026, citing Google’s announcement.

Carbon arrives roughly 2 weeks after Argon, giving Google another high-end model release in quick succession. The company says the system is designed for longer, multi-step work and has been observed running for days while working through difficult engineering tasks. Ars Technica described Carbon as a relatively incremental improvement over Argon, with benchmark advantages that remain narrow across several competing frontier models.

Cybersecurity is Gemini 4 Carbon’s defining focus. Google says the model makes a “significant leap” in vulnerability detection, validation, remediation, and secure code generation. The company also says Carbon is twice as likely as Argon to produce code free of vulnerabilities, while Chrome’s team found it generated high-quality patches at 10 times the rate of non-Google models.

Carbon ranks first on SEC-bench Pro, which evaluates discovery of real-world software bugs, and on ExploitGym, which measures whether models can turn vulnerabilities into exploits. Those results give the model a stronger security profile than a general coding upgrade would provide. They also underline the dual-use nature of the release: the same capabilities that help defenders identify and repair flaws can assist researchers in reproducing attacks.

Google places Gemini 4 Carbon at the top of the Artificial Analysis Intelligence Index, ahead of Claude Opus 5.5 and GPT-6 Astra. Carbon also gained 89 points on WebDev Arena, reaching fourth place on that leaderboard. The results put Google’s latest model near the front of the current competition in coding and general intelligence evaluations.

The margins remain relatively small. Ars Technica noted that Carbon outperforms Argon by slim amounts across the tests Google disclosed, while the leading models cluster in the same general range. Benchmark leadership therefore establishes a competitive signal, but it doesn't by itself demonstrate a broad real-world advantage. Performance on security tasks, sustained engineering work, and deployment conditions will determine whether Carbon’s gains translate beyond controlled evaluations.

Gemini 4 Carbon is built to persist through extended engineering tasks, according to Google. The company says it handles dead ends more effectively, avoids shortcuts, resists prompt injection attacks, and collaborates more closely with users. Google also says Carbon cuts unsupported claims by 2 times compared with Argon, framing the change as an improvement in transparency and honesty.

Those features matter for cybersecurity systems that inspect code, investigate vulnerabilities, or propose patches across large repositories. A model that continues working across multiple steps can reduce the need for constant human prompting, while stronger resistance to prompt injection helps protect workflows that process untrusted code or external data. Human review remains central because vulnerability discovery and exploit generation carry direct operational consequences.

Gemini 4 Carbon begins with restricted access rather than a broad public release. Selected cybersecurity defense companies are testing the model first, and other partners are expected to gain access through Google’s Fairwind Program. That rollout gives Google a controlled setting for evaluating a system with both defensive and offensive security capabilities.

The model will launch at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens. Pricing keeps Carbon within reach for organizations running high-volume coding or security workflows, but access restrictions mean developers and ordinary users cannot immediately test the model directly. The staged release also leaves open how quickly Google will expand availability and whether future versions will be offered through wider developer products.

Gemini 4 Carbon’s next test is deployment rather than leaderboard position. Cybersecurity partners will provide evidence about vulnerability triage, patch quality, exploit analysis, prompt-injection resistance, and the amount of human review required for safe operation. Those results will shape the case for expanding access through Fairwind and other Google channels.

Google’s rapid progression from Argon to Carbon also reflects the intensity of frontier-model competition. The model’s advantages are meaningful in security benchmarks, but the disclosed gaps over other leading systems are narrow and the public rollout is limited. Carbon’s long-term importance will depend on whether its security improvements hold up against real software, unfamiliar vulnerabilities, and adversarial users outside Google’s evaluation environment.