Google Announces Gemini 4 Carbon Frontier Model
New AI model prioritizes cybersecurity defense and long-horizon engineering capabilities.
1Launch of Gemini 4 Carbon
Google DeepMind unveiled Gemini 4 Carbon, its latest frontier model designed for advanced cybersecurity and long-horizon engineering tasks. Announced on October 28, 2026, the model follows the release of Gemini 4 Argon by just two weeks. It currently ranks number one on the Artificial Analysis Intelligence Index with a score of 59.
2Cybersecurity Performance Metrics
Carbon demonstrates significant leadership in security benchmarks, achieving 88.2% on SEC-bench Pro for vulnerability detection and 54.8% on ExploitGym for exploitation tasks. Additionally, it scored 71.0% on CWE-bench for patching. These results highlight the model's specialized capability in managing the entire cybersecurity defense lifecycle, from initial detection to final remediation of vulnerabilities.
3Agentic Workflow Capabilities
Engineered for sustained agentic workflows, Carbon excels at complex, multi-step engineering challenges. Internal testing shows the model can operate for days, persisting through dead ends to resolve subtle bugs in large codebases. This capability marks a shift toward models that can handle low-level systems and long-horizon tasks without constant human intervention.
4Web Development Advancements
Beyond security, Carbon shows major improvements in full-stack web development and multimodal understanding. The model achieved a notable 89 Elo point increase, reaching a ranking of 1768 on the WebDev Arena. This jump to fourth place reflects enhanced output quality and artifact generation, making it a powerful tool for modern software engineering teams.
5Real-World Security Impact
The Chrome team reported that Carbon performed ten times better than other frontier models when generating high-quality patches. Furthermore, the cloud security platform Wiz observed that Carbon leads in web penetration testing benchmarks. These practical applications demonstrate the model's effectiveness in generating secure code and supporting critical infrastructure protection initiatives like Scan for Good.
6Safety and Alignment Strategy
Building on the safety stack of the Argon model, Carbon incorporates real-time misuse defenses and prompt-injection resilience. Google trained the model to maintain alignment even during complex or high-pressure tasks. The company is also engaging in the U.S. government’s voluntary pre-release testing process to ensure reliability before a broader public rollout occurs.
7Phased Deployment Approach
Google is adopting a deliberate rollout strategy, starting with a select group of cyber defenders this week. Access will gradually expand to trusted partners through the Fairwind Program. Feedback gathered during this phase will inform final guardrails before the model becomes available to general developers, businesses, and consumers via standard product surfaces.
8Pricing and Future Outlook
Gemini 4 Carbon launches with an introductory price of $2 per million input tokens and $10 per million output tokens. Cached input tokens receive a 95% discount. After the introductory period, rates will double to $4 and $20 respectively, reflecting the model's advanced engineering and security-focused capabilities for enterprise-grade applications.