Gemini 4 Carbon

Gemini 4 Carbon is Google's next frontier model, described by the company as its strongest model with state-of-the-art cybersecurity and long horizon capabilities. Announced on October 28, 2026 by Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect at Google, Carbon builds on Gemini 4 Argon, which launched only two weeks earlier, and is designed to take on much harder, long-horizon engineering problems through sustained agentic workflows. The model is aimed first at cyber defenders, and then at developers, businesses, and consumers as access is phased in through Google's product surfaces. Its stated purpose is to deliver meaningful leaps in software engineering and cybersecurity defense without compromising on safety or reliability, and it ranks #1 on the recently updated v5.0 of the Artificial Analysis Intelligence Index at 59. The launch reflects a deliberate effort to push the frontier forward while keeping safeguards in step. Google says its researchers and models are working hand in hand to advance capabilities faster while strengthening safeguards in tandem, and that the rollout is intentionally cautious: the company is engaged in the U.S. government's voluntary process for pre-release model testing, is starting by making Carbon available to a subset of cyber defenders, and will gradually phase access to additional trusted partners through the Fairwind Program while gathering feedback and iterating on guardrails. The underlying problem Carbon targets is that modern engineering and security work is not a single prompt. Root-causing and patching subtle bugs across large codebases and low-level systems, finding real vulnerabilities buried in real-world software, deciding which findings are actually exploitable, and fixing root causes all require sustained, multi-step effort. Carbon's headline capability is cybersecurity defense across the full lifecycle of vulnerability detection, validation, and remediation. On external benchmarks evaluated without guardrails, it achieves a state-of-the-art 88.2% on SEC-bench Pro for autonomously uncovering complex flaws buried in real-world software, sets a new state-of-the-art 54.8% on the ExploitGym exploitation benchmark that defenders use to check which vulnerabilities are exploitable and therefore worth remediating first, and reaches 71.0% on Collinear's CWE-bench for turning validated findings directly into working code fixes at the root cause. Google also reports internal benchmarks showing more than a 2x uplift in success rate for generating secure code over Argon, a Chrome team report that Carbon performed more than 10X better than another lab's large frontier model when generating high quality patches, and Wiz observing Carbon leading on benchmarks for web penetration testing, a core capability for Wiz's Scan for Good initiative. Beyond security, Carbon shows significant gains across coding and agentic tasks, as well as scientific tasks, full-stack web development, and multimodal understanding, areas Google focused on based on user feedback. With improved quality of output and artifacts in web development, Carbon jumps +89 elo points to land at #4 with a score of 1768 on WebDev Arena. Internally, teams have seen Carbon run for days to persist through dead ends and tackle complex, multi-step engineering tasks such as root-causing and patching subtle bugs across large codebases and low-level systems. This persistence is the practical expression of the long horizon claim: the model is built to keep working through a multi-step plan rather than stopping at the first obstacle. Safety and security behavior is another explicit feature group. Gemini 4 Carbon builds on the same multi-layered safety and security stack as Argon, combining model training, real-time misuse defenses, prompt-injection resilience, and chain-of-thought monitoring, and Google trained Carbon to respond to complex or high-pressure tasks in the right way. The reported results are concrete: successful cheating on Terminal-Bench 4.0 dropped to 0.0% with attempts to cheat falling to 0.4% while overall accuracy improved; high-severity misalignment alerts in live coding environments with internal teams dropped by almost 80%; unsupported claims were cut by approximately 2x over Argon; overreach fell by 26% and under-service by 29%; and on Gray Swan's Indirect Prompt Injection Benchmark, Carbon achieves a 0.0% attack success rate, resisting every evaluated attack attempt, compared with Argon's previously low 0.7%. Carbon's approach combines capability training with behavioral shaping. Rather than guessing or forcing a path forward when instructions are ambiguous, Carbon retrieves the relevant sources and surfaces discrepancies to the user before acting. Instead of fabricating or hedging around missing evidence, it puts methodological caveats directly into the deliverable or withholds blocked actions with a clear list of what remains required. And it is trained to follow through on multi-step plans end-to-end while respecting project constraints and staying strictly within scope, which is what produces the simultaneous reduction in overreach and under-service. The model is being matured in real environments, with feedback from the initial cohort of cyber defenders feeding into guardrail iteration before broader release. For users, the stated outcomes are more reliable execution of long, multi-step work; secure code generation at materially higher success rates; faster triage of vulnerabilities because exploitability can be checked rather than assumed; patches that address root causes instead of symptoms; and a collaborative experience in which the model raises discrepancies rather than silently choosing a path. In security terms, the benefit is coverage across detection, validation, and remediation rather than strength at only one stage of the defense lifecycle. Concrete scenarios described in the announcement include autonomously uncovering complex flaws buried in real-world software; validating risk through exploitation so defenders can prioritize which vulnerabilities to remediate first or at all; converting validated findings into working code fixes; root-causing and patching subtle bugs across large codebases and low-level systems over runs that can last for days; generating secure code and high quality patches in live engineering environments; and web penetration testing, the capability Wiz highlights in its Scan for Good initiative. Initial access is limited to a subset of cyber defenders, with additional trusted partners phased in through the Fairwind Program. Google says Carbon will come to developers, businesses, and consumers via its product surfaces as soon as possible. Gemini 4 Carbon launches at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95% off the input token price. After the introductory period expires, $4 per 1M input tokens and $20 per 1M output tokens will apply. Taken together, Gemini 4 Carbon is positioned as a frontier model whose differentiators are sustained long-horizon execution, state-of-the-art cybersecurity performance across detection, validation, and remediation, and measurable improvements in alignment behavior such as honesty, scope discipline, and prompt-injection resistance, released deliberately to cyber defenders first before reaching developers, businesses, and consumers.