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 is designed to take on much harder, long-horizon engineering problems and to support sustained agentic workflows. Its stated purpose is to deliver meaningful leaps in software engineering and cybersecurity defense without compromising on safety or reliability. Carbon builds directly on Gemini 4 Argon, which launched two weeks earlier, and illustrates what Google describes as the pace at which it is pushing the frontier forward, with researchers and models working hand in hand to advance capabilities faster while strengthening safeguards in tandem. Carbon targets a class of work that is hard to complete in a single step. The announcement describes complex, multi-step engineering tasks such as root-causing and patching subtle bugs across large codebases and low-level systems, and notes that Google's internal teams have seen Carbon run for days to persist through dead ends on exactly this kind of problem. Security work is framed around a defense lifecycle of vulnerability detection, validation, and remediation, and Carbon is evaluated against each stage of that lifecycle. The release also reflects a governance problem: Google states it is taking a deliberate approach to rollout, engaging in the U.S. government's voluntary process for pre-release model testing and phasing access to trusted partners rather than releasing broadly at once, so that capabilities can advance while safeguards are strengthened alongside them. Building on user feedback, Google focused Carbon's improvements across coding, agentic and scientific tasks, as well as full-stack web development and multimodal understanding. Internally, teams have seen the model run for days to persist through dead ends on complex, multi-step engineering tasks such as root-causing and patching subtle bugs across large codebases and low-level systems. On full-stack web development, improved quality of output and artifacts lifted Carbon by 89 elo points to #4 with a score of 1768 on WebDev Arena. Carbon also ranks #1 on the recently updated v5.0 of the Artificial Analysis Intelligence Index with a score of 59, and shows significant gains across coding and agentic tasks. These results matter because sustained agentic work is where models typically lose coherence over long runs; Google's claim is that Carbon keeps working through multi-step plans without stalling. Carbon's headline capability is cybersecurity defense, where Google reports state-of-the-art performance across the defense lifecycle. On external benchmarks evaluated without guardrails, Carbon scores 88.2% on SEC-bench Pro for autonomously uncovering complex flaws buried in real-world software, 54.8% on ExploitGym for validating which vulnerabilities are exploitable, and 71.0% on Collinear's CWE-bench for turning validated findings directly into working code fixes. Exploit validation matters because cyber defenders increasingly rely on models to check which vulnerabilities are exploitable so they can prioritize which ones to remediate first, or at all. Internally, Google reports more than a 2x uplift in success rate for generating secure code over Argon; the Chrome team reported Carbon performed more than 10X better than another lab's large frontier model when generating high quality patches; and Wiz, Google's leading cloud security platform, observes Carbon leading on benchmarks for web penetration testing, a core capability for its Scan for Good initiative. 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. With Carbon, Google set out to go a step further on alignment, training the model to respond to complex or high-pressure tasks in the right way, and reports several concrete results. Carbon reduced successful cheating on Terminal-Bench 4.0 to 0.0%, with attempts to cheat dropping to 0.4%, all while improving overall accuracy; in live coding environments with internal teams, high-severity misalignment alerts dropped by almost 80%. Carbon cuts unsupported claims by approximately 2x over Argon, putting methodological caveats directly into the deliverable or withholding blocked actions with a clear list of what remains required rather than fabricating or hedging around missing evidence. When facing ambiguity, it retrieves the relevant sources and surfaces discrepancies to the user before acting. It simultaneously reduced overreach by 26% and under-service by 29%, and it resisted every evaluated attack attempt on Gray Swan's Indirect Prompt Injection Benchmark, achieving a 0.0% attack success rate versus Argon's 0.7%. Google describes a deliberately staged rollout rather than a broad launch. Carbon is being released under the U.S. government's voluntary process for pre-release model testing, and starting in the announcement week it was made available to a subset of cyber defenders, with access gradually phased to additional trusted partners through the Fairwind Program. Google says it is continuing to gather feedback and iterate on guardrails before making Carbon available to developers, businesses, and consumers via its product surfaces as soon as possible. Underneath that rollout, the model's approach combines the same multi-layered safety and security stack as Argon with training aimed at handling complex or high-pressure tasks correctly, an emphasis on transparency when evidence is missing, and behavior that retrieves sources and surfaces discrepancies rather than guessing or forcing a path forward. Stated outcomes for users include much more reliable execution of multi-step plans end-to-end while respecting project constraints and staying strictly within scope, fewer unsupported claims, and a better collaborative experience when requirements are ambiguous. For security teams, the payoff is a model that can help detect, validate and remediate vulnerabilities rather than stopping at a list of findings. For engineering teams, the reported benefit is persistence on long-running problems: working through dead ends across large codebases and low-level systems instead of giving up after the first failed attempt. Google also points to a stronger internal capability to generate secure code, and to reduced misalignment and fabrication rates in live coding environments with its own teams. The announcement frames several concrete scenarios where Carbon is used. It autonomously uncovers complex flaws buried in real-world software, checks which vulnerabilities are exploitable so defenders can prioritize remediation, and turns validated findings directly into working code fixes. It is applied to sustained engineering work such as root-causing and patching subtle bugs across large codebases and low-level systems over multi-day runs. It supports full-stack web development through improved quality of output and artifacts. It is also used for web penetration testing, the capability behind Wiz's Scan for Good initiative, and for generating high quality patches inside Google's own teams, including work reported by the Chrome team. The initial audience is a subset of cyber defenders, followed by additional trusted partners through the Fairwind Program, with developers, businesses, and consumers to follow via Google's product surfaces. 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. Google notes that after the introductory period expires, the price of $4 per 1M input tokens and $20 per 1M output tokens will apply. The model is presented as part of the Gemini 4 family, following Gemini 4 Argon, which was launched two weeks earlier. Gemini 4 Carbon's primary value proposition is combining frontier-level coding and agentic reasoning with state-of-the-art cybersecurity defense, while holding the line on safety through a deliberately staged release. Google positions it as its strongest model to date, leading the Artificial Analysis Intelligence Index and the SEC-bench Pro, ExploitGym and CWE-bench cybersecurity benchmarks, and as a step toward delivering meaningful leaps in software engineering and cybersecurity defense without compromising on safety or reliability.