Google on October 28, 2026 announced Gemini 4 Carbon, which it calls its strongest model yet, with a focus on cybersecurity and long-horizon software engineering. The release arrives two weeks after Gemini 4 Argon, and Google is not opening it up broadly: starting this week, Carbon goes to a subset of cyber defenders, with access phased out to more trusted partners through the company's Fairwind Program before any general release to developers, businesses or consumers.
In a blog post, Koray Kavukcuoglu, senior vice president at Google DeepMind and Google's chief AI architect, said Carbon ranks first on version 5.0 of the Artificial Analysis Intelligence Index with a score of 59. The Verge reported that the index places Carbon ahead of Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Astra. Google also says Carbon climbed 89 Elo points to fourth place on WebDev Arena.
The cybersecurity numbers are the centerpiece. Google reports that Carbon scores 88.2% on SEC-bench Pro, which tests finding real vulnerabilities in software; 54.8% on ExploitGym, which measures turning a vulnerability into a working exploit; and 71.0% on Collinear's CWE-bench, which tests patching flaws at the root cause. Google says those results were measured without guardrails enabled. These figures are self-reported by Google, and DigiEditorial has not independently tested the model.
Google also cites internal results it has not published in detail: more than double Argon's success rate at generating secure code, and a report from its Chrome team that Carbon produced high-quality patches more than ten times better than another lab's large frontier model, which Google did not name. The company says its security unit Wiz sees Carbon leading on web penetration-testing benchmarks.
Beyond security, Google is pitching Carbon for sustained agentic work. It says internal teams have watched the model run for days on multi-step engineering problems, such as root-causing and patching subtle bugs across large codebases. On safety, Google says Carbon reduced successful cheating on Terminal-Bench 4.0 to 0.0%, cut unsupported claims roughly in half compared with Argon, reduced overreach by 26%, and achieved a 0.0% attack success rate on Gray Swan's indirect prompt injection benchmark. Those are also Google's own measurements.
The staged rollout reflects how dual-use these capabilities are. Google says it is taking part in the U.S. government's voluntary pre-release model testing process. According to Google DeepMind's program page, Fairwind already counts more than 650 partners, prioritizes governments, critical-infrastructure operators and core technology platforms, runs background checks on applicants, and requires phishing-resistant multifactor authentication, with access limited to internal security, incident-response and penetration-testing teams and no resale or sharing.
Carbon will launch at Argon's introductory price of $2 per million input tokens and $10 per million output tokens, with cached input 95% cheaper; after the introductory period, pricing rises to $4 and $20. Google has not given a date for wider availability. Google's October 28 blog post is the primary source for this report; DigiEditorial corroborated the launch, staged access and pricing against reporting from Reuters and The Verge.