Google Launches Gemini 4 Carbon, a Frontier Model Built for Cybersecurity Defense
Google DeepMind has announced Gemini 4 Carbon, its strongest model yet, with state-of-the-art scores across vulnerability discovery, exploitation and patching, and a staged rollout starting with cyber defenders.
Google DeepMind has announced Gemini 4 Carbon, its next frontier model. Announced on October 28, 2026, Carbon is the successor to Gemini 4 Argon, which Google launched two weeks earlier. DeepMind calls it the company’s strongest model, with state-of-the-art cybersecurity and long-horizon capabilities.
The headline number is an index ranking. Carbon ranks first on version 5.0 of the Artificial Analysis Intelligence Index, with a score of 59. That index aggregates performance across a range of evaluations, and Google presents the top ranking as evidence that Carbon leads the current frontier field.
The more specific claims are in cybersecurity. DeepMind says Carbon is state of the art across the defense lifecycle: finding vulnerabilities, validating them through exploitation, and patching them at the root cause. On SEC-bench Pro, a benchmark for autonomously uncovering flaws in real-world software, Carbon scores 88.2 percent. On ExploitGym, which tests whether a model can exploit a vulnerability to confirm it is real, it scores 54.8 percent. On CWE-bench, which measures turning validated findings into working code fixes, it reaches 71.0 percent. Google notes these external benchmarks were evaluated without guardrails.
Internal results point the same way. On Google’s own benchmarks, Carbon generates secure code at more than twice the success rate of Argon. The Chrome team reported that Carbon performed more than ten times better than another lab’s large frontier model at generating high-quality patches, according to Google. Wiz, Google Cloud’s security platform, observes Carbon leading on benchmarks for web penetration testing, a capability behind its Scan for Good initiative.
Built for long, hard engineering jobs
Beyond security, Carbon is designed for long-horizon work: tasks that run for hours or days rather than minutes. DeepMind says its internal teams have seen Carbon persist through dead ends to tackle complex, multi-step engineering jobs, such as root-causing and patching subtle bugs across large codebases and low-level systems.
That emphasis shows up in other benchmarks too. On WebDev Arena, which ranks models on full-stack web development, Carbon gained 89 Elo points to land fourth, at 1769. DeepMind says output quality and artifacts in web development improved alongside coding, agentic and scientific performance and multimodal understanding.
Safety and alignment get a dedicated section of the announcement, and the numbers are unusually concrete. Carbon reduced successful cheating on Terminal-Bench 4.0 to 0.0 percent, with cheating attempts falling to 0.4 percent, while improving overall accuracy. In live coding environments with Google’s own teams, high-severity misalignment alerts dropped by almost 80 percent. Internal testing found Carbon cuts unsupported claims by roughly half compared with Argon, reduces overreach by 26 percent and under-service by 29 percent, and resists every evaluated attack on Gray Swan’s Indirect Prompt Injection benchmark, a 0.0 percent attack success rate against Argon’s 0.7 percent.
A staged rollout, defenders first
Carbon will not be generally available at launch. Google says it is taking a deliberate approach to rollout and is engaged in the U.S. government’s voluntary process for pre-release model testing. Starting this week, Carbon goes to a subset of cyber defenders, with access gradually phased to additional trusted partners through DeepMind’s Fairwind Program. Developers, businesses and consumers will get access through Google’s product surfaces as soon as possible, the company says, without giving a date.
Pricing is set to match Argon’s introductory terms: 2 dollars per million input tokens and 10 dollars per million output tokens, with cached input tokens priced at 95 percent off the input rate.
The timing matters for the frontier race. Argon arrived two weeks ago, and Carbon follows it within the same month, a pace DeepMind frames as researchers and models working hand in hand to advance capabilities faster while strengthening safeguards in tandem. Reuters, reporting the launch on October 28, described it as Google stepping up competition with OpenAI and Anthropic in coding and cybersecurity systems.
The staged rollout also follows the pattern Google set with Argon, which went to cybersecurity partners first before reaching wider availability. The logic is straightforward: a model that is unusually good at finding and exploiting vulnerabilities is dual-use by nature, and Google is choosing to put it in defenders’ hands first while it gathers feedback and iterates on guardrails. That sequencing is itself part of the story. Frontier releases are no longer just benchmark announcements; the order of access, and who gets the model first, has become a safety decision in its own right.
What is still unknown is how Carbon performs outside Google’s chosen benchmarks and its staged rollout. The external scores are strong, but they are defense-oriented by design, and general availability has no date yet. For now, the concrete fact is the launch itself: Google’s second frontier model in two weeks, aimed squarely at the people who defend software rather than only those who build it.