Gemini 4 Carbon strengthens cybersecurity and long-format code

Google launches Gemini 4 Carbon, its most ambitious frontier model, initially reserved for a limited circle of cyber defenders. The company highlights gains in code, security, and alignment, with a gradual rollout.

Google presented Gemini 4 Carbon on October 28, 2026, as its most powerful frontier model, with deployment initially limited to a subset of cyber defenders. The company says it wants to test its safeguards before a wider release to developers, businesses, and then individuals. The challenge is clear: push performance without loosening usage conditions too quickly, an exercise where AI sometimes loves to run before tying its shoelaces.

A model designed for long tasks and stubborn bugs

Gemini 4 Carbon, Google DeepMind’s new model, targets the longest engineering problems, with agent workflows capable of lasting several days. According to Google, internal teams have seen it persevere on complex tasks, such as identifying and fixing subtle bugs in large source codes and low-level systems. Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect at Google, describes it as “the company’s strongest model.”

The model succeeds Gemini 4 Argon, launched two weeks earlier. Google says it targeted code, agent tasks, scientific uses, full-stack web development, and multimodal understanding. On the Artificial Analysis Intelligence Index version 5.0, Carbon takes first place with a score of 59, according to the announcement. On WebDev Arena, Google says it gained 89 Elo points to reach 1768 and fourth place, with improved output quality and web artifacts.

Why Google initially reserves it for cyber defenders

The rollout follows a cautious logic. From the week of the announcement, Google is to open Carbon to a small group of cyber defenders, then gradually expand access to other trusted partners via its Fairwind Program. The company says it participates in the US government’s voluntary process for model testing before release and continues to adjust its safeguards before making it available on its products.

Fairwind Program, operated by Google DeepMind, grants early access to prioritized actors, including governments, health players, and telecom operators. Google DeepMind says it works with more than 650 partners worldwide. The rules are strict: only authorized defensive uses are allowed, such as threat simulation, malware analysis, or reverse engineering for defensive and academic research purposes. Malicious uses are banned.

Strong cybersecurity figures, but under control

On cyber defense, Google reports results placing Carbon at the top on several test benches. The company claims the model scores 88.2% on SEC-bench Pro, 54.8% on ExploitGym, and 71.0% on Collinear’s CWE-bench. These measures cover vulnerability detection, exploitability verification, and root cause correction. External evaluations were reportedly conducted without safeguards, according to Google.

The firm also reports that Carbon more than doubles secure code generation success compared to Argon on internal tests, and that the Chrome team observed a result “more than 10X better” than another major frontier lab model for producing quality patches. Wiz, Google’s cloud security platform, also sees Carbon leading on web penetration testing benchmarks, a key point for its Scan for Good initiative.

Alignment, safeguards, and introductory pricing

Carbon uses the security stack employed for Argon, with multiple layers: model training, real-time abuse defenses, prompt injection resistance, and reasoning chain monitoring. Google says it reduced cheating success to 0.0% on Terminal-Bench 4.0, while dropping attempts to 0.4% and improving overall accuracy. Serious misalignment alerts reportedly fell by nearly 80% in real coding environments.

On multi-step tasks, Google says it reduced overshooting by 26% and underserving by 29%, while unfounded responses were halved compared to Argon. The company adds that Carbon more clearly indicates what is missing when an action is blocked, instead of inventing a shaky answer. On Gray Swan’s Indirect Prompt Injection Benchmark, Google announces a 0.0% attack success rate, after Argon had already reached 0.7% on this test.

Regarding pricing, Carbon starts at the same level as Argon: 2 dollars per million input tokens and 10 dollars per million output tokens, with cached input tokens billed 95% cheaper. After the launch period, the price will rise to 4 dollars and 20 dollars per million tokens. Google says it wants to observe how the first cyber defenders use the model before expanding access to developers, businesses, and consumers.

By the numbers

  • October 28, 2026 — date of Gemini 4 Carbon announcement by Google.
  • 650 partners — global network claimed by Google DeepMind.
  • 88.2% — score announced on SEC-bench Pro.
  • 2 dollars / 10 dollars — launch price per million input/output tokens.
  • 0.0% — attack success rate on Gray Swan’s Indirect Prompt Injection Benchmark.

To read