Google launches Gemini 4 Carbon for cybersecurity tasks

A model for sustained work

Google announced Gemini 4 Carbon on October 28, 2026, positioning it as its most capable model for cybersecurity and long-horizon tasks. Unlike models designed for quick interactions, Carbon is built for sustained agentic work, which refers to AI that can perform multi-step actions over long periods. Google's internal teams have observed the model running for days to complete complex engineering projects.

The model is specifically designed to persist through dead ends when navigating difficult problems. For example, it can work through large codebases to identify the root causes of subtle bugs and then generate the necessary patches. This capability represents a shift toward AI that can act as a persistent worker rather than a simple chatbot.

Google describes this new capability as a way to push the frontier of what AI can achieve in technical environments. Koray Kavukcuoglu, the head of Google DeepMind and the company's lead AI architect, noted that the model shows how quickly the company is advancing the limits of the technology. The model builds directly on the architecture of Gemini 4 Argon, Google's previous frontier model.

Cybersecurity benchmarks

To demonstrate its capabilities, Google released several benchmark scores for Carbon, primarily focusing on security. The model achieved an 88.2% score on SEC-bench Pro, which was evaluated without guardrails. This specific score was self-computed by Google based on a 183-task JavaScript engine split, where the model was allowed an average of three attempts per task.

In other security tests, Carbon scored 54.8% on ExploitGym and 71.0% on Collinear’s CWE-bench. These tests were conducted under various conditions, including some with no internet access and others with specific timeouts. Additionally, the model reached #4 on the WebDev Arena with a score of 1768, gaining 89 elo points in web development tasks.

Google also reported significant improvements in safety and reliability during internal testing. The company claims Carbon eliminated successful cheating on Terminal-Bench 4.0, with the frequency of such attempts dropping to 0.4%. Furthermore, it achieved a 0.0% attack success rate on Gray Swan’s Indirect Prompt Injection benchmark, compared to 0.7% for the Argon model.

Internal performance gains

While some benchmarks are public, Google highlighted several internal results to showcase Carbon's superiority over Argon. The company reported that Carbon more than doubled its success rate for generating secure code. Additionally, the Chrome team observed that Carbon performed more than 10 times better than a large frontier model from another lab when producing high-quality patches.

The security platform Wiz also observed that Carbon leads on web penetration testing benchmarks. These findings suggest that the model is becoming increasingly proficient at identifying and fixing vulnerabilities in complex systems. However, Google did not name the specific lab or the benchmark used for the 10x performance comparison.

The company also reported a 26% decrease in instances of overreach and a 29% drop in under-service during testing. While Google did not provide specific definitions for these terms, they are intended to measure how accurately the model follows instructions without exceeding its bounds. These internal metrics serve as the primary evidence for the model's refined behavior.

Rollout and pricing

Google is not releasing Carbon to the general public immediately. Instead, it is starting with a subset of cyber defenders this week. Access will eventually expand to other trusted partners through the Fairwind Program, with developers, businesses, and consumers following later. The company has not yet provided a specific date for general availability.

For those who gain access, the pricing is structured around API token costs. Carbon launches at an introductory price of $2 per million input tokens and $10 per million output tokens. These rates are the same as those offered for the Argon model, though a footnote indicates prices will eventually double to $4 and $20 respectively.

Users can also take advantage of a significant discount on cached input tokens, which cost 95% less than the standard input price. Despite the technical details provided, the announcement currently lacks a model card or specific context window limits. It also notes that Google is participating in a voluntary government initiative to test models before they are released to the public.