Google launches Gemini 4 Carbon for cybersecurity tasks

A specialized tool for defenders

Google has introduced Gemini 4 Carbon, a frontier model specifically engineered for cybersecurity and long-horizon work. This new model follows the release of Gemini 4 Argon by only two weeks and aims to handle complex, multi-step engineering tasks. It is designed to support sustained agentic workflows, which are processes where AI can perform actions autonomously over long periods.

The model is intended to assist with various technical fields including coding, science, and full-stack web development. It is built to understand multimodal inputs and navigate large codebases or low-level systems. Google's teams have observed the model running for days at a time to complete these types of difficult tasks.

Initial access will be limited to a subset of cyber defenders this week. Google plans to expand access gradually through its Fairwind Program, which involves trusted partners. Eventually, the model will be available to developers, businesses, and consumers after Google iterates on the necessary guardrails.

Performance in security testing

The model has undergone significant testing to measure its ability to identify and fix security flaws. As testingcatalog.com reported, Carbon scored 88.2% on SEC-bench Pro for vulnerability discovery in evaluations conducted without guardrails. It also achieved a 54.8% score on ExploitGym for exploitation and a 71.0% score on Collinear’s CWE-bench for producing working fixes.

In internal coding tests, Google found that Carbon performed significantly better than its predecessor, Argon. The model showed more than twice the success rate of Argon for generating secure code. Furthermore, the Chrome team noted that Carbon generated high-quality fixes at a pace exceeding 10 times that of a large frontier model from another lab.

Beyond raw performance, the model shows high reliability in following instructions. It recorded a 0.0% successful cheating rate on Terminal-Bench 4.0, despite having a 0.4% attempt rate. These metrics suggest a high level of adherence to the intended tasks without trying to bypass system constraints.

Safety and reliability metrics

Google DeepMind is integrating the safety and security stack originally introduced for Gemini 4 Argon into the Carbon model. This includes defenses against real-time misuse, prompt-injection resistance, and chain-of-thought monitoring. These features are designed to ensure the model remains helpful while minimizing risks during complex operations.

The model also shows improvements in factual accuracy compared to its predecessor. Google reports that Carbon has roughly half as many unsupported claims as Argon. It also demonstrates 26% less overreach and 29% less under-service, which helps in maintaining the balance of the AI's responses.

Security improvements are also evident in how the model handles indirect prompt injections. The attack success rate on Gray Swan’s Indirect Prompt Injection Benchmark dropped from 0.7% with Argon to 0.0% with Carbon. Additionally, internal coding work saw a reduction in high-severity misalignment warnings of nearly 80%.

Web development and pricing

Carbon has also shown notable improvements in web development capabilities. Its performance rose by 89 Elo points, placing it in fourth place on the WebDev Arena with a score of 1769. This indicates a significant leap in how the model handles front-end and back-end web construction tasks.

The pricing for the model will match the introductory rates set for Gemini 4 Argon. It will cost $2 per million input tokens and $10 per million output tokens. To help with efficiency, Google will offer cached inputs at a cost that is 95% less than standard input tokens.

The staged release of the model is a deliberate move by Google. The company is participating in a voluntary pre-release testing process with the US government. This allows them to test safeguards with trusted cyber defenders before the model reaches a wider audience.