Google launches Gemini 4 Carbon for complex engineering and security

Built for long-horizon tasks

Google DeepMind has introduced Gemini 4 Carbon, a model designed to tackle difficult engineering problems that require sustained effort. Unlike many AI models that focus on quick interactions, Carbon is built for long-horizon agentic workflows. These are tasks where the AI must work through multiple steps over a long period to reach a final goal.

Internal testing shows that Carbon can run for days while completing complex tasks. This includes identifying the root causes of bugs and patching them across large codebases and low-level systems. By maintaining focus over time, the model can handle the heavy lifting of software engineering that usually requires human oversight.

The model builds on Gemini 4 Argon, which was released two weeks ago. While Argon serves as a foundation, Carbon is specifically tuned for more demanding technical work. It currently holds the top spot on the Artificial Analysis Intelligence Index version 5.0 with a score of 59.

Advancements in cybersecurity

A primary focus of Gemini 4 Carbon is its improved ability to handle security-related tasks. As Pulse 2.0 reported, the model achieved an 88.2% score on SEC-bench Pro, which measures the ability to find real-world software vulnerabilities. It also scored 54.8% on ExploitGym for validating those vulnerabilities and 71% on CWE-bench for creating working fixes.

Google noted that Carbon achieved a success rate more than twice that of Gemini 4 Argon on internal benchmarks for generating secure code. Additionally, the Chrome team found that Carbon was significantly more effective than a competing frontier model at producing high-quality software patches, outperforming it by a factor of more than 10.

Beyond just finding bugs, the model is being tested for its ability to resist malicious inputs. On the Gray Swan Indirect Prompt Injection benchmark, Carbon recorded a 0% attack success rate. This means it successfully resisted every evaluated attempt to hijack its instructions through indirect prompts.

Reasoning and web development

While security is a major pillar, Carbon also shows broad improvements in general reasoning and development. Google reported gains in scientific reasoning, multimodal understanding, and full-stack web development. These capabilities allow the model to understand more complex contexts across different types of data and programming languages.

In the realm of web development, the model gained 89 Elo points on WebDev Arena. This brought its score to 1,768, placing it in fourth place on the leaderboard. These scores reflect the model's ability to generate functional and high-quality code for web-based applications and user interfaces.

The model also shows progress in reducing common AI errors. Internal testing showed that high-severity misalignment alerts fell by nearly 80% compared to the Argon model. Furthermore, the company reported that Carbon reduced model overreach by 26% and under-service by 29% during testing.

Safety and phased rollout

Google is taking a cautious approach to releasing Carbon to the public. The company is currently participating in a voluntary pre-release model testing process with the U.S. government. Initially, access is being granted only to a specific subset of cybersecurity defenders to gather feedback and refine safety protocols.

The model will eventually be available to a wider audience through Google's Fairwind Program. Once the company has gathered enough data and refined the safeguards, it will expand access to broader groups of trusted partners. Only after this phase will the model be made available to the general public, including businesses and consumers.

Safety metrics also show significant improvements in the new model. For example, successful cheating on Terminal-Bench 4.0 was reduced to 0%, while attempted cheating fell to 0.4%. Google stated the model is intended to combine autonomous engineering capabilities with these stronger safeguards as it moves toward a wider release.

Pricing and availability

For those who eventually gain access, Google has set a clear pricing structure for Gemini 4 Carbon. The costs are identical to those of the Gemini 4 Argon model. This includes a price of $2 per million input tokens and $10 per million output tokens for the service.

To make the model more efficient for large-scale tasks, Google is offering a discount on cached data. Cached input tokens will cost 95% less than the standard input-token price. This is particularly useful for long-horizon tasks where the model may need to reference large amounts of previous information repeatedly.

The rollout remains a gradual process focused on security and reliability. By starting with cybersecurity experts, Google aims to ensure the model can handle high-stakes engineering tasks before it reaches the general market. The goal is to provide a tool that can work autonomously while maintaining high standards of safety and accuracy.