Google releases Gemini 4 Carbon for complex coding and security

A new frontier for reasoning

Google DeepMind has introduced Gemini 4 Carbon, a model specifically engineered to manage intricate coding jobs and cybersecurity threats. This new release represents a significant step forward in long-horizon capabilities, which refers to the ability of an AI to stay on task over extended periods. Unlike models that might give up quickly, Carbon is built to navigate through difficult engineering obstacles that require sustained effort.

The model's performance has already been validated by third-party testing. As Yahoo Tech reported, Carbon secured the top position on the Artificial Analysis Intelligence Index v5. It achieved a score of 59, surpassing other major models including OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5. This ranking highlights its strength in complex reasoning compared to its peers.

While it is a major leap, the model is not perfect in every single category. It still trails some rivals on certain specific coding benchmarks. However, its overall trajectory is positive, particularly in how it handles large-scale codebases. Google's internal teams have observed the model working on these big projects for days at a time to reach a solution.

Persistence in the codebase

One of the defining characteristics of Gemini 4 Carbon is its persistence when facing 'messy' coding tasks. Google notes that the model can push through various dead ends to find hidden bugs and develop fixes across massive amounts of code. This behavior is a key differentiator from the previous Gemini 4 Argon model, which served as its foundation.

To demonstrate this endurance, Google points to the WebDev Arena leaderboard. On this platform, Gemini 4 Carbon saw a significant jump of 89 points. This suggests that the model is becoming much more effective at navigating the nuances of web development and complex software architecture. It is designed to act more like an agent that can follow a workflow rather than just a chatbot.

The development of Carbon involved specific improvements in reasoning and agent-based workflows. By refining these areas, Google aims to create a tool that can handle the 'long-horizon' nature of software engineering. This means the AI can plan out a multi-step fix and execute it without needing constant human intervention at every single turn.

Cybersecurity and vulnerability

Cybersecurity is positioned as the primary selling point for the Gemini 4 Carbon model. It has been tested against several industry-standard benchmarks to see how well it can identify and repair security flaws. On the SEC-bench Pro, the model scored 88.2% for finding vulnerabilities, showing a high level of proficiency in spotting risks.

The model also performed well in other security-related tests. It achieved a score of 71% on CWE-bench, which measures the ability to fix flaws, and 54.8% on ExploitGym for testing how exploitable those flaws actually are. These numbers provide a concrete look at how the model can be used by defenders to harden their systems.

By focusing on these specific metrics, Google is targeting a professional audience. The goal is to provide a tool that can automate the tedious parts of security auditing. Instead of a human having to manually scan every line of code for a specific type of error, Carbon can be tasked with finding and fixing those issues across the entire codebase.

Safety and reliability updates

Google has implemented several changes to make Gemini 4 Carbon more reliable than its predecessor, Argon. One of the main goals was to reduce unwanted behaviors that can plague large language models. These behaviors include the tendency to invent information, known as hallucination, or taking shortcuts that bypass user instructions.

The company worked to ensure the model stays within the bounds of what a user actually asks it to do. By making these adjustments, Google hopes to provide a more stable experience for professional users who require accuracy. Reliability is a critical factor when the AI is being used to handle sensitive tasks like cybersecurity or large-scale software updates.

These safety improvements are part of the broader effort to make Gemini 4 Carbon a 'frontier' model. A frontier model is one that pushes the boundaries of what AI can do while maintaining a level of control. Google is balancing these high-level capabilities with the need for the model to follow specific, safe instructions without deviating into unpredictable actions.

Access and pricing details

Gemini 4 Carbon is not yet available to the general public in its full capacity. Currently, it is being offered to select cybersecurity defenders through a specific initiative called the Fairwind Program. This limited rollout allows Google to evaluate the model's real-world performance and safety before a wider release.

For those who do eventually get access, Google has established an introductory pricing structure for its API. The starting rates are $2 per million input tokens and $10 per million output tokens. These costs are intended to be lower during the initial phase to encourage adoption and testing by developers.

Google has indicated that these prices will eventually increase to $4 per million input tokens and $20 per million output tokens. However, the company has not yet announced the specific date for this price hike. The timeline for broader availability will depend on how Google evaluates the model's safeguards and its performance in various practical applications.