Hyderabad: Google DeepMind has announced Gemini 4 Carbon, its newest and most capable Artificial Intelligence (AI) model. It is built for complex engineering tasks and advanced cybersecurity work. Google describes Carbon as the strongest model the company has ever released, which already tops global AI benchmarks.
Pricing and availability
Gemini 4 Carbon is priced at $2 per million input tokens and $10 per million output tokens, which is the same introductory pricing as Argon. Cached input tokens are available at 95 per cent off the standard input price.
The model is currently being rolled out first to a select group of cybersecurity professionals. Access will expand gradually through Google's Fairwind Programme before becoming available to developers, businesses, and everyday users in the coming weeks.
What makes Gemini 4 Carbon different?
Carbon is the direct successor to Gemini 4 Argon, which launched just two weeks earlier. Despite the short gap, the new model represents a significant step forward. It ranks first on the Artificial Analysis Intelligence Index v5.0, scoring 59 — placing it ahead of every other AI model currently assessed on that scale.
In other words, Carbon is faster, smarter, and more reliable than its predecessor, Argon, particularly when it comes to handling long, difficult tasks that require sustained effort over time.
Built for complex, long-running tasks
Google says Carbon is designed to operate on harder and long-horizon engineering problems. It is built to conduct sustained agentic workflows. The tech giant's internal teams have reported that the model can run continuously for days, pushing through dead ends to fix tricky bugs buried deep within large codebases.
Carbon has been improved across coding, scientific tasks, full-stack web development, and multimodal understanding, meaning it can handle text, images, and other types of data together.
State-of-the-art cybersecurity capability
One of the headline features of the Carbon is its performance in cybersecurity. The model has been designed to support the full cycle of digital defence. It includes finding weaknesses in software, checking whether those weaknesses can be exploited, and then writing code to fix them.
On SEC-bench Pro, which tests a model's ability to find real vulnerabilities in real-world software, Carbon scored 88.2 per cent — the highest of any model evaluated. On ExploitGym, which tests exploration — checking whether a flaw can actually be used by an attacker — it scored 54.8 per cent, again setting a new record. On CWE-bench, which measures how well a model can patch vulnerabilities at their root cause, Carbon reached 71.0 per cent.
Google's Chrome team also reported that Carbon performed more than 10 times better than a rival frontier model when generating high-quality security patches. Cloud security platform Wiz noted similar gains in web penetration testing.
Safer and more secure model
Alongside raw performance, Google has focused on making Carbon more trustworthy in how it behaves. The model has been trained to be more honest and transparent — cutting unsupported claims by roughly half compared to Argon.
It is also better at knowing its own limits. Rather than guessing or making things up when information is missing. Carbon flags the gap, surfaces the relevant sources, and asks the user before proceeding. High-severity alignment issues dropped by almost 80 per cent in internal testing.
Carbon also resists prompt injection attacks — attempts by malicious inputs to hijack the AI's behaviour — achieving a perfect 0.0 per cent attack success rate on an industry benchmark, improving on Argon's already strong 0.7 per cent.