Gemini 4 Carbon Introduced as Google’s Latest Frontier AI Model With Enhanced Cybersecurity Capabilities

Gemini 4 Carbon was introduced by Google on Wednesday as its latest frontier AI model. The Gemini 4 Carbon model arrives weeks after the tech giant released the Gemini 4 Argon. Google's latest frontier model is currently available to a select group of cybersecurity experts. Google is also engaging authorities for voluntary pre-release model testing. However, the tech giant plans to release Gemini 4 Carbon to general users, developers, and enterprises. Gemini 4 Carbon outperforms the Argon model on multiple fronts, including coding and reasoning for complex, long-horizon tasks. It is also better at finding cybersecurity vulnerabilities.

Gemini 4 Carbon Pricing, Availability

Google has announced that the Gemini 4 Carbon model will be available at the same introductory price as Argon. The AI model is being offered at $2 (roughly Rs. 191) per million input tokens and $10 (about Rs. 953) per million output tokens. Cached input tokens for Gemini 4 Carbon and Argon are priced at 95 per cent off the input token price.

The Mountain View-based tech giant is currently working with the US government on a voluntary pre-release model testing process. The company will start releasing the Gemini 4 Carbon AI model to a “subset of cyber defenders”. Google also says it will gradually offer phased access to “additional trusted partners” through its Fairwind Program. Soon, developers, businesses, and consumers will access Gemini 4 Carbon through different Google products and services.

Gemini 4 Carbon Capabilities

The Gemini 4 Carbon AI model brings enhanced cybersecurity capabilities. Google claims that, in internal benchmarks, the frontier model achieved more than a 2x uplift in success rate for generating secure code over Gemini 4 Argon. Further, the company says its Chrome team said Gemini 4 Carbon delivered more than 10x better performance in generating high-quality patches than a competitor. Google's Cloud security platform, Wiz, claimed it observed that Gemini 4 Carbon outperforms other AI models on various web penetration testing benchmarks.

On SEC-bench Pro, which tests how well an AI model can autonomously detect vulnerabilities, the Gemini 4 Carbon model managed to score 88.2 percent, outperforming OpenAI's GPT-6 Astra and Gemini 4 Argon, which scored 85.4 percent and 83.4 percent, respectively. Similarly, on ExploitGym, which tests risk validation capabilities of AI models, Gemini 4 Carbon scored 54.8 percent, beating GPT-6 Astra's 42.4 percent score.

To test how well the AI model patches the root cause, Gemini 4 Carbon was tested on Collinear's CWE-bench, where it scored 71 percent, which is higher than Grok 4.7, Gemini 4 Argon, and GPT-6 Astra AI models.

In addition, the tech giant mitigated shortcut attempts during training to ensure high-quality coding performance for Gemini 4 Carbon as a security and safety measure. Google also claims Gemini 4 Carbon cut unsupported claims by about 2x compared with Argon during internal testing. It is also better at resisting prompt injection, scoring 0.0 percent attack success rate on Gray Swan's Indirect Prompt Injection (IPI) Benchmark, an improvement over Argon's 0.7 percent score.