On Wednesday, Google unveiled its latest cutting-edge AI model, Gemini 4 Carbon—arriving just weeks after the tech giant launched its predecessor, Gemini 4 Argon. Currently, this new model is available exclusively to select cybersecurity experts. Google is also collaborating with relevant organizations to conduct voluntary pre-release testing. The company intends to eventually roll out Gemini 4 Carbon to the general public, developers, and enterprise users. Gemini 4 Carbon outperforms the Argon model in several areas, including coding and reasoning capabilities for complex, long-horizon tasks, as well as in the identification of cybersecurity vulnerabilities.
Pricing and Availability
Google has announced that Gemini 4 Carbon will launch at the same price point as Argon. The model is priced at $2 per million input tokens and $10 per million output tokens. For both Gemini 4 Carbon and Argon, the cost for cached input tokens is 95% lower than the standard input token rate.
The Mountain View-based tech giant is currently working with the U.S. government on a voluntary pre-release testing process. The company will initially grant access to Gemini 4 Carbon to a select group of “cybersecurity defenders.” Furthermore, Google stated that it will provide phased access to “additional trusted partners” through its Fairwind program. In the near future, developers, enterprises, and consumers will be able to access Gemini 4 Carbon across Google’s various products and services.
Gemini 4 Carbon Capabilities
The Gemini 4 Carbon AI model significantly enhances cybersecurity capabilities. Google claims that, in internal benchmarks, this cutting-edge model achieves a success rate in generating secure code that is more than double that of Gemini 4 Argon. Furthermore, Google reports that its Chrome team found Gemini 4 Carbon to be over ten times more effective than competing models at generating high-quality patches. Wiz, a Google Cloud security platform, also noted that Gemini 4 Carbon outperformed other AI models across various web penetration testing benchmarks.
In the SEC-bench Pro test—which evaluates an AI model’s ability to autonomously detect vulnerabilities—Gemini 4 Carbon scored 88.2%, surpassing both OpenAI’s GPT-6 Astra (85.4%) and Gemini 4 Argon (83.4%). Similarly, in the ExploitGym test, which assesses an AI model’s capacity for risk verification, Gemini 4 Carbon achieved a score of 54.8%, outperforming GPT-6 Astra’s 42.4%.
To evaluate how effectively an AI model can remediate root causes, we tested Gemini 4 Carbon using Collinear’s CWE-bench; it achieved a score of 71%, outperforming the Grok 4.7, Gemini 4 Argon, and GPT-6 Astra AI models.
Furthermore, to ensure high-quality coding performance in Gemini 4 Carbon, the tech giant limited the use of “shortcuts” during the training process—a measure implemented as a safety precaution. Google also claims that, in internal testing, the number of instances where Gemini 4 Carbon issued a “refusal” (stating it could not fulfill a request) was roughly half that of Argon. It also demonstrated superior resilience against prompt injection, achieving a 0.0% attack success rate on Gray Swan’s Indirect Prompt Injection (IPI) benchmark, outperforming Argon’s 0.7%.