Google has introduced Gemini 4 Carbon, a new flagship large language model aimed at cybersecurity and complex software engineering. The release comes only two weeks after the company made its previous flagship, Gemini 4 Argon, broadly available, underscoring the speed at which major AI developers are updating their frontier systems.
Google said internal testing found that Carbon was more than twice as effective at generating secure code as an earlier model. The company also reported that Carbon surpassed GPT-6 Astra and Claude Opus 5.5 across several benchmarks, although those comparisons reflect Google's testing and the evaluation configurations it selected.
Cybersecurity is the model's central use case
Carbon's security capabilities are designed to cover several stages of software defense. Google said the model can identify weaknesses in complicated software systems, assess whether attackers could exploit them and repair the underlying causes. The company said Carbon also performed better than Argon when creating patches for vulnerabilities it found.
On SEC-bench Pro, an evaluation of an AI system's ability to locate weaknesses in complex software, Carbon set a new record, according to Google. The Chrome team reported that the model performed substantially better than another unnamed frontier model when generating high-quality patches.
Google also said Carbon achieved a record result on the Artificial Analysis Intelligence Index, a composite evaluation that combines more than a half-dozen large-language-model tests. The version released Wednesday adds two benchmarks and removes questions that Google described as too easy for leading models.
| Evaluation area | Google's reported result |
|---|---|
| Secure-code generation | More than twice as effective as the earlier model in internal testing |
| Artificial Analysis Intelligence Index | Record score in Google's testing |
| Live coding environments | Nearly 80% fewer misalignment alerts, according to internal teams |
| Unsupported claims | Approximately 2x reduction compared with Gemini 4 Argon |
Access begins with Google's internal safeguards
Carbon includes an improved version of the safeguards Google previously built into Argon to reduce hallucinations and help mitigate cyberattacks. Google said teams using the model in live coding environments reported a nearly 80% decline in misalignment alerts. It also said Carbon cut unsupported claims by approximately two times compared with Argon.
Those results are company-reported findings rather than an independent assessment. Google configured Carbon to avoid filling gaps in its knowledge with unsupported answers, according to Kavukcuoglu. When the model cannot provide a reliable response, the company said it may include methodological limitations in its output or withhold certain actions while explaining the reason.
Google's rollout places cybersecurity organizations at the front of the queue. The model is being introduced through Google's cyber-focused program, while the company has not indicated in the published announcement when Carbon will become broadly available to developers, businesses or consumers. That restricted deployment gives Google a way to evaluate the model's performance in security work before wider access.
- Gemini 4 Carbon is Google's new flagship AI model.
- The model is designed for vulnerability discovery, exploit assessment and software repair.
- Google reported stronger safeguards against misalignment and unsupported claims than in Gemini 4 Argon.
