Google Rolls Out Gemini 4 Carbon to Selected Security Defenders
Google says its latest AI model improves vulnerability analysis and complex coding tasks, but the initial rollout is restricted rather than public.
SAN FRANCISCO — Google has unveiled Gemini 4 Carbon, a new AI model with a limited initial rollout to selected cybersecurity teams. The 28 October announcement emphasizes the model's ability to investigate software flaws and work on complex coding tasks, but it is not a general public launch.
Announced on Wednesday and drawing fresh attention on Thursday, Carbon is designed for long-running coding and engineering tasks as well as security work. Google says it performs better than Gemini 4 Argon, introduced earlier this month, on several internal and external evaluations.
The important distinction for prospective users is access. Carbon is not being released immediately to everyone with a Gemini account. Google says selected security organisations will receive early access, followed by additional trusted partners through its Fairwind programme.
What Google is claiming
In its technical launch post, Google reports an 88.2% score on SEC-bench Pro, a test involving software vulnerabilities, and 71.0% on CWE-bench, which assesses repairs to security weaknesses. Those are the company's reported evaluations, not a guarantee of production performance. Google also says Carbon is substantially better at producing secure code in its own evaluations.
Those are meaningful claims for security researchers, but benchmark scores do not guarantee identical results in ordinary production environments. Performance may vary with the codebase, the tools granted to the model and whether its proposed changes are checked by engineers.
The same capabilities that help defenders validate risks could be misused if deployed without controls. That dual-use concern is part of the stated rationale for the restricted initial release. Google says it is testing safeguards and taking part in a US government process for pre-release evaluations.
Pricing and broader release
Google listed an introductory price of $2 per million input tokens and $10 per million output tokens, matching the introductory price of Argon. It has not given a firm date for general consumer or enterprise availability.
A key commercial question is whether the model can handle the kinds of lengthy software tasks that often defeat shorter AI demonstrations: tracing bugs across multiple services, making a safe fix and checking that the result does not break something else.
Access remains limited to the groups described by Google, and no verified date for general availability was provided. Readers should distinguish benchmark performance from independently established results in their own codebases.
Sources
Musa Ersoy
Musa Ersoy is an editor and writer at DEFUXA Global News, covering developments in the United Kingdom and around the world. He researches news stories using official statements, credible sources and independently verifiable information. His editorial work focuses on factual ac…
Author profile and articles