Google releases Gemini 4 Carbon, a cybersecurity-focused AI model
Google has released Gemini 4 Carbon, an AI model aimed at cybersecurity work such as vulnerability discovery, exploitation and software patching. The company says Carbon generates secure code at more than twice the success rate of its predecessor, Gemini 4 Argon, on internal benchmarks. Carbon scored 59 on the Artificial Analysis Intelligence Index v5.0 and leads that benchmark, according to The Verge. Google priced the model at $2 per million input tokens and $10 per million output tokens. Access is initially limited to a select group of cybersecurity defense companies, with wider availability planned through the Fairwind Program.
Key facts
- Carbon scored 88.2 percent on SEC-bench Pro for finding vulnerabilities, and 54.8 percent on ExploitGym for exploitation.
- Carbon scored 71.0 percent on CWE-bench for patching tasks, according to The Hacker News.
- Gemini 4 Argon was launched two weeks before Carbon, per Engadget.
- Google says Carbon is twice as likely to generate vulnerability-free code than Gemini 4 Argon.
Summarised with AI from 5 sourcesRead our AI policy
Why it matters
A model aimed at security tasks could change how defenders find and fix software flaws, if its benchmark results hold up outside Google's testing.
What changed
First published from CNA, Anadolu Agency, Ars Technica, The Verge and Engadget