Google's Gemini 4 Carbon claims top scores in coding and cybersecurity

[ForumGremlin]

Google has released another frontier AI model just two weeks after the last one. Engadget reports that Google has announced Gemini 4 Carbon, which it describes as a significant leap in AI coding and cybersecurity

Google says Carbon scored 59 on the Artificial Analysis Index version 5.0, the highest listed score, just ahead of Claude Opus 5.5 and GPT-6 Astra. It claims a near sweep across agentic coding, computer use, knowledge work, and science and maths benchmarks

Cybersecurity is the headline claim. Google says Carbon beats rival models by fairly wide margins at finding, prioritising and fixing vulnerabilities at their root cause, and is twice as likely as its predecessor Gemini 4 Argon to write secure code. It also claims Carbon takes fewer shortcuts, is more honest and is more resistant to prompt injection

Access is being rolled out carefully. Selected cybersecurity defence companies get it first for pre-release testing, followed by trusted partners through Google's Fairwind Program, and only then the general public. No pricing or public release date has been given

All the benchmark claims come from Google and have not been independently verified. The staged release for security uses reflects growing concern about powerful models helping attackers as well as defenders. The pace of releases among the big labs is relentless right now

Are AI models now the best tool for finding security bugs? And is a staged release the right approach for powerful cybersecurity capabilities?

Googlewww.engadget.com