AI NewsWaveGoogle Unveils Gemini 4 Carbon With Stronger Security and Long-Running Task Abilities
Key points
- Google unveiled Gemini 4 Carbon on the 28th
- Stronger coding, cybersecurity and long-running task capabilities
- Follow-up model two weeks after Gemini 4 Argon
Google unveiled Gemini 4 Carbon, its latest AI model with stronger coding and cybersecurity capabilities, on the 28th (local time). Chosun Biz and Boannews (보안뉴스) reported that it came two weeks after the company introduced its frontier model Gemini 4 Argon. According to Boannews, Google described Carbon as its most powerful model.
Several outlets reported on its performance ranking. Digital Today, citing The Verge, reported that Carbon ranked first on the Artificial Analysis Intelligence Index (v5.0). AI Newspaper (인공지능신문) wrote that it took first place in the world with an intelligence score of 59, and GeekNews also noted that it ranked first with 59 points on the same index. AI Times reported that Carbon topped Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Astra, and that Google had retaken first place in the overall performance ranking after one year.
Each outlet stressed slightly different points. GeekNews summarized that Carbon is designed to keep carrying out complex multi-step engineering tasks and has stronger coding and cybersecurity defense capabilities. Boannews reported that it sharply improved cybersecurity and long-running task capabilities. AI Newspaper observed that global AI competition is expanding beyond model reasoning performance into agentic AI, meaning software development, cybersecurity and tasks carried out autonomously over long periods.
News of external testing came before the release. AI Times reported on the 24th that word had emerged that Gemini 4 Carbon, which Google was developing, had entered testing with outside partners. The same report said output from a model believed to be Carbon appeared about a week after Argon's release, and early users judged its 3D work to be improved compared with Argon. Comments that processing time was long were also raised. This was a pre-release user assessment and cannot be assumed to match the performance of the released model.
Glossary Agentic AI refers to AI that takes a goal and plans and carries out multi-step tasks on its own, without a person giving instructions at each step.