Google DeepMind Launches Gemini 4 Carbon, Cybersecurity Focused AI Model
Google DeepMind announced the Gemini 4 Carbon on October 28, 2026 — billed as their most powerful model — just two weeks after the Gemini 4 Argon. The announcement was widely discussed on social media.
AI creates a new explanation from RingkasNow facts and source metadata—it does not copy or fully translate the source article.
News explanation
What happened
Key facts
Why it matters
Context
Who is involved / affected
Gemini 4 Carbon is focused on cybersecurity and long-term agent coding capabilities. According to Google, this model is ranked first in the Artificial Analysis Intelligence Index v5.0 with a score of 59, above Claude Opus 5.5 (58) and GPT-6 Astra (57).
Google claims a number of benchmark results: SEC-bench Pro 88.2% for finding vulnerabilities, ExploitGym 54.8% for exploit testing, and CWE-bench 71.0% for patching. Secure code generation is said to be more than twice as good as Argon, with an indirect prompt injection attack success rate of 0.0% in Gray Swan testing. Please note, these numbers are internal Google claims that have not been reproduced independently.
The model was rolled out in stages through the Fairwind Program in partnership with more than 650 institutions — governments, cyber authorities, and critical infrastructure — before being made available in Vertex AI and Google AI Studio. The introduction price is 2 US dollars per million input tokens and 10 US dollars per million output tokens.
On social media, this announcement sparked debates about benchmarks, complaints about limited initial access, and jokes about model naming schemes inspired by the periodic table.
What to monitor next
Gemini 4 Carbon menjadi model frontier terbaru Google di tengah persaingan AI yang kian ketat.
Ask AI about this story
Something still unclear? Ask a specific question. Answers are limited to recorded source evidence.
AI may make mistakes. Check original reporting before making decisions.