Gemini 4 Carbon: Google's Strongest Model Goes to Cyber Defenders First
Google announced Gemini 4 Carbon last week with top scores in security tasks, limited access and mixed independent results on accuracy.
Last week, Google announced Gemini 4 Carbon, which it calls its next frontier model, meaning its most capable one. Most people cannot use it yet. A small group of cybersecurity defenders gets access first, followed by trusted partners through Google’s Fairwind program, and developers, businesses and consumers later, with no date given. Its sibling Gemini 4 Argon is available now.
On paper the model does well. Google reports top scores in 15 of 17 comparison rows, including 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for checking whether an exploit works, and 71.0% on CWE-bench for patching. Independent figures from Artificial Analysis broadly agree but show a narrower lead: Carbon scores 59 on its Intelligence Index, against 58 for Claude Opus 5.5 and 57 for GPT-6 Astra. It trails Opus 5.5 with Claude Code on a coding-agent index (59 versus 62). The introductory price is $2 per million input tokens and $10 per million output tokens (tokens are the small chunks of text models bill by), rising later to $4 and $20, with no end date given for the intro rate.
What’s behind it: The restricted rollout is the interesting part. A model that is good at finding and fixing security holes is, by the same skill, good at attacking. Giving defenders a head start is a way to let them patch before the tool is widely available, which is also why Google’s tests were reported without guardrails and why the limited release is a sensible hedge. There are open questions. Google has not published a technical report, parameter count or API details. Its own results come without uncertainty estimates. And accuracy claims conflict: Google says unsupported claims roughly halved, while Artificial Analysis measured hallucinations (confident wrong answers) rising from 4% to 11% on one test.
What this means for you: Today, nothing, because you cannot use it. When it arrives, treat the headline rankings as a starting point and try it on your own tasks; small leads rarely decide real work. If you run websites or software, the broader message is that security tools powered by AI are improving quickly, so keeping your software updated matters even more. Also note the price: an introductory rate that later doubles is worth planning around.
Sources
"Egoless Mathematics": A Guest Post on Terence Tao's Blog Asks What Maths Is For in the AI Era
Three authors argue on Terence Tao's blog that AI should push mathematicians to value understanding over output, rank or credit.
