Google’s Gemini 4 Carbon Is Its Strongest Security Model Yet. Almost Nobody Can Use It.
Google’s new frontier model tops harder benchmarks for finding and fixing security flaws — but access is limited to selected defenders, and its introductory price doubles soon.
1 What happened
Google has introduced Gemini 4 Carbon, its next frontier model, built for sustained coding work and cybersecurity. The company says Carbon retakes the top spot on updated, harder benchmarks, and the scores it cites are striking: 88.2 percent on SEC-bench Pro, which tests vulnerability discovery, and 71 percent on CWE-bench, which tests patching.
But you probably cannot try it yet. Access is limited to selected cybersecurity defenders and trusted partners while Google gathers feedback and tightens the model’s guardrails. A broader release will follow, but Google has not given a date. The pricing is also on a clock: during the introductory period, Carbon costs two dollars per million input tokens and ten dollars per million output tokens, with a 95 percent discount for cached inputs. When that period ends, those prices double to four and twenty dollars.
In plain English, Carbon is a model designed to do long, careful technical work — reading code, hunting for flaws, and writing fixes — rather than just answering questions. That is why Google is starting with the people whose job is defending networks, not with the general public.
2 Why it matters
First, security work is becoming a flagship use case for frontier models. Finding vulnerabilities in real software used to be slow, expert, human work. A model that scores this highly on discovery and patching benchmarks suggests AI can now do meaningful parts of that work — which cuts both ways, since the same capability in the wrong hands makes attackers faster too.
Second, the limited rollout is itself a signal. Google is treating Carbon as powerful enough to need guardrails first, feedback second, and wide release third. That cautious sequence is becoming the norm for the most capable models, and it means the gap between what labs announce and what everyone can use is growing.
Third, the pricing tells you who this is for. At two and ten dollars per million tokens, with a steep discount for cached inputs, Carbon is priced for teams running sustained, high-volume work, not casual chat. When the price doubles, smaller teams will feel it first, which is why the introductory window matters.
3 What to watch next
- Wider access Watch who gets in next, and when. A broader release date — and any changes to the guardrails — will show how confident Google is in the model’s safety.
- Real-world results Benchmarks are a promise, not proof. The test that matters is whether defenders using Carbon find and patch real flaws faster than before.
- The price jump Introductory pricing ends at some point. Watch whether the doubled price sticks, or whether competition forces Google to extend the discount.
- The rivals Every major lab is racing on coding and security models. Expect competing announcements, and judge them by independent results, not launch-day claims.
This post is based on Google’s announcement: “Gemini 4 Carbon: our next frontier model” — Google, October 29, 2026. Figures are as reported by Google.
This story led Episode 19 of the Daily Tech Briefing podcast — your daily dose of tech news in about 10 minutes, read by Melody, your AI anchor. Listen on Spotify or Listen on Apple Podcasts.