Most of us give up on a bug around the time the coffee runs out. Google says its newest model kept going for days. Meet Gemini 4 Carbon, the coworker who never asks for a stand-up meeting and apparently never sleeps.
Two Weeks After Argon, Here's Carbon
On October 28, Google announced Gemini 4 Carbon, calling it "our strongest model with state-of-the-art cybersecurity and long horizon capabilities." It arrives just two weeks after Gemini 4 Argon. Google says it's built for "sustained agentic workflows," and that internally Carbon has "run for days to persist through dead ends" while root-causing and patching bugs in large codebases.
The benchmark numbers are big. Google reports 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym, 71.0% on CWE-bench for patching, and the top spot on the Artificial Analysis Intelligence Index at 59. It also claims more than 2x improvement over Argon in secure-code generation. Introductory pricing is $2 per million input tokens and $10 per million output, and a footnote says that doubles after the intro period.
Good Guys First, Everyone Else Later
You can't just sign up and start pointing Carbon at things. Google is releasing it first to "a subset of cyber defenders," then to trusted partners through its Fairwind Program. Developers, businesses, and consumers get it "as soon as possible," with no date attached. Google also says it's taking part in the U.S. government's voluntary pre-release model testing. A model that scores well on exploitation benchmarks is useful to defenders and tempting to attackers, so the gated rollout is the grown-up move here, not just a marketing velvet rope.
The detail most people skip past is the honesty work. Google says Carbon cuts unsupported claims by about 2x over Argon, and that successful cheating on Terminal-Bench 4.0 dropped to 0.0%. If a model is going to work unattended for days, "doesn't make things up and doesn't game the test" matters more than another leaderboard point. An agent that grinds through a codebase all weekend is only helpful if it's honest about what it fixed on Monday. Google also says Carbon scored 0.0% attack success on the Gray Swan indirect prompt injection benchmark, which is reassuring for anything that reads untrusted web content while it works.
Keep the usual caveat in mind: these are Google's own numbers, from Google's own blog. Independent testing will tell us how much of this holds up in messy real-world codebases, which tend to be less tidy than benchmark suites.
The AI security arms race has officially moved from "can it find bugs?" to "how many days can it keep finding them?" Defenders get the head start this time, and they should use it.
If models like Carbon have you wondering where AI-assisted code review or vulnerability triage could fit in your own stack, we're happy to talk through what's practical today.
Source: Google Blog