Google introduced Gemini 4 Carbon on Wednesday. Gemini 4 Argon is barely two weeks old, and the successor is already here, built around something no frontier model has leaned on this hard before: cybersecurity.
What Carbon puts on the board
Carbon tops the Artificial Analysis Intelligence Index at 59 on the new v5.0 of the test. That version shipped the same day, adds two benchmarks and drops several saturated ones.
The security numbers are the real story:
- SEC-bench Pro: 88.2% for finding vulnerabilities in real-world software
- ExploitGym: 54.8% for exploitation
- CWE-bench: 71.0% for patching
Koray Kavukcuoglu, SVP at Google DeepMind, writes that the Chrome team saw Carbon perform more than ten times better than another lab’s large frontier model at generating high-quality patches. Internally, the success rate for producing secure code doubles against Argon.
One figure is easy to skip past. On Terminal-Bench 4.0, successful cheating falls to 0.0%, with attempts down to 0.4%. In Google’s own live coding environments, high-severity misalignment alerts dropped by almost 80%, and unsupported claims were cut roughly in half.
Who actually gets it
Almost nobody, for now. Starting this week Google hands Carbon to a subset of cyber defenders and will phase access to further trusted partners. Developers, businesses and consumers come later. The company says it is taking part in the US government’s voluntary pre-release testing process.
Pricing stays where Argon started: $2 per million input tokens, $10 per million output tokens, cached input at a 95% discount.
In web development Carbon jumps 89 elo points to #4 at 1,768. For long work, Google lets its own teams describe it: the model keeps running for days on multi-step engineering problems instead of giving up at dead ends.
Access is the news here, not the benchmark
Two weeks between frontier models is a cadence only press releases enjoy. The ordering interests me more. Google ships the model to defenders first and to everyone else second, and a model that finds 88% of vulnerabilities finds them for both sides of that fence.
For Claude users the picture has turned awkward. SiliconANGLE reports that Carbon beats GPT-6 Astra and Claude Opus 5.5 across multiple benchmarks. While the model sits with a small circle of partners, though, that comparison stays a number on a slide. Checking it takes hands on the thing.