Google's Gemini 4 Carbon grabs the No. 1 AI spot and a security edge

Google has a new flagship model, and it went straight to the top of the board. Gemini 4 Carbon landed a score of 59 on the Artificial Analysis Intelligence Index, the ranking that sorts the leading AI models by overall smarts, according to @BullTheoryio. That score puts it at number one, ahead of Claude Opus 5.5 at 58 and OpenAI's GPT-6 Astra at 57.

The pitch from Google is endurance as much as raw intelligence: the company says the model can grind on a complex coding problem for days without tapping out. But the flashiest number in the launch isn't about code at all. It's cybersecurity, where Gemini 4 Carbon apparently runs away from the field.

The model was put through the three core stages of defending software. When hunting for security flaws in real software, it found them 88.2% of the time. Asked whether those flaws could actually be exploited in an attack, it scored 54.8%, well clear of GPT-6 Astra at 42.4% and Claude Opus 5.5 at 34.5%. And when it came time to patch the holes, it wrote working fixes 71.0% of the time.

Google's Chrome team says the model made patches more than 10 times better than another lab's top offering, per @BullTheoryio. That's a bold internal claim, and it lands as AI labs increasingly sell security chops as the reason to pick their models over a rival's.

The catch: Google is keeping the keys close. The company says Gemini 4 Carbon is so effective at finding flaws and turning them into attacks that only a small number of cybersecurity groups are cleared to use it for now. For everyone else, the wait continues.

Add it up and Google gets a rare double: the highest overall intelligence score on the board, plus a locked-down security tool it can hand-pick customers for. Rivals OpenAI and Anthropic now have a benchmark to chase, and the next round of model launches will show how fast they close the gap.