Gemini 4 Carbon Tops the Leaderboard, But You Can't Try It Yet

Google released Gemini 4 Carbon on October 28. It's the new number one on the Artificial Analysis Intelligence Index, and almost nobody can use it. Both of those things are true at once, so I wanted to write down what we actually know.

The score. Carbon gets 59 on version 5.0 of the index. Claude Opus 5.5 is one point behind at 58, and GPT-6 Astra is at 57. That's close enough that I wouldn't read much into the order.

Where it wins. The big gains are in security work. According to Google, it scores 88.2% on SEC-bench Pro (finding real bugs in software), 54.8% on ExploitGym (exploiting them) and 71% on CWE-bench (fixing them). It also does well on Terminal-Bench 4.0 at 65.2%.

Where it doesn't. Google's own comparison table has Opus 5.5 ahead on computer use (87.2% vs 81.5% on OSWorld-2.1) and GPT-6 Astra ahead on FrontierSWE 2 (65.5% vs 61.2%). So it's not a clean sweep.

Why it's locked. Google says the security scores were measured with guardrails switched off. A model that's this good at finding and exploiting bugs is useful to attackers too. So access starts with a small group of cyber defenders this week, then widens through Google's Fairwind Program. No date has been given for regular developers.

Price. $2 per million input tokens and $10 per million output, the same as Gemini 4 Argon, which came out just two weeks earlier.

My take: the cautious rollout makes sense for a security-heavy model. But until it's open, it shouldn't change anyone's plans. If you're choosing a model today, the options you can actually call haven't changed.

Sources: Google's official blog, Google DeepMind's Gemini page, Artificial Analysis, The Verge, SecurityWeek.

Comments

Popular posts from this blog

Planning a small bakery product shot with Seedream 5.5: keep list first, prompt second

상세페이지 이미지 세 장을 한 세트로: Seedream 텍스트로 이미지에서 프롬프트 뼈대 쓰기

Seedance 3.0, Kling or Vidu: Choose Around the Job