Google Unveils Gemini 4 Carbon, Topping a Major AI Benchmark

Google has released Gemini 4 Carbon, a new frontier AI model that has taken the top spot on one of the most-watched independent AI benchmarks. Unlike most launches, though, it is not opening to everyone right away.

Top score on the Intelligence Index

According to independent testing by Artificial Analysis, reported by Tom’s Hardware on November 3, Gemini 4 Carbon scored 59 on the Artificial Analysis Intelligence Index V5, narrowly ahead of Anthropic’s Claude Opus 5.5 at 58 and OpenAI’s GPT-6 Astra at 57. It ranked first on four of the index’s nine evaluations.

It also appears to be efficient: Carbon used about 123,000 output tokens per task, 29% fewer than Claude Opus 5.5 in the same tests.

Best knowledge score, but not perfect

On the AA-Omniscience knowledge benchmark, which rewards correct answers and penalizes made-up ones, Carbon achieved the highest score of any model at 54 on a scale from -100 to +100. Its accuracy rose to 58% from 48% for the earlier Gemini 4 Argon. However, its hallucination rate also rose, to 11% from 4%, so even the best model is far from always right.

Google says Carbon is designed to decline tasks it cannot complete instead of inventing answers, and to show sources and suggested actions more often.

Not the best at everything

The same analysis found that Carbon trails Claude Code with Opus 5.5 at programming and lags several rivals on some knowledge-work tasks. At lower price points, open-weight models from Chinese labs still offer better value.

A careful rollout

Instead of a public launch, Google is first giving access to selected “cyber defenders” through its Fairwind program, with wider availability through trusted partners later. Google says it ran voluntary pre-release testing with the U.S. government and wants to strengthen its cybersecurity safeguards before opening the model to developers. “We’re taking a deliberate approach to rollout,” the company said.

What it means for you

• Most users cannot try it yet. Keep using today’s Gemini, ChatGPT or Claude models for everyday work.
• Benchmarks are close. A one-point lead rarely changes real-world results; the best model depends on your task.
• Always check facts. Even the top-scoring model still makes mistakes.