Google's new Gemini 4 Carbon AI edges past rivals in tests

Google has unveiled Gemini 4 Carbon, a new AI model that scores 59 on a well-known independent test, one point ahead of Anthropic's Claude Opus 5.5. It also ranks first on a separate test of facts and made-up answers. Even so, the top score on that test is only 54 on a scale that runs from -100 to +100.

What changed

The tests come from a group called Artificial Analysis. On its main Intelligence Index, Carbon beats OpenAI's GPT-6 Astra by two points, Meta's Muse Wonder Full by nine and Claude Fable 5.1 by 10. The article calls the release a comeback for Google, which had put out models only occasionally for much of the year.

On the second test, AA-Omniscience, Carbon scores 54. Claude Opus 5.5 scores 48. The test tracks how well a model recalls information and how often it invents answers.

The catch on made-up answers

Google says Carbon cuts unsupported claims twice as much as Gemini 4 Argon, which it released two weeks earlier. Google also says Carbon will decline to act rather than invent an answer. The test data is more mixed. Carbon answers more questions correctly than Argon (58% versus 48%). But its hallucination rate, meaning how often it gives a wrong answer instead of holding back, is higher (11% versus 4%). Claude Opus 5.5 and GPT-6 Astra are both at 50%.

What it costs

Running one task on the Intelligence Index costs $7.76 with Carbon. That is about 23% less than Claude Opus 5.5 at $10.04, and Carbon uses 29% fewer output tokens, the small chunks of text an AI writes. GPT-6 Astra is cheaper at $6.25. The open-weight DeepSeek V4.1 Flash 1017 costs about 13 times less per task, but it scores only 32.

Carbon is not the best at everything. It trails Claude Code with Claude Opus 5.5 at programming. It falls behind Grok 4.7, Muse Wonder and Kimi K3.1 Pro on knowledge work. At lower price points, Chinese open-weight models such as GLM 5.3-Flash and DeepSeek V4 Pro take over.

Who can use it

Not most people yet. Google is opening Carbon first to a group of cyber defenders, the security teams who protect systems from attacks. It will then widen access through trusted partners in its Fairwind Program. Developers come later, after Google tightens its safeguards against hacking uses. Google says it is taking part in a voluntary U.S. government process for testing models before release.

What to do

Treat the scores as a sign of progress, not proof of reliability. If you use AI for work or money decisions, check important facts against a trusted source before you act on them.