Artificial Analysis, an independent firm that regularly benchmarks large language models, has released version 5 of its Intelligence Index and reports that Gemini 4 Carbon, a new model from Google, now sits at the top of the ranking. The original announcement is brief and does not break down the scores across the individual sub-benchmarks that make up the composite index.
The Intelligence Index combines results from several benchmarks covering reasoning, coding, mathematics, and multimodal understanding, among others, to produce a single score meant to capture general model capability rather than performance on any one task. Version updates to this index happen periodically as new benchmarks are added and older, saturated tests are dropped — a methodology change that can itself shift rankings even without any new model being released.
The "Carbon" designation appears to refer to a specific variant within Google's Gemini 4 family, though technical details such as model size, architecture, context window, or inference cost are not specified in the available source. Leaderboards of this kind are closely watched by developers and businesses selecting models for their applications, but they warrant some caution: topping a composite index does not necessarily translate into superiority across every task or in real-world production settings.
The announcement fits into an ongoing pattern of competition among major labs — Google, OpenAI, Anthropic, and others — which routinely ship updates to their flagship models, each accompanied by fresh performance claims on public benchmarks. Without further detail on the actual scores, the real magnitude of Gemini 4 Carbon's lead over previously top-ranked models remains unclear.