2026年10月28日，分析机构Artificial Analysis发布了智能指数（Intelligence Index）第五版，谷歌推出的Gemini 4 Carbon以59分的总成绩位列榜首 1。该模型在复杂推理、知识储备及科学任务方面表现领先 1。尽管目前仅向部分用户开放，但其在编码代理领域的实力同样强劲，虽以59分的成绩略低于Claude Opus 5.5 (max)的62分，仍稳居该细分榜单第二 1。
在成本与性能平衡方面，Gemini 4 Carbon每个任务的平均费用为7.76美元，这一价格比Claude Opus 5.5 (max)低约23%，但高于GPT-6 Astra (max)的6.25美元 1。该模型拥有100万令牌（1M tokens）的上下文窗口，标准定价设定为每百万输入/输出令牌4美元和20美元，并正在提供50%的发布折扣，其中缓存输入令牌可享受高达95%的优惠 1。值得注意的是，其标准定价与Gemini 4 Argon相同，但实现了显著的性能跃升 1。然而，在AA-Omniscience v1.1测试中，Gemini 4 Carbon得分为54，伴随11%的幻觉率，相比之下，前代Argon的幻觉率为4% 1。
On October 28, 2026, Artificial Analysis released the updated version of its Intelligence Index, where Google’s newly launched model, Gemini 4 Carbon, secured the top position with a score of 59 1. The model demonstrates leading performance in complex reasoning, knowledge tasks, and scientific benchmarks, establishing itself as the current leader on the v5 index 1. While it ranks second on the Coding Agent Index with a high-tier score of 59, it trails behind Anthropic’s Claude Opus 5.5 (max), which achieved 62 points 1.
In terms of operational metrics and cost-efficiency, Gemini 4 Carbon supports a context window of 1 million tokens 1. Its average cost per task stands at $7.76, representing approximately a 23% reduction compared to Claude Opus 5.5 (max), though it remains higher than GPT-6 Astra (max) at $6.25 1. Regarding reliability benchmarks, the model scored 54 on the AA-Omniscience v1.1 test with an hallucination rate of 11%, a notable increase from the 4% rate recorded by its predecessor, Argon 1. Currently available only to select users, Gemini 4 Carbon is offered at standard pricing identical to Gemini 4 Argon—$4 per million input tokens and $20 per million output tokens—but includes a temporary 50% launch discount and a 95% discount for cached input tokens 1.
评论
还没有评论，欢迎留下第一条。