Gemini 4 Carbon reportedly tops Artificial Analysis Intelligence Index v5
A secondary report citing Artificial Analysis says Gemini 4 Carbon placed first in the Intelligence Index v5 with 59 points, leading four of nine evaluations: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA. Carbon also scored 59 on the coding-agent index, three points behind Claude Opus 5.5, while its reported hallucination rate rose from Argon’s 4% to 11%. The report says Carbon remains limited to selected users, with no public availability timetable.
The reported results position Carbon as a leader in complex reasoning, scientific tasks and knowledge reliability, but not across every evaluation dimension. Its higher hallucination rate, comparatively strong coding score and reported per-task cost of $7.76—below Claude Opus 5.5’s $10.04 but above GPT-6 Astra’s $6.25—make workload-specific trade-offs and limited availability important considerations.
榜单第一名的模型,你现在花钱也买不到。
10 月 29 日,第三方评测机构 Artificial Analysis 把谷歌的 Gemini 4 Carbon 放上智能指数 v5 的榜首,59 分,压过 Claude Opus 5.5 和 GPT-6 Astra。Hacker News 上这条消息拿到三百三十八个赞、两百四十七条评论,但被顶得最高的讨论不是它有多强,而是它到底什么时候才开放。和上一代 Argon 发布时一样,Carbon 目前只推送给选定用户,公开时间表根本不存在。更有意思的是数据里藏着一个反常识的细节:新模型的幻觉率从 4% 涨到 11%,翻了将近三倍,可这个数字放到全场对比,依然是第二低。这篇文章把官方评测数据拆开,看看这个榜首到底值多少钱,以及你我现在该做什么。
榜单长什么样:九个单项拿下四个第一
先看总分。Artificial Analysis 的智能指数 v5 用九个评测项给模型打分,Carbon 拿到 59 分成为新榜首。这九个单项里它拿了四个第一,分别是 HLE-Diamond、Terminal-Bench-Science、AA-Omniscience 和 AutomationBench-AA,覆盖复杂推理、科学任务和知识可靠性这些硬骨头领域。这个指数由第三方机构维护,九个维度独立计分,目的就是避免单一 benchmark 被针对性刷分。评测方给出的定性是,它在复杂推理、知识和科学任务上领先,编码和知识工作也有大幅进步,但并不是所有维度都领先。
#Gemini 4 Carbon#Google#Artificial Analysis#Claude Opus 5.5#Claude#GPT-6 Astra#GPT#Gemini 4 Argon#Antigravity CLI#Claude Code