换一把尺子重排整个前沿：AA 智能指数 v5、上线即登顶的 Gemini 4 Carbon，与被摊开的 NVIDIA 反垄断调查

信源说明：本文基于 The AI Daily Brief（主持人 Mohammad）一期 Headlines + 主话题节目整理（视频 ID Dz4_db_rRDE，全长约 26.7 分钟，英文转写 14 段、0s–1604s 已完整读取，文中数字与引语均依英文转写整理；720s–960s 为赞助商口播段，未纳入正文）。全部事件、数字、股评与人物言论均为主持人陈述或节目引用的财报、社交媒体帖文与媒体报道，本文未做独立核实，归属按「主持人称／节目引用／当事人认为」分层标注；ASR 疑似错词以「转写疑作」标注，文末附校正表与完整转写原文。同系列前文：Google 组合式世界模型与 Gemini 4 Carbon 的预告见《Genie 4 把「全家桶」装进世界模型》，AI 泡沫与 capex 信号的上一轮讨论见《Anthropic IPO 前的泡沫辩论》，Kimi K3.1 与开源距前沿的口径见《开源距前沿五周、网络末日常态化》。

先说结论

这一期节目表面是三条新闻加一个基准更新，实际是同一个问题的四种切面：当前沿模型彼此咬得很紧、单点分数失去区分度时，行业用什么重新建立秩序。Artificial Analysis 的答案是换尺子——智能指数 v5 过半改用私有数据集、把编码与科学评测全面 agentic 化，一换尺子，开源与闭源的差距从 3 分拉大到 12 分，头部实验室与其余所有玩家之间出现了 10 分断层；Google 的答案是换节奏——Gemini 4 Carbon 距 Argon 发布仅两周就拿下新榜第一，「有新底座、就能用强化学习快速迭代」从口头逻辑变成了可观测的发布速度；微软和谷歌的财报则显示，云与 AI 收入仍在狂奔，但搜索货币化被自家 AI 模式稀释、capex 撞上电力与组件的物理上限；而 DOJ 对 NVIDIA 的调查从「绕开反垄断审查的 acquihire」嫌疑摊开成覆盖风投、销售与客户施压的全链条调查。对读者的可迁移判断：看模型排名先看尺子怎么变的，看大厂竞争先看迭代周期与物理约束，看监管先看调查范围是否在扩。

头条一：Gemini 4 Carbon——当 RL 迭代速度本身成为竞争力

节目称，Google 在 Gemini 4 Argon 发布仅两周后推出下一个前沿模型 Gemini 4 Carbon——主持人强调这是超大规模模型版本间隔最短的发布之一。逻辑此前在节目里反复讲过：同一基座意味着 Google 可以用强化学习快速迭代，而不必为每个新模型重跑漫长的预训练。Google CEO Sundar Pichai 在周三（节目口径）财报电话会上的表述被直接引用：「这种快速进展凸显了我们惊人的模型势头。我们正处在 AI 帮助我们提高训练与发布速度的节点上。因为 RSI，我预计未来 12 个月会继续加速。」

基准分数（节目引用的官方与 Artificial Analysis 数字）：编码上 Carbon 在 Terminal Bench 4.0 涨 8 分、Frontier Suite（转写名）涨 6 分；网络安全上 Exploit Gym 涨 12 分、CWE Bench 涨 3 分；计算机使用上 OS World 涨 7 分、Agents Last Exam 涨 8 分——八项关键基准中六项登顶，压过 GPT-6 Astra 与 Opus 5.5。在新的 AA 智能指数 v5 上，Carbon 以 59 分（新量表上跳升 9 分）成为登顶者；AA 编码指数上排第二，与 Sonnet 5.5 并列、落后 Opus 5.5 3 分。成本口径：节目引用 AA 测得每任务 7.76 美元，比 Opus 5.5 便宜 23%、比 Astra 贵 20%，处于前沿模型中游。主持人的判断是：两周前 Argon 发布时不少人认为 Google 错过了窗口——Opus 5.5 与 GPT-6.1 Sol（转写原作 GPT-610 Soul）已把前沿推到让 Google 与 Meta 一起落入第二梯队；这版 Carbon 若实测匹配分数，Google 是「真正回到了前沿竞赛」。

但开放节奏明显放慢，主持人称之为「不太好的部分」：Google 遵循（美国）政府的前置发布测试指南，采取所谓「审慎的 rollout」——本周先向一部分网络防御者开放，再扩展到网络防御项目里的可信伙伴；对开发者、企业与消费者的开放「在尽快完成护栏迭代后进行」。少数拿到早期访问的是评测方 Arena：其渲染与游戏设计测试视频被普遍批评「不如 Opus 5.5 漂亮」——主持人重申他并不认为 3D 设计是衡量 LLM 的合适标尺，但也承认在该用例上 Carbon 确实未到对手水平；Arena 的 Peter Gosted（转写原文如此）的总体印象则是「比 Argon 好很多，这次 Google 是来真的，希望他们保持这个发布节奏」。

两条内部信号被节目认为比分数更有信息量。其一，DeepMind 研究员 Zachary Nado 回应网友「Fable 5.5 还没来」时写「Gemini Redacted 也还没来」（Redacted 为遮蔽代称），X 用户 Chubby 据此评论「事情要变疯了——他们还有一个比 Carbon 更强的未发布 Gemini 模型，而 Carbon 今天已经 AA 智能指数第一」。其二，Google 后训练研究员 Xiao Zhang 描述的内部状态：不到一年前 Gemini 3 Pro 发布还是「全团队庆祝第一的 big bang 事件」，如今团队连续数周内部试用（dogfooding）、「几乎已经适应了新的智能水平」——大胆的代码重写被视为理所当然、整夜挂循环自动跑几十个实验去定位修复疑难 bug、用模型集群检查修复 RL 数据，「发布现在只是周期性的检查点」。主持人补充的社区情绪：X 用户 Benjy’s Hessen（转写原文如此）调侃「为什么 Google 突然像他们发明了 Transformer、拥有半个互联网的数据、坐在 TPU 海洋上一样狂发模型？哦等等，本来就是」；Latent Spaces 的 Twix（转写原文如此）则说 Gemini 在 AAv5 登顶让他「很久以来第一次打开 Antigravity」，若 Google 真的做成，「Meta、Grok（转写原作 Croc）和现在 GDM 的回归季就正式开始了」。主持人自己的立场：拿到模型前 suspend judgment，但「看起来有希望」这件事本身就已经让人兴奋。

头条二：双云财报的另一面——搜索被自家 AI 稀释，capex 撞上物理上限

微软与谷歌同季财报（节目引用的数字均为公司口径）：微软 Azure 同比增长 48% 至 330 亿美元——这是微软首次把 Azure 单列为一行收入，主持人认为在如此增速下这是个好决定；Azure 带动总营收增长 19% 至 923 亿美元，两项均小幅超出分析师预期。微软还披露 OpenAI 为 Azure 贡献了 100 亿美元收入、约占该项业务三分之一；财报后微软盘后一度涨 4.8% 创历史新高。谷歌云收入同比翻倍至 307 亿美元，几乎追平微软、远超华尔街的高预期；但「问题在其余一切」（Aptus Capital 的 David Wagner 语）：搜索收入增速降至 12%，较上季的 17%、再上季的 19% 大幅放缓；YouTube 增速连续第三个季度放缓至 9.2%；TPU 销售的利润率问题被高管在电话会上回避；总营收增长 25% 大体符合预期。Pichai 亲口确认了自我稀释机制：「值得注意，AI 模式查询的强劲增长产生了影响，因为我们在 AI 模式上看到的货币化率更低。」谷歌还因 SpaceX 持仓减值而在利润上 miss。乐观如 Dan Ives 试图定调「这是我们一直在等的云财报，比空头今晚要咬的疲软搜索数字重要得多」，但盘后股价仍一度暴跌 10%——部分分析师认为 Google 已到「被迫蚕食自家搜索业务」的时点，因为它不再是可靠的增长引擎。

主持人把更多的篇幅给了更慢的一条线：数据中心建设。两家公司今明两年 capex 指引均维持不变，但本季实际支出低于预期；谷歌指引明年 capex 增 50% 至 3000 亿美元（节目称约数），目前却在低兑现今年的数字；Oracle 与 Meta 处境类似——主持人据此提出「AI 建设可能已到达峰值速度」：不是需求不足，而是组件、电力、人力乃至社区支持的物理约束先到顶。这里有一个对「AI 泡沫论」的重要反证：空头一直把 capex 减速当作刺破泡沫的催化剂，其逻辑是需求疲软迫使超大规模厂商刹车；但当前看到的更像「厂商想继续加速却被物理约束卡住」。（此段有一句转写损坏：科技分析师 Luke Longo（转写疑作 Luke Lango）「认为减速更多在于融资」的完整观点在转写中断裂重复，本文不引申。）主持人预告这是趋势早期，谷歌下季财报会深入谈 capex。

头条三：DOJ 对 NVIDIA：从 acquihire 嫌疑摊开成全链条反垄断

节目引用路透社：美国司法部就 NVIDIA 收购 Hugging Face 130 亿美元（9 月宣布、预计 2027 年上半年完成交割）询问该交易是否会让 NVIDIA 控制AI 模型的关键分发节点。NVIDIA 当时的口径是 Hugging Face 仍将保持平台开放、继续与各类模型开发商、芯片厂商与云厂商合作。关键增量在于调查的全貌：《纽约时报》9 月曾报道 DOJ 调查 NVIDIA 与芯片创业公司 Groq（转写原作 Grok；xAI 的 Grok 并非芯片公司，依上下文疑作 Groq）的授权加人才协议——那是一笔 acquihire（收购式雇佣，转写原作 aqua hire）交易，此类结构可追溯到 Google 2024 年与 Character AI 的交易，因不构成正式收购而不触发自动反垄断审查、理论上交割更快；外界原本以为 DOJ 只是借 NVIDIA 之案整肃这一惯例。但路透社消息源本周称，调查比 Groq 交易更早、范围也宽得多：NVIDIA 对数据中心优化公司 SkedMD（转写原文如此，未核实）与 Run:ai（转写原作 RunAI）的收购也在范围内——而这两笔是传统收购，并非 acquihire；近期的问询还覆盖 NVIDIA 的风投布局、销售惯例，以及是否对客户施加压力。主持人的归纳：这已经不是一条「绕开审查」的窄案，而是 DOJ 对 NVIDIA 的自上而下全面反垄断调查——目前没有任何指控被提出、调查也可能以不作为收场，因此不构成 NVIDIA 行为不当的暗示，但「比大多数人以为的要更值得担忧」。

主话题：AA Intelligence Index v5——基准的猫鼠游戏与一次大重排

主话题回到本播客的「元问题」。主持人先亮立场：基准「天然有限但重要」——没有它们，就只能靠纯体感（vibes）区分模型；但也几乎不应把任何单一基准、甚至聚合分当作圣经。Artificial Analysis 刚刚把其智能指数升到 v5，等于给前沿 AI 换了一把新尺子。

**这把尺子是怎么进化到今天的。**节目回溯：在 agentic coding 时代之前，LLM 多半是「高级搜索/写作助手」，chatbot arena 这类体感榜就是最高检验，因为它告诉开发者用户想要什么；随着研究与编码成为主用例，更客观的基准出现。进化一直有两种动力：一是饱和——所有模型挤在榜顶时，出题方就加新题难题出迭代版，直到再次饱和；二是用例迁移——agentic 计算机使用成为模型能力的基本盘后，评测经历了更彻底的重构。AA 在每次变化中换基准、调权重，过程并不总是平滑：指数一度严重饱和失去区分度；「刷榜」（benchmark maxing，主持人定义为把评测题混进训练集，让分数反映「对正确答案的记忆」而非能力）一度让部分开发者排出远超真实体验的名次——主持人点名这是中国开源权重模型的「一贯故事」（节目口径）。最近的教训是 9 月初的 v4.2 提前更新：GPT-6 Astro 重仓计算机使用能力，而 v4.1 对 agentic 基准权重太轻，导致 Astro 分数「不仅令人失望、而且与大家的实际体验对不上」；更新后 Astro 仍略落后 Fable 5.1，但指数更贴近真实使用了。

**v5 改了什么。**节目引 AA 官方：v5 首次让过半基准建立在私有数据集上——理论上无法「对着考题训练」，直接打击刷榜。新增两条：Frontier Code 2.0（Cognition 私有的真实世界编码基准，评分标准是「仓库维护者会不会合并这段代码」）与 Terminal Bench Science（agentic 科研基准）；Humanity’s Last Exam 升级到更难的 HLE Diamond；AA Omniscience（知识与幻觉）和物理基准 Crypt PT（转写名，未核实）做了饱和化更新。出局的有：衡量真实场景白领工作的 GDPval-AA（转写原作 GDP Val Double A）、科研基准 SciCode、长上下文推理基准 AA LCR；后者由 AA Briefcase、Automation Bench AA 与 GDP PDF（转写名）沿更细粒度的分工接棒。

主持人认为这份增删清单本身就是一份「模型进化史」：v4 发布的今年 1 月，长上下文推理还是最关键能力——大代码库作业必需，而很多编码模型上下文窗口有限、深受 context rot 之苦；此后压缩（compaction）与记忆的进步让长上下文变得理所当然，于是长上下文基准出局。同理，指数内的编码基准现在只测 agentic 表现——agentic 已被固化为开发者使用 AI 的方式；科学基准也从考既有知识改为测真实 agentic 科研作业。权重上，AA 首次把指数在四个类别间等权（转写表述为 knowledge work、coding、science 与 general——后者涵盖通识知识与幻觉率，每类约 25%）；主持人猜测科学占 25% 是对 Anthropic 与 OpenAI 近来互相抢科学头条的前瞻性押注。

**重排结果（节目引用的 AA v5 分数）。**Gemini 4 Carbon 周三空降第一（59 分）之前，Opus 5.5 以 58 分领先；GPT-6 Astra 跳升第二（57 分，原第六——计算机使用与科学权重上调的直接受益者）；GPT-5.1 Sol 第三（原第九）；Sonnet 5.5 从第二滑到第四；GPT-6.1 Terra 第五（原第十，主持人称这说明它比本周早些时候外界以为的要好）；Gemini 4 Argon 升一位、与 Meta Wonder（转写原作 Wonderful）并列第六——Wonder 从第三掉下来是显著降级，但主持人说许多人本来就认为那个第三与真实体验不符；Fable 5.1 从第五滑到第八；跌幅最大的是 Kimi K3.1 Pro，从第四直落第九。

两个结构性结论被节目重点放大。其一是开源权重差距骤然拉大：AA 官方口径「开源权重差距从 3 分扩大到 12 分」——Kimi K3.1 Pro 以 46 分居开源首位；在已发布权重的模型里，DeepSeek V4.1 Flash 与 GLM 5.3 以 32 分并列领先，落后榜首 26 分，甚至低于 36 分的 Claude Haiku 5.5。其二是头部与其余的断层：Anthropic、OpenAI、Google、Meta 四家都有 50 分以上的模型，Fable 5.1 与 Kimi K3.1 Pro 卡在线外，其后到「领先其余梯队」的 Claude 5.5（转写原文如此，疑与前述 Sonnet 5.5 同物或另有所指）之间有 10 分空档——主持人的解读是：把更多 agentic 任务放进指数后，前沿与其余模型的分离第一次如此清晰。成本层面，节目引 AA 称前十模型的每任务成本差异巨大：最便宜的 GPT-6.1 Terra 约 0.96 美元/任务，最贵的 Fable 5.1 约 13 美元/任务（转写作 13.04 cents，按与 Opus/Sonnet 约 10 美元同量级疑为美元口径）；Opus 与 Sonnet 约 10 美元，GPT-6 Astra 6.24 美元，另一处口径 GPT-6.1 为 1.27 美元（与前述 0.96 并存，两处数字未调和，均为节目口径）；Muse、Wonder、Grok 4.7 与 Carbon 都在 7 美元档、略贵于 Astra。

**舆论的接受与反驳。**看多更新的声音：OpenAI 战略前瞻负责人 Dean Ball 称「这是 AA 排名第一次大致接近我本人对模型的主观排序」；Haseeb Qureshi 认为重排「贴合体感得多、清掉了大量刷榜，也显示出 Anthropic 与 OpenAI 和其余玩家之间更深的沟，以及闭源与开源更大的沟」；Ethan Malick（转写疑作 Ethan Mollick）说这符合「我们还没见到 Fable 级开源模型」的普遍体感；内容创作者 Theo 的总结更刻薄：「今天是『开源模型和 Opus 5.5 一样好』guy 的糟糕一天。」为开源说话的最有力者是布鲁金斯学会研究员 Kyle Chan，节目完整转述了他的三点：其一，AA 之前就有私有基准，这次只是份额从 45% 提到 60%；其二，看具体科目比看总分重要——Kimi K3.1 Pro 的明显弱项在科学（Terminal Bench Science 得分 34%，GPT-6 Astra 为 63%），但它在 agentic 编码上仍极强，Terminal Bench 4.0 以 67% 领先所有模型；其三，这算刷榜证据吗？不一定——Kimi 失分最多的恰恰是非私有基准，而它在私有的 Automation Bench AA 上赢了 Claude。Kyle Chan 的结论被主持人引为更平衡的叙事：「中国模型能在特定 agentic 任务上追平甚至击败美国前沿模型，但在高级科学推理与科研工作流上仍有显著差距。」

主持人自己留下两条批评。一是权重之争：OpenAI 与 Anthropic 确实在把重心压向科学发现，但对普通用户的实际用例，科学权重 25% 是否合理？这既可以论证「应该换一套权重」，也可以论证「别看总分、只看你在乎的类别对应的单基准」。二是设置之争：榜单按 max 档跑，而「现实中没人用 max 跑」（X 用户 3.0 World 语，转写名）——AA 在成本分析里确实把 effort 档位纳入了计算，但主持人认为更实用的是让指数既能按模型、也能按 effort 档位组织。最后一条新增量：AA 同步推出了 Progress Index，把 2022 年 11 月以来所有模型放到同一条能力轴上看各实验室头牌的纵向进步。节目引用 X 用户 Lasan Al-Gayb（转写原文如此）用它做的两个观察——「OpenAI 要么真有最好的强化学习，要么他们的模型不开推理就很难用」，以及换一种画法后估算中美差距约 3.3 个月且在收窄。主持人收束全期：基准重要且天然有限，他乐见更多力气花在改进基准上，连这些争论本身也在增进我们对模型差异的理解。

机制、影响与启发

**链条一：评测供给决定竞争叙事。**变化（AA v5 私有数据集占多数＋agentic 全面加权）→机制（刷榜成本上升、长上下文等已「商品化」能力被移出指数）→受益者（重仓 agentic 与科学的美国闭源实验室、以及榜单话语权本身回归 AA）→受损者（依赖旧权重排位的开源阵营与第二梯队）→二阶影响（模型选型预算向「按用例看单基准」迁移，总分的营销价值波动）→观察指标（私有基准份额是否继续上升、各实验室是否转向发布「按 effort 档」的成绩、Kimi/DeepSeek/GLM 下一代在科学类基准的分数）→启发（引用 Kyle Chan：总分变动时先查权重与科目明细，再下「差距拉大」的结论）。

**链条二：基座＋RL 让发布速度本身成为竞争变量。**变化（Carbon 距 Argon 两周、且据称还有更强未发布版本）→机制（同一超大基座上做后训练迭代，跳过预训练周期；内部 dogfooding 提前消化新能力）→受益者（拥有超大基座与自研算力的 Google；用户更快拿到增量改进）→风险（发布节奏快于安全评测与一般可用性，Carbon 的「审慎 rollout」正是这一张力的表现）→观察指标（下一版 Gemini 间隔、是否常态化「小步快跑」、其他实验室是否跟进基座复用策略）→启发（判断实验室竞争力时，把「迭代周期」与「单点分数」并列为一级指标——这是本次节目内外部信号共同的指向）。

**链条三：云高增长与物理约束并存，泡沫论需要换抓手。**变化（Azure +48%、谷歌云翻倍，但 capex 实际支出低于指引）→机制（AI 收入高速货币化的同时，搜索被自家低货币化率的 AI 模式分流；数据中心扩张受电力/组件/人力硬约束）→受益者（云与算力持有者短期、电网与供应链环节）→受损者（依赖搜索广告旧模式的资产定价、社区对数据中心的容忍度）→二阶影响（「capex 减速刺破泡沫」的空头逻辑失效，检验点转向需求侧与货币化率）→观察指标（谷歌 AI 模式货币化率的变化、明年 3000 亿 capex 指引的兑现度、OpenAI 对 Azure 收入贡献的可持续性）→启发（节目口径：别把「支出低于指引」直接读成需求走弱——先分辨「不想建」与「建不动」）。

**链条四：监管对平台型公司的打法是扩大射程。**变化（DOJ 对 NVIDIA 的调查从 acquihire 惯例扩到风投、销售、客户施压与传统收购）→机制（acquihire 结构不触发自动审查，监管以「全链条调查」回应结构性权力而非单笔交易）→影响（Hugging Face 130 亿美元交易的 2027 年交割时间表承压；acquihire 惯例的隐性成本上升）→观察指标（是否出现正式指控或和解条件、后续 AI 交易是否更多走完整审查）→启发（节目口径：调查范围比调查存在本身更重要；对平台公司，分发节点＋算力＋资本的组合才是监管真正盯住的东西）。

ASR 校正表

（完）

转写原文（修正版）

[0s - 120s] So frequently on this show, you will hear me ramble off a bunch of numbers that are supposed to tell you something significant about a new model that has just come out. And more often than not, those benchmark numbers are in some ways related to Artificial Analysis, the company that has become the leading provider of the overarching benchmark that people use to judge AI models. And as AI models become more significant, the benchmarks that tell us how they perform relative to one another consequently get more important. Today, we are looking at a major update to Artificial Analysis’s Intelligence Index, as well as a new leader for the benchmark, which comes, believe it or not, from Google. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. All right, friends. Quick announcements before we dive in. First of all, thank you to today’s sponsors, KPMG, Sea Data, Granola, and Blitzie. To get an ad-free version of the show, go to patreon.com/aiDailyBrief, or you can subscribe on Apple Podcasts. And if you want to learn more about sponsoring the show, send us a note at sponsors@aidailybrief.ai. On aidailybrief.ai, you can also find links to everything else going on in the ecosystem, free trainings, paid trainings like our new Champions program at Superintelligent, or of course the new How We Got Here podcast and the first episode out this week called How Tech Lost America. Again, you can find all of those links from aidailybrief.ai. Welcome back to the AI Daily Brief Headlines Edition. All the daily AI news you need in around five minutes. We kick off today with a story that could have easily been the main, and honestly, probably would have been if more people had had their chance to get their hands on this model. Google has unveiled their next frontier model, Gemini 4 Carbon, just two weeks after the release of Gemini 4 Argon. As I have mentioned on the show a number of times since Argon arrived, having a new base model means Google can iterate quickly with reinforcement learning rather than needing to complete a lengthy pre-training run to put out a new model. And this is definitely a quick pace, one of the shortest gaps we’ve seen between versions of ultra-large models. CEO Sundar Pichai even suggested the pace would increase from here, commenting on Wednesday night’s earnings call. This rapid progress.

[120s - 240s] Underscores our incredible model momentum. We are at the point where AI is helping us increase our training and launch velocity. I would expect the next 12 months to see continued acceleration because of RSI. So Google is iterating at a record pace, but I’m sure you are sitting there asking yourself the same question that everyone is, which is, is the model any good? In particular, is it a significant improvement on the underwhelming release of Argon? Certainly, when it comes to the benchmarks, the answer is a definitive yes. In coding, Carbon gained eight points on Terminal Bench 4.0 and six points on Frontier Suite. Cybersecurity saw a 12-point increase on Exploit Gym and a three-point increase on CWE Bench. For computer use, Carbon picked up seven points on OS World and eight points on Agents Last Exam. On six of those eight key benchmarks, Gemini 4 Carbon is now the top of the leaderboard, ahead of Astra and Opus 5.5. Gemini 4 Carbon is now the front runner on the new Artificial Analysis Index v5, which we will be discussing in the main part of the episode. It scored 59 points, representing a nine-point jump on the new scale and placing it ahead of Opus 5.5 in first place. On the AA Coding Index, Carbon landed in second place, tied with Sonnet 5.5 and three points behind Opus 5.5. Now these are huge jumps for Google. When Argon was released two weeks ago, some felt that Google had missed their window. The model might have been impressive a month or two earlier, with benchmark scores topping Fable, but the release of Opus 5.5 and GPT-6.1 Sol had pushed the frontier sufficiently that it left Google in the second tier alongside Meta. This release, if people’s experience with it matches the benchmarks, puts Google genuinely back in the race for the frontier. Cost is basically unchanged compared to Argon, with Artificial Analysis measuring cost per task at $7.76. For reference, that’s 23% cheaper than Opus 5.5 and 20% more expensive than Astra, so right in the middle of the road for frontier models. Now here we come to the less great part. We don’t have many firsthand testimonials, as Google unveiled the model before it’s ready for general release. They’re following the administration’s pre-release model testing guidelines and said their

[240s - 360s] Taking what they call a deliberate approach to the rollout, the model will be made available to a subset of cyber defenders beginning this week, with access extended to trusted partners in their cyber defense program. As for the rest of us, Google wrote, “We’re continuing to gather feedback and iterate on guardrails before making Carbon available to developers, businesses, and consumers via our product surfaces as soon as possible.” One of the few groups talking about their experience with the new model was Arena, who had access for testing purposes. They put out a video review full of render and game design tests that were generally slammed for not being as pretty as Opus 5.5. Now, as I have said before, I am not particularly convinced that 3D design is the right yardstick for LLMs, no matter how much social media users like to focus on that. But certainly, at least when it comes to that particular use case, Carbon is not at the same level as its rivals. Arena’s Peter Gosted did give a few more details, however, posting, “I have used Carbon a bit before the release, and I can tell you it feels a lot better than Argon. This is a real deal from Google this time. I hope they keep up with the rapid pace of releases from now on.” And that does seem to be the plan. Responding to T. R. Taxes commenting that Fable 5.5 isn’t here yet, though, DeepMind researcher Zachary Nado wrote, “Gemini Redacted isn’t here yet, though.” Chubby on X commented, “Okay, now things get crazy. They have another unreleased Gemini model that’s even better than Carbon, and Carbon has already reached number one in the AA Intelligence Index today.” Gemini post-training researcher Xiao Zhang added that Google AI is out of their funk and operating like a well-oiled machine. Less than a year ago, he wrote, “Launching Gemini 3 Pro was a big bang event where the whole team celebrated number one benchmark numbers. Today, Gemini 4 Carbon hits number one on the AA Index, yet the feeling inside is distinctly different. We’ve been dogfooding continuously for weeks, and we’ve almost adapted to the new intelligence, taking ambitious code rewrites for granted, leaving loops overnight that launch dozens of experiments to repo, root cause, and fix mysterious bugs, and using swarms to inspect and fix RL data. Launches just feel like periodic checkpoints now.” And despite taking 10 months off the early adopter AI conversation, you can see that Google still has lots of people’s benefit of the doubt.

[360s - 480s] Benjy’s Hessen writes, “Why is Google suddenly dropping models like they invented transformers, own half the internet’s data, and sit on an ocean of TPUs? Oh wait.” Latent Spaces Twix writes, “Gemini number one on AAV5. This is making me open up anti-gravity for the first time in a long time. If Google actually pulls this off, it is officially comeback season with Meta, Croc, and now GDM being cool again.” For my part, I think we will suspend judgment until we get our hands on the actual model, but it looks promising, and that in and of itself is exciting. Now, speaking of Google, the other story was in there and Microsoft’s earnings reports. Both companies delivered another incredibly strong quarter in cloud growth, which is where, of course, they contain their AI revenue. Microsoft said their Azure division grew at 48% over the past year to reach 33 billion. This is the first time Microsoft has broken out Azure revenue as a separate line item, which seems like a very good decision with growth this strong. Azure drove overall revenue growth of 19% to reach 92.3 billion for the quarter. Each of those figures was a slight beat over analyst estimates. Microsoft also disclosed that OpenAI contributed 10 billion to Azure revenue, around a third of the total business. Microsoft stock was up as much as 4.8% to reach a new all-time high in after-hours markets. Google’s cloud division, meanwhile, doubled revenue compared to year ago figures, reaching 30.7 billion for the quarter, and meaning they’ve almost entirely closed the gap with Microsoft. This was well above Wall Street’s high expectations. However, as David Wagner of Aptus Capital Advisors put it, “The problem is everything else. Search revenue growth came in at 12%, a massive slowdown from 17% in the previous quarter and 19% the quarter before that. YouTube revenue also added a third continuous growth slowdown to reach 9.2%. Even TPU sales was a bit of a question mark, with executives dodging a question about profit margins for the chips during the earnings call. Overall revenue growth still came in at 25%, roughly in line with expectations. But to some analysts, it looks like Google might have reached the moment where they’re forced to cannibalize their own search business because it’s no longer a reliable growth driver. In fact, CEO Sundar Pichai acknowledged this dynamic.”

[480s - 600s] During the earnings call, commenting, “It’s worth noting that the strong growth in AI mode queries had an impact because of the lower monetization we’re seeing in AI mode. In addition, Google missed on bottom line earnings largely due to a markdown on their SpaceX stock.” Tech analyst Dan Ives, ever the optimist, tried to strike a positive tone, writing, “This is the cloud print we have been waiting for, and in our view, it matters more than a lighter search number the bears will lean on tonight.” And lean on it they did, sending the stock plummeting as much as 10% in after-hours markets. Much of the coverage focused on the cloud race between the two companies, which are now neck and neck, but the pace of the data center buildout might be even more consequential. Both companies held capex forecasts steady for this year and next, but actual capex spend for the quarter came in under expectations. Google, by way of example, is guiding a 50% increase in capex next year to 300 billion, but they’re currently undershooting this year’s numbers. Oracle and Meta are in a similar position, so it looks like the AI buildout may have reached peak velocity. That’s not at all surprising given physical constraints around components, power, and labor, let alone community support running on empty. Now we are quite early in this trend, and Google said that they would have more in-depth discussions on capex during the next quarterly earnings. But what makes this interesting is that many AI bears suggested that a capex slowdown would be the catalyst that pricks the AI bubble. What we’re seeing, however, doesn’t exactly line up with that theory. The logic was that demand would falter and force the hyperscalers to slow down. Instead, it looks like the hyperscalers, if they had their druthers, would like to continue accelerating their data center buildout, but are running into more fundamental and physical constraints. Tech analyst Luke Longo thought the slowdown was more about finance, but are running into more fundamental and physical constraints. Lastly, today, more antitrust scrutiny for NVIDIA as the DOJ queries their acquisition of Hugging Face. Reuters reports that the Department of Justice has made inquiries about how the acquisition would impact competition in the AI industry. In particular, the DOJ has asked whether ownership of Hugging Face would allow NVIDIA to control a key distribution point for AI models.

[600s - 720s] When they made the thirteen billion dollar acquisition in September, Nvidia maintained that Hugging Face would remain free to offer a platform and partner with a wide range of model developers, chip manufacturers, and cloud providers. They said the deal was expected to close in the first half of 2027, pending regulatory approval. These inquiries form part of what is now revealed to be a multi-year antitrust investigation into Nvidia. In September, the New York Times reported that the DOJ had opened an investigation into Nvidia’s licensing and talent acquisition deal with chipmaking startup Grok. That deal was structured as an acquihire deal, which has been a common practice in the AI industry dating back to Google’s 2024 deal with Character AI. Because these deals are not technically acquisitions, they don’t trigger automatic antitrust reviews and can therefore theoretically close much faster. The assumption had been that the DOJ was using an investigation into Nvidia as a way to crack down on this practice more broadly. However, sources told Reuters this week that the investigation is much broader and seems to predate the Grok deal. Acquisitions of data center optimization companies SkedMD and Run:ai are also within scope for their impact on the competitive landscape in AI infrastructure. And both of those deals were traditional acquisitions, not acquihire deals. Furthermore, sources said that recent questioning from the DOJ covered Nvidia’s venture investments, sales practices, and whether the chip designer has exerted pressure on its customers. In other words, rather than just a narrow investigation about circumventing antitrust rules, it now seems the DOJ is embarking on a full top to bottom antitrust investigation of Nvidia. Obviously, no charges have been filed at this stage, and the investigation could conclude with no action. So there’s no implication that Nvidia has done anything improper. Still, compared to what most people thought this was, it appears to be a much more concerning investigation for the dominant chip maker. Something to keep an eye on as we go forward for sure. But for now, that is going to do it for today’s AI Daily Brief. Next up, the main episode. A new study from KPMG and the University of Texas at Austin found that when people work with AI, similar skills don’t guarantee similar outcomes. Researchers studied more than 500 early career professionals and found that the best performers consistently amplified the value of AI by guiding

[720s - 840s] Evaluating and refining its outputs. These top performers, called AI amplifiers, weren’t defined by what they knew alone, but by how they worked with AI. Learn more about what separates AI amplifiers from everyone else at KPMG.com/us/AI-amplifiers. This episode is brought to you by CData. Token spend is climbing, security is holding agents back, and companies rolling out AI initiatives might not know how it’s truly being used. The root cause: AI is cut off from the systems and context your business runs on. That’s why CData built Connect AI Gateway, one control point between your AI and your systems. Connect AI Gateway is not just another static model router. Prompts get intelligently routed to the right model with exactly the tools and data it needs, governed down to the record. The result: accurate answers at a fraction of the token cost with a context engine that learns your business with every interaction. Built on the connectivity layer powering Palantir, Google, Salesforce, and hundreds more. Give it a try and start your free trial at CData AI. That’s C as in Connect Data AI. You’re in back-to-back meetings all day. You’re trying to stay present, but you’re also worried you’ll forget the decision, the action item, the important next step. This is a scenario that we absolutely live here at the AI Daily Brief. Given how fast things move and change, and how quickly we’re evolving everything that we do, that’s where Granola comes in. Granola is an AI-powered notepad for meetings. You jot down rough notes like you always do, and in the background, Granola transcribes and turns them into clear, useful notes when the meeting ends. No bots joining your calls, no distractions, just a clean notepad that helps you focus. During or after the call, you can chat with your notes, ask Granola to pull out action items, help you negotiate, write a follow-up email, or even coach you using recipes which are pre-made prompts. Once you try it on a first meeting, it is very hard to go without. Head to Granola AI AI Daily and use code AI Daily to get 100% off for three months for new users. That’s Granola AI AI Daily and code AI Daily to get 100% off for new users.

[840s - 960s] For three months, Blitzy’s deep codebase understanding unlocks the thing every roadmap owner cares about: shipping new features. Here’s the truth about building inside a massive enterprise codebase. Writing code was never the bottleneck. Context is. Which system does this touch? Which contracts can’t break? Which standards apply? Blitzy already knows because it reverse-engineered your entire codebase into a dynamic knowledge graph before feature work began. With that complete picture, Blitzy builds features end-to-end. Architecture, APIs, UI, and tests all validated against your existing systems. One Blitzy customer built an AI-native application from scratch with 100% autonomous completion, saving over 2,700 engineering hours. Features that respect your codebase instead of fighting it. Stop letting your backlog grow faster than your team. Accelerate your roadmap at blitzy.com. That’s b-l-i-t-z-y.com. Welcome back to the AI Daily Brief. Today we are talking about benchmarks, specifically a new set of benchmarks and how they’re shaking up our sense of the AI field. Now the eagle-eared among you may wonder exactly what I think about benchmarks. Given that on the one hand, with every new model release, I implore you to only give them so much attention, and yet at the same time still rattle off a whole bunch of them as though they matter. In short, I think that benchmarks are inherently limited, but also important. Without them, we don’t have good tools other than strictly vibes to understand the difference between different models. And yet, in almost no circumstances do I think we should take the performance of one model on any one benchmark or even in the aggregate as the gospel. Done well, they are a useful guide to helping us understand where different models fit into our overall stack. But as we will see today, there is more and more effort to addressing some of the inherent limitations of benchmarks and trying to make them better. Whenever there is a new model, you’ll hear me talk about the scores that it gets on the independent test administered by Artificial Analysis. Well, that company has just overhauled the way that they test model, giving us a brand new yardstick for Frontier AI with version five of their Intelligence Index. The index takes around a dozen key benchmarks and distills.

[960s - 1080s] Them into a single score, and while the index is now entrenched in how we understand the AI race, leaning so hard on that single score is a relatively new phenomenon. Prior to the era of agentic coding, when LLMs were in many cases still glorified Google or writing assistants, vibes reigned supreme. Ranking on the chatbot arena leaderboard was the pinnacle test of a new model because it told developers what users wanted. But as research and coding became more important use cases, we got more objective benchmarks giving a clearer indication of how each model would perform on real work. These evaluations have evolved over time. First by adding harder questions as benchmarks became saturated, in other words, as all the models crowded the very top score of a benchmark, the creators of said benchmark would go add in a bunch of new or harder questions, creating an iterative version that would hopefully provide more granularity on the difference between different models until that model got saturated as well. As agentic computer use became a fundamental part of model performance, the evaluations got an even more complete rethink. Through each of these changes, artificial analysis also evolved, switching out benchmarks and rebalancing their weightings as different model abilities became comparatively more important. This hasn’t always been a smooth process. There have been moments where the intelligence index looked extremely saturated and stopped giving us a good indication of which models were best. There were moments when benchmark maxing allowed developers to game the system and rank much higher than real world testing would support. In fact, this has been a fairly consistent story when it comes to Chinese open weight models. In early September, artificial analysis pushed an early update with version 4.2, pulling forward many of the changes they had planned for V5. That decision was triggered by the release of GPT-6 Astro, which focused heavily on improving computer use. The problem was that V4.1 of the intelligence index still had relatively light weightings for agentic benchmarks, leading GPT-6 Astro to have a slightly disappointing score. And more than just disappointing, one that didn’t really comport with people’s lived experience of the model. After the update, Astro was still ranked slightly behind Fable 5.1, but the index had become a closer reflection on how people were actually using the.

[1080s - 1200s] New crop of models, and this has been the story of the intelligence index—a constant cat-and-mouse game of updating the mix of benchmarks to match the rapidly evolving real-world use cases for AI. Introducing version five of the index, artificial analysis highlighted that for the first time, the majority of the benchmarks are built on private datasets. The idea is to make benchmark maxing more difficult, as model developers theoretically can’t train their models to the test, and that, by the way, is what we mean by benchmark maxing when developers include those sort of evaluation tests in their answers in the training dataset, leading to an artificially inflated score, as the benchmark ceases to measure what the model can do and instead reflects simply a memory of the correct answers. For version five, AA added two new benchmarks: Frontier Code 2.0, which is the brand new version of Cognition’s private benchmark of real-world coding tasks. This test features scoring based on whether a repository maintainer would merge the code. They also added Terminal Bench Science, which is an agentic scientific research benchmark. Humanity’s Last Exam, once pitched as a near-impossible knowledge search task, has been upgraded to HLE Diamond, the even more difficult version. Similar saturation updates have been made to AA Omniscience, which measures knowledge and hallucinations, as well as Crypt PT, a physics test. GDP Val Double A, the measure of agentic work in real-world settings, is out, with the index now relying on AA Briefcase, Automation Bench AA, and GDP PDF to measure white-collar work along more granular lines. Also removed are Sci Code, a scientific research test, and AA LCR, which tests long-context reasoning. Now, rather than just a list of names and acronyms, these adjustments say a lot about how AI models have changed since v4 of the index was released in January. At that time, long-context reasoning was one of the most critical features, used for working on large codebases. Many coding models still had limited token context windows and suffered from fatal issues with context rot. Since then, improvements to compaction and memory have allowed us to take a lot of that long-context performance for granted. Agentic work is the other huge change that’s evident from the redesign of the index. The

[1200s - 1320s] Coding benchmarks included in the index now only measure agentic performance, as that has been cemented as the way developers use modern AI. And the science benchmarks are now testing real-world agentic research work rather than pre-existing scientific knowledge. For the time, artificial analysis has chosen to equally weight the index across four categories: knowledge work, coding, science in general, which encompasses both general knowledge as well as hallucination rate. In some ways, it feels like that 25% weighting for science is a forward-looking bet based on how Anthropic and OpenAI have been trading science and math headlines over recent weeks. So, what did this change in the scores? In version five, Opus 5.5 was still in the lead with a score of 58, at least until Google dropped Gemini 4 Carbon on Wednesday. GPT-6 Astra jumped into then second place with a score of 57, and Astra had previously been in sixth place, showing there’s a pretty big jump based on the index giving more weight to computer use in scientific tasks. GPT-5.1 Sol takes third place up from ninth, followed by Sonnet 5.5, which slipped from second to fourth. GPT-61 Terra was in fifth up from tenth, suggesting the model is a little better than many assumed earlier in the week. In just a moment, we’ll get into what I think that reflects. Gemini 4 Argon gained a rank to tie Meta Wonder in sixth place, which is a pretty notable drop for Wonder, which was previously in third. Although many felt that that ranking didn’t line up with real-world experience. Fable 5.1 is looking a little long in the tooth in eighth place, sliding from fifth, and in ninth place we have one of the most notable drops in the new index, with Kimi K 3.1 Pro falling all the way from fourth. In fact, artificial analysis noted that the gap between open-weight models and U.S. proprietary models had opened up massively after the index rebalancing. They wrote, “The open weights gap widens from three points to 12. Kimi K 3.1 Pro is the top open weights model at 46. Among models with released weights, DeepSeek V41 Flash and GLM 5.3 lead at 32, 26 points behind the leader, and below Claude Haiku 5.5 with 36 points. Now, in addition to the new gulf between U.S. and Chinese models.

[1320s - 1440s] There was also a clear drop off between the handful of leading labs and everyone else. Anthropic, OpenAI, Google, and Meta all have models with scores of 50 and above. Fable 5.1 and Kimi K 3.1 Pro are just outside that cutoff. However, there’s then a 10 point gap between Kimi K 3.1 and Claude 5.5, which leads the rest of the pack. Basically, by adding more agentic tasks to the index, there’s now a clear separation between frontier models and the rest. Artificial Analysis noted that costs vary wildly among the top 10 models that arguably make up the frontier. That’s largely because the top 10 is a mix of large models, token-intensive models, and a few smaller, more efficient models. At each end of the spectrum, we have GPT 6.1 Terra as the cheapest at 96 cents per task, and Fable 5.1 is the most expensive at 13.04 cents per task. For a more apples-to-apples comparison based on model size, Opus and Sonnet both cost around 10 dollars per task, compared to 6.24 for GPT 6 Astra and only a buck 27 for GPT 6.1. Muse, Wonder, Grok 4.7, and Gemini 4 Carbon are all slightly more expensive than GPT 6 Astra in the high seven dollar range. Now, most people found this to be a welcome update. OpenAI head of strategic futures Dean Ball wrote, “This is the first time an artificial analysis ranking has even roughly approximated my own subjective ranking of models.” Haseeb Qureshi wrote, “Artificial Analysis re-ranking their intelligence score seems to match the vibes a lot better and clear out a lot of bench maxing. Shows a deeper gap between Anthropic, OpenAI, and everyone else. Also shows a bigger gap between closed and open weights.” Ethan Malick wrote that the analysis quote matches the general vibes that we haven’t seen a Fable class open model. And AI content creator Theo drove that point home, writing, “Terrible day to be an open weights models are just as good as Opus 5.5 guy.” Now, there were a few people who spoke up for the open weights models. Brookings Institute fellow Kyle Chan pointed out that even before this, artificial analysis had some private benchmarks. It’s just that the share went from 45 percent to 60 percent. Using Kimi 3.1 Pro score as an example, he suggested that while the headline news.

[1440s - 1560s] Is the gap with the top U.S. model widening from three points to twelve? That it’s more important to look at the specifics. Science, he points out, is a major weakness, with Kimi scoring just 34% on Terminal Bench Science versus 63% for GPT-6Astra. But he writes, Kimi is still extremely strong at coding agents, pointing out that it actually leads all models on Terminal Bench 4.0, scoring 67%. Kyle’s conclusion is this evidence of benchmark gaming? Not necessarily. Some of Kimi’s biggest deficits are on non-private benchmarks, while it actually beats Claude on the private automation bench Double A evaluation. The bigger story, he says, Chinese models can match or beat U.S. frontier models on certain agentic tasks, but a significant gap remains in advanced scientific reasoning and research workflows. And ultimately, even if in general people think that this is a positive update, no one is going to agree on everything. In particular, I think one of the biggest debates will become what the relevant weighting should actually be. Yes, OpenAI and Anthropic are clearly emphasizing scientific discovery, but relative to the average individual use case, does that matter as much as agentic coding as it does in the weighting of the new V5 index? Now that might be an argument for a different weighting, or it might be an argument that instead of paying attention to the headliner number, we’d be better suited to focus on the individual benchmarks for the categories of use cases that we care about. One critique that I do think is worthy of consideration is the fact that the leaderboards are run based on max settings. 3.0 World writes, no one runs max in real life. Where’s the medium slash high slash extra high leaderboard? And to be clear, it’s not like artificial analysis ignores that, especially in their analysis that involves cost. The effort setting does become part of the calculus. But I do think that on a practical level, people might be better served by the ability to organize the index not just by model but also by effort. Now one last note. Alongside the new intelligence index, artificial analysis introduced a new progress index. The index places all models released since November 2022 on a capability scale to see how each lab’s leading model improves over time. It’s designed less to be about the current batch of models and more to give a

[1560s - 1604s] Long-term view of model progression. Still, the new measure gave people easy access to a bunch of data to play with. Lasan Al-Gayb, for example, looked at how the switch to reasoning models impacted the major labs, commenting, “OpenAI really does have the best reinforcement learning, or their models just suck without reasoning.” And by charting the data in a different way, they estimated that the gap between the U.S. and China is now 3.3 months and closing. Summing up, like I said at the beginning, I think benchmarks are important and inherently limited, but I love that more effort is going into improving them, and I think even the debates that follow are also extremely value-accreitive when it comes to how we understand how different models compare to one another. For now, however, that is going to do it for today’s AI Daily Brief. Appreciate you listening or watching as always, and until next time, peace.