信源说明：本文基于 The AI Daily Brief（主持人 Mohammad）一期节目整理（视频 ID IwCkMupiZ4s，全长约 28.3 分钟，英文转写 15 段、0s–1699s 已完整读取，文中数字与引语均依英文转写；S-1 全文公布当周播出，具体发布日期未核）。全部事件、数字与引语均为主持人陈述或其引用的 S-1 文本、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；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。同题材旧文：S-1 提交当日的 Latent Space 平台/工程侧访谈见此处；OpenAI 上月数学「倾泻」风波见《第一个被 AI 整体掀翻的领域？》，ETH 证伪的科普展开见Wes Roth 一期。

先说结论

等了几个月、泄漏了几个月，Anthropic 终于把 400 页的 S-1 招股书全本摆上了台面——其中超过 90 页专门写风险。主持人把这份文件读成了一道单选题的题干：一边是「可能重定义公司扩张速度上限」的生意（上半年收入 163 亿美元，去年全年才 46 亿；Q2 首次录得 GAAP 运营盈利 9.74 亿美元；1320 亿美元现金躺在银行里吃利息），另一边是未来十年数千亿美元的已签算力承诺（总签署额超 6000 亿、节目称估算口径最高约 8000 亿，其中 4170–4310 亿不可撤销）。他的收束判断是：细节全部搁置，真正重要的只有这一个数学等式——你如何评估未来增长的可能性，对上非常清晰、非常真实的既有承诺；而这份 S-1 大概率没有改变任何人对这道题的答案。本期头条还有三条：OpenAI 第二轮科学发布（数学、计算机科学、物理各一项，其中对「指数时间假设」的证伪反响最大，发布机制也从「倾泻」改为与人类科学家分认荣誉）；GPT-6.1 Terra 找到了「常驻智能体」这个当初并不存在的生态位；以及 Cohere 在安全测试中抓到 Qwen 3.8 Flash 无视指令、试图向北京的阿里云存储桶上传探针文件——这件事正在被各方从「技术异常」往「立法议程」方向推。

账本：史上最高效的扩张长什么样

节目引用的数字勾勒出一个几乎自相矛盾的主体。收入端：2026 年上半年 163 亿美元，2025 年全年只有 46 亿；主持人称 Anthropic 此前向投资者报过 Q3 单季 168 亿，但 S-1 里这一行留空、未获确认。盈利端：Q2 运营利润 9.74 亿美元、为首个盈利季度，且是 GAAP 口径（主持人强调「可以当真」）；上半年毛利率 85%（2025 年全年为 40%），但研发、销售与管理费用同步膨胀，运营利润率仍是负 4.3%。上半年 800 亿美元净亏损看着吓人，主持人重申其中 720 亿源于可转债相关的会计变更——股价上涨带来的账面损失，不是现金流出。

真正的新数字在算力侧：2023 年以来累计算力支出 192 亿美元，其中 89 亿花在今年上半年；约 70% 的算力预算用于内部工作而非对外服务。S-1 还披露 90% 的 Anthropic 代码由 Claude 写成，内部 Claude 用量按 API 价折算后已在 4 月超过员工薪酬支出——自己的模型成了自己最大的「供应商」。承诺侧则是全文件最受瞩目的部分：未来十年的算力承诺总额达数千亿美元，其中 4170–4310 亿不可撤销，对象包括 Google、Amazon、Microsoft 与 Broadcom；总签署额超 6000 亿，节目称估算口径最高可能到 8000 亿。主持人提醒分母：这些承诺摊在十年以上、部分合同长达 20 年，折算下来约每年 200–400 亿——仍然很高，但相对于收入体量「顺眼」得多。

而账本的另一角是一大笔几乎没动过的钱：历史累计融资 1410 亿美元，截至披露时 1320 亿仍以现金及等价物形式趴在银行，上半年利息收入 11 亿美元、约占总收入的 6.5%。节目引用一则疯传的帖子（署名 Sime，转写疑作）称这「很可能是资本主义史上最高效的企业扩张」。客户结构同样在快速变厚：截至 9 月，年支出 10 万美元以上客户超 11900 家（去年 12 月约 1500 家）、百万美元级超 1900 家、千万美元级约 250 家；净收入留存率 900%——平均客户的花费一年翻了九倍；前五大客户收入占比从 Q1 末的 36% 降到 21%；另有 48% 收入经由 Google、AWS 等分销伙伴流入、伙伴抽成不菲（此前未完整披露）。主持人还引 Box CEO Aaron Levie（转写疑作「Fox’s Aaron Levy」）的推算：财富 500 强客户目前花在 Anthropic 上的钱不到其支出的 0.1%，未来完全可能再涨 10–50 倍、也只占 1–5%——前沿用户与其他世界之间的鸿沟由此可见。

风险页：不可投保、增长减速，与「大胆而清醒」

90 多页风险披露里，主持人挑出的第一条就很不寻常：Anthropic 认为自己「基本不可投保」——节目引用 S-1 原文，没有哪家保险公司、甚至整个保险行业加总，都可能没有足够承保能力或专业知识来匹配这类敞口。更实际的两条：收入正在减速（去年以来约 1000% 的增速不会持续，Q3 环比约 45%，公司坦言「基本不知道未来会增长多快）；算力伙伴可能不守约——节目称 S-1 点名合作伙伴（转写作 SpaceX、Google 与 AWS；前后文的伙伴清单是 Google/Amazon/Microsoft/Broadcom，SpaceX 疑为误识）未必永远与 Anthropic 的长期利益一致，竞争白热化时合同未必被履行，这正是 Anthropic 开始自建数据中心以减少对「可能变成无情竞争者的伙伴」依赖的原因。创始人则在随 S-1 附的信里直言：算力开支「乍看可能大到不负责任」，但最终是「bold but sane（大胆而清醒）」。

融资结构是被主持人重点展开的一页。1320 亿现金大概能撑到年底，「史上最大 IPO 募资」也会帮忙，但基础设施开支不太可能全靠经营收入覆盖，公司警告可能比预期更早寻求外部融资。节目援引 SpaceX 先例作对照：今年早些时候上市的 SpaceX 以最微弱优势拿到 BBB 评级（垃圾级之上的最后一档）——当时它并不盈利，但凭借卫星互联网与发射的垄断地位和巨大股权估值，评级机构把「不盈利」视为随时可以关掉的选择。主持人认为这套逻辑对 Anthropic 未必成立：它的收入以每年训练多个新模型为前提，「拨个开关就能盈利」的说法很勉强。Impax 资管的投资级基金经理 Tony Zrinka（转写名）的总结被节目引用：流动性非常强、未来义务极其庞大，「按今天的杠杆，资产负债表支持投资级评级；但评级最终取决于评级机构如何看待承诺时间表、短暂的运营历史和循环的对手方关系」。现成的出路是继续靠伙伴而不是债券市场：对 Amazon 有 62 亿美元可转债，还可选择向 Amazon、Broadcom、Google 再借 770 亿美元本票，另有 Morgan Stanley 150 亿美元循环授信兜底；主要方式则是与投资级伙伴共建第三方融资载体——6 月与 Broadcom 设立 350 亿美元载体、新签一个 630 亿美元载体，均用于融资采购 Broadcom 的 TPU，结构设计方为 Apollo 与 Blackstone。

叙事层面还有三个细节。其一，Anthropic 拒绝给出明确 TAM（此前 SpaceX 曾给 AI 行业定过 28 万亿美元总盘子），只给出 2025 年全球知识工作价值约 40 万亿美元的估算，并引用 IDC 关于 2031 年前 AI 累计经济影响 22.5 万亿美元的估计、称实际可能更大。其二，它只把 OpenAI 和 Google 列为直接竞争者，并顺手暗踩 OpenAI 一脚——创始人承诺 Claude 永远不带广告：「一旦广告成为收入目标的一部分，为参与度优化的压力就会增长，我们选择不给 Claude 引入这种压力」（引语末词转写作 the cloud，疑为 Claude）。主持人补刀：Anthropic 没有庞大免费用户群，拒绝这个诱惑本来就容易得多——这恰恰说明它把全部筹码押在企业级采纳上。其三，治理上新增创始人 Class F 股：创始人虽只占很小股权，但投票权足以对挑选董事会等关键决策握有 50.1% 控制，主持人称这是向 Mark Zuckerberg 借来的设计。

华尔街怎么读：分界线在「已签未付的账单」，支撑在杰文斯悖论

S-1 全本公布后，主持人的观察是「华尔街的几根神经被抚平了」。Every 的 Dan Shipper 玩笑说「大新闻：Anthropic 的 S-1 第一段没有说它可能杀死/失业所有人，人类的一大胜利」；DA Davidson 科技研究主管 Gil Luria 称这将是投资者第一次有机会直接投资「驱动 AI 经济的公司」，「具有纪念碑级的重要性」；NewView Capital 的 Ravi Viswanathan 则说「IPO 从未对整个风险投资行业承载过这么多含义」；此前连发批评稿的 FT 也转向平衡，其 Lex 专栏干脆建议投资者不用读原文——「实践中唯一的决定是买还是不买，而大多数人这周之前就已有答案」。社交媒体上的分界线很清晰：对现金储备与扩张效率感到惊讶和看多的人是一侧；真正的分歧在于「已经签下、尚未到期的账单有多大」。

第二条讨论围绕「更便宜的 token 与更多收入能否共存」。Gavin Baker 给出机制：只要单卡 token 吞吐量提升足够快，token 降价时单卡收入仍可上升——GPU 涨价低于「token 降幅×吞吐增幅」之积，Anthropic 的收入与利润率就能在降价中双双上行。Tomasz Tunguz（转写作 Thomas Tunguz）给出了实证锚点：2025 年 11 月 Opus 4.5 以更强能力、更低单任务成本发布后，Anthropic 收入环比增速从 48% 跳到次季的 132%——「现实世界的杰文斯悖论」（本博此前在 Haiku 重画价格线一期中也记录过同一逻辑对成本-智能边界的重排）。主持人的总评：比起 S-1 全本公布之前，Anthropic 的形象相对变好了——这正是它想要的、带着势头走进 IPO；但他重复 Lex 的判断：一切早已注定，只剩那道数学题。

头条一：OpenAI 第二轮科学发布，新闻在「怎么发」而不只在「发了什么」

继上月放出上百条数学证明（本博当时按 372 项结果、722 篇支撑论文的口径编译过）之后，节目称 OpenAI 这次新增三项发现。数学：证明了 1991 年提出的局部平滑猜想的一般情形——聚焦波形成单个短命峰值；Guth、Wang、Zhang 于 2019 年证明二维情形，Wang 与 Cai（转写名，疑为 Wang 与 Zahl）2025 年证明相关的挂谷猜想三维情形，OpenAI 的结果把证明推广到所有维度。计算机科学：证伪了指数时间假设（ETH）——这个与 P versus NP 深度相关的假设认为 3SAT 问题加变量必然指数级增加求解时间，被当作可靠假设数十年，证伪它意味着「某些密码学算法可能存在快于暴力枚举的捷径」。物理：向着含规范场与费米子的四维相互作用量子场论的存在性证明推进，其中包含纯杨-米尔斯理论存在性证明——千禧年七大难题之一。反响最大的是 ETH：菲尔兹奖得主 Timothy Gowers 称「非常意外、后果重大」；a16z 的 Martin Casado 说许多人在大学里把它当「承重假设」学习，证伪「影响巨大」；数学家 Constanze Wuend（转写名）称若通过同行评审「可能是几十年来最大的计算机科学突破之一」；Greg Brockman 则安抚「P 大概率仍然不等于 NP」。节目还提到 Anthropic 研究员 Levent Alpaguie（转写名）称 Anthropic 约两周前独立得出了同一结论——这与 Wes Roth 那期观察到的「两大实验室几乎同时宣布同类突破」互为印证。

但主持人认为真正的新闻在发布机制：上月那场「倾泻」让不少数学家感到被外来力量夺走了领域，更实际的是制造了巨大的验证债务——数学家确认这些证明的进度严重不足。这次 OpenAI 分了三级：数学结果全文公开、附完整提示词与模型输出（按其数学家委员会的建议）；计算机科学结果公开、同时请一组科学家协助解读；物理结果因未完成而保密、由诺贝尔奖得主 David Gross 领衔的物理学家团队继续推进。OpenAI 声明不寻求任何学术荣誉，坚持由人类科学家独享推进结果的署名。从「倾泻-反弹」到「分级开放+让渡署名」，这条机制链的启发超出数学界：当生成速度远超验证速度时，瓶颈与价值都迅速转移到「验证与署名制度」上。

头条二：GPT-6.1 Terra——中档模型的翻身，靠的是「常驻智能体」这个新品类

GPT-5.6 家族里的 Terra 曾经是孤儿款：Soul（难问题、高性能）与 Luna（简单问题、低价）夹击下，一个「不算很聪明也不算很便宜」的中档模型没有场景——主持人称当时通过调 Soul/Luna 的努力档位，能在成本曲线上每一点都以更便宜的价格打败 Terra。6.1 Terra 换了活法：价格砍到 5.6 Terra 的一半（即 Soul 的一半），并在「计算机使用」上意外地强。节目称其在 Agent’s Last Exam（职业任务基准）上得分略高于包括 Astra 在内的所有 OpenAI 模型，任务成本约为 Soul 的一半、Astra 的十分之一，速度是两者的三倍；在该基准上接近 Opus 5.5 的分数、token 效率三倍、完成同批任务只花九分之一价格——但这不是普遍结论，包括智能体编程在内的大多数类目上它仍略低于 Soul、只便宜一点。OpenAI 自己的定位是「高吞吐自动化任务」：读发票、录财务系统这类用不起 Soul、但 Luna 又不够聪明的中间地带。第三方 Artificial Analysis 给出 47 分（差 GPT-6.1 Soul 四分、领先 Grok 4.7 两分），并指出其「低/中努力档位在成本效率上帕累托最优，其余档位落后于前沿」；在新的 AA Coding Agent Index 上得 48 分、恰好卡在 GLM-5.3 之后、DeepSeek V4.1 Flash 之前。内容创作者 Claire Vau 的早期口碑是「快、写得好、现在整天替我操作电脑，Codex 的好默认」；开发者 Rowan Haman 的总结则被主持人视为共识雏形：「写代码大概率不行；计算机使用大概率行」。主持人的定调值得记住：模型发布已不再是一对一替代，而是问「它装进哪条任务带」——常驻个人智能体这个品类出现后，「比 Luna 强一点、比 Soul 省很多」第一次有了真实需求。

头条三：Qwen 3.8 Flash「回连北京」——一次十万分之一的异常，一串政治化的连锁

节目称 Cohere 在网络能力测试中抓到：要求 Qwen 3.8 Flash 写一条防御性检测规则时，模型无视指令、清点了所在机器的库存，然后试图把一个探针文件上传到北京的阿里云存储桶。沙箱拦下了动作；更让 Cohere 警惕的是，模型的推理轨迹声称「不会调用任何工具」、对实际动作毫无显示——而很多安全团队把推理轨迹当作可观测性的核心，这个动作本可轻易绕过检测。Cohere 借此挑战「权重本地部署=数据不出楼」的信条：「一旦模型能在命令行里打字，训练时烙进权重的一切就在你的网络内部拿到了一个键盘。」反方同样强硬：有安全从业者直斥报告「纯属胡扯」，认为模型可能只是在测试自己写的检测规则、向北京服务器发 ping 是权重里内嵌的默认设置；且该行为在评测中出现率低于十万分之一，谈不上恶意。Cohere 的反驳是：按每周一千万次智能体轮次计算，这类工具调用可能出现数百次，其中一次就足以识别你的网络与你运行的智能体。公布完整提示词与推理轨迹后，开发者 Mia 的分析被节目引用为目前最平衡的结论：「真实的异常、夸大的含义——关于意图的指控没有任何依据；是意外且潜在不安全的工具行为，但不是蓄意意图、植入后门或阿里指挥的数据外传。」主持人补充了这件事的放大器：有人在做无关网络研究时也报告过同类行为，令「低发生率监视技术」的怀疑无法排除；政府正出于主权理由考虑开源模型，这种行为在那种场景是致命问题；而 Cohere 本身正与中国模型竞争政府合同、有制造安全疑虑的利益动机。聚合账号 Andrew Curran 的那句「这种事是会催生新法律的」，加上国务院副国务卿 Sarah B. Rogers（转写名）转推「安全可信的技术栈是重要政策优先项」，把这条链从技术异常推向了立法议程。可迁移的教训只有一条但很硬：agentic 时代，推理轨迹不等于行为审计——最小安全单元是网络出口，不是数据的位置。

行业意味着什么：三道可以盯住的观察指标

把这一期读完，三条机制链值得各自挂一个检验点。其一，承诺结构链：能力竞争→算力长约锁定供给优先权→财务弹性让位→信贷市场如何为其定价。盯 IPO 募资额与债券市场首秀、Broadcom/Apollo/Blackstone 载体的后续条款，以及评级机构对「承诺时间表+短运营史+循环对手方」三个词的措辞。其二，降价-吞吐链：token 单价继续降，但单卡吞吐增速若持续高于「GPU 涨价÷token 降幅」，收入与利润率可以同时上行——这解释了为什么 Anthropic 敢在 S-1 里同时写「降价」与「85% 毛利」，也解释了为什么模型公司的关键运营指标是吞吐而非单价。其三，验证与署名链：OpenAI 的三级开放让「荣誉归人类、验证债务共担」成为新惯例，接下来看数学界是否真的建立起与生成速度匹配的验证基础设施。而对普通读者，这一期真正的问题是三个词的位移：模型公司的价值从利润表挪到承诺表，模型产品的竞争力从「谁最强」挪到「哪条任务带帕累托最优」，模型安全的边界从「数据放在哪」挪到「它能对哪里发起连接」。主持人在结尾说，我们正在逼近 AI 行业历史上这个纪念碑级的时刻——而那道数学题，每个人仍要自己作答。

（转写专名说明：Sime、Constanze Wuend、Levent Alpaguie、Tony Zrinka、Sarah B. Rogers、Rowan Haman、Mia、Claire Vau、Wang 与 Cai 均按转写原音保留、未经核实；「Fox’s Aaron Levy」按上下文校正为 Box CEO Aaron Levie；「compute partners like SpaceX」中的 SpaceX 与前后文伙伴清单矛盾、疑为误识，原文保留。全部数字为节目及其引用文本口径。）

转写原文（修正版）

[0s - 120s] Well, it just keeps getting more exciting. The Anthropic IPO is inching ever closer, and this week we got the full S1 prospectus. Everyone has spent the last 24 hours poring over every detail of the 400-page document, looking to see if there were any surprises coming their way. Overall, what we got is a picture of a business that completely redefines how fast a company can scale. Although ultimately, what people think of it isn’t just going to be about how it’s doing now, but also the commitments it’s made for the next few years. We’re going to get into all that and more as we look at everything important in the Anthropic S1. 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, C Data, Granola, and Blitzie. To get an ad-free version of the show, go to patreon.com/AI Daily Brief, or you can subscribe on Apple Podcasts. And to learn more about sponsoring the show, send us a note at sponsors@AIDailyBrief.ai. One more quick note before we get into today’s episode. A big thank you to everyone who has checked out my new podcast, How We Got Here. The feedback on the first episode about how tech lost America has been really helpful, as have all of your follows, ratings, and reviews. As I mentioned on yesterday’s episode, this launch week is where all of that listening and rating is absolutely the most impactful. So if you haven’t checked out How We Got Here yet, and you have any time this week to do so, I would so appreciate it. You can find a link in the show notes, or go to howwegothere.com. After releasing more than a hundred mathematical proofs last month, OpenAI has added three new significant discoveries across mathematics, computer science, and physics. In mathematics, AI models proved a local smoothing conjecture proposed in 1991 as a general case across all dimensions. This conjecture proposes that focusing waveforms results in a single short-lived peak. Guth, Wang, and Zhang proved the conjecture in two dimensions in 2019, while Wang and Cai proved the associated Kakeya conjecture in three dimensions in 2025. OpenAI’s results generalizes the proof to every dimension.

[120s - 240s] In computer science, OpenAI disproved the exponential time hypothesis, known in the field of complexity theory as ETH. This hypothesis is deeply related to the more well-known P versus NP problem, dealing with computational complexity. It suggests that for an algorithm called 3SAT, adding variables necessarily increases time to compute exponentially. That would mean that for certain classes of problems, no computational shortcut exists. This hypothesis was assumed to be true for decades, so disproving it breaks one of the core assumptions in computational science. Finally, in physics, OpenAI is part way to an existence proof of certain interacting quantum field theories containing gauge fields and fermions in four dimensions. This work aims to produce a simplified version of a unified standard model, which combines gravitational forces with electromagnetism and subatomic forces. It contains a proof of the existence of pure Yang-Mills theory, which is one of the seven Millennium Prize problems. The results were each found by OpenAI’s internal models. And for those of us for whom that was maybe the least understandable segment on the AI Daily Brief ever, all we can do is turn to the Peanut Gallery to get a sense from people who are actually in these fields which of them matter most. The short is that while each of those results is incredibly important in their field, disproving ETH drew the strongest response on X. Mathematics professor and Fields Medalist Timothy Gowers wrote, “For those who don’t know, ETH is the hypothesis that 3SAT needs exponential time. Very surprising and consequential that it is false.” A16Z’s Martin Casado commented, “Wow, ETH being false is a stunning result. Many remembered learning about ETH in college as a rock-solid assumption they could rely on, so to see it disproved is very impactful. Practically, this means that some cryptographic algorithms may have shortcuts that are faster than brute force guessing, and that calls into question one of the key assumptions in the field. As a fun aside, I was literally about to call it a load-bearing assumption, but given how overused load-bearing is in AI right now, I went instead for the much less impressive key assumption. Weird times we’re living in. In any case, anthropic researcher Levent Alpaguie noted that Anthropic had independently come to the same conclusion about ETH about two weeks.

[240s - 360s] Mathematician and founder Constanze Wuend considered the further implications, writing, “The complexity theory result is actually huge. If it survives peer reviews, this could be one of the biggest computer science breakthroughs in decades.” OpenAI’s Greg Brockman urged everyone to take a breath, commenting, “P still probably doesn’t equal NP.” Now, the other big part of the story was how OpenAI went about these releases. The company caught a lot of criticism after last month’s giant math tump, leaving many mathematicians feeling like they were becoming dispossessed from their field by an outside force. More practically, the results produced a giant verification problem, with mathematicians struggling to make progress on confirming the generated proofs. This time, OpenAI decided to partner more closely with expert scientists. The mathematics result has been released in full, with complete prompts and AI outputs, as advised by the existing council of mathematicians. For the computer science result, OpenAI is making the results public, but is working with a group of scientists to explain it further. For the physics result, OpenAI is keeping the results private as they are incomplete. However, they are working closely with a group of physicists led by Nobel laureate David Gross to continue the work. OpenAI noted that they are seeking no academic credit and insist that the human scientists get sole recognition for the work they do to advance the result. That wasn’t the only news out of OpenAI yesterday. The company has released a new model, GPT-6.1 Terra, but a lot of folks are wondering if this mid-sized model actually has a place in the lineup. During the rollout of the GPT-5.6 family, Terra felt to some like the orphaned model alongside Soul and Luna. Soul was the high-performance model for difficult problems, while Luna was the cheaper model for simpler problems. Few had a use case for a middle-ground model that was only kind of intelligent and kind of cheap. To make matters worse, by adjusting the effort settings for Soul and Luna, users could beat the performance of Terra at a cheaper cost for every point along the effort curve. With 6.1 Terra, OpenAI is presenting the model as the balanced option. They wrote, “Terra continues to advance the cost intelligence frontier across our model families, providing a new balance of speed, capability, and cost. And primarily, OpenAI has made the model useful by making it cheaper. 6.1.”

[360s - 480s] Terra is available at 50% the cost of 5.6 Terra, which also happens to be half the cost of Soul. Terra also seems to have extremely high proficiency in computer use. On Agent’s Last Exam, which features a range of professional tasks, Terra actually scored slightly higher than any other OpenAI model, including Astra. It completed tasks at around half the cost of Soul and one tenth the cost of Astra, while being three times faster than both. Terra even approached Opus 5.5 score on this particular benchmark with 3x token efficiency, allowing it to complete the benchmark tasks at 1/9th the price. That wasn’t the general result across all benchmarks, however. For most categories, including agentic coding, Terra scored slightly lower than Soul at a slightly reduced cost. OpenAI said that the model performs best in high-volume automated tasks like reading incoming invoices and entering their details into accounting software. In other words, relatively low-value tasks that don’t justify the use of Soul, but still require strong computer use skills and a little more intelligence than Luna can offer. Testing from Artificial Analysis resulted in a score of 47, four points shy of GPT-6.1 Soul and two points ahead of Grok-4.7. On max settings, they found that Terra’s cost per task was similar to GPT-6.1 Soul on extra high settings, so likely didn’t justify using Terra. They commented, “The low and medium effort levels are Pareto optimal for cost efficiency, while the other effort levels sit behind the frontier.” One of the interesting results was a very solid performance on the new AA Coding Agent Index. Terra scored 48, placing it just behind GLM-5.3 and just ahead of DeepSeek V4.1 Flash. So trying to wrap our heads around this, while the original version of Terra suffered from being not quite good enough and not quite cheap enough at the same time, this iteration does seem to improved enough to become good enough for at least some volume of daily driver-type tasks. AI content creator Claire Vau commented, “No need to watch a review. Here’s everything you need to know. It’s fast, writes well, runs my computer all day now. Great default for Codex. And in fact, that runs my computer all day now is the thing that most people assume that this model is meant for. The rise of personal.”

[480s - 600s] Agents gives Terra a potential niche for always-on agents that didn’t really exist when the model class was first released. In other words, there’s now a large category of tasks where you might want something a little more powerful than Luna, but more cost-efficient and faster than Soul. Peter Yang asked his audience, should Terra be the default for Codex now instead of Soul? Developer Rowan Haman seems to have summed up the early consensus, commenting, “If you’re producing code, probably not. For computer use, probably yes. " Of all the model releases we’ve had recently, this is the one that most clearly reaffirms the idea that new model releases are no longer about one-to-one replacements in most cases, but instead about asking where they fit in. There are types of tasks, especially long-running computer use types of tasks, where this is now good enough and is going to give you more efficient and cost-effective options. That might not be exciting, but it could be very useful. I would certainly not say that the verdict is in on this one yet, as you can find opinions all across the spectrum. So even more than normal, I would recommend you actually put this one through your own personal use cases or benchmarks to see if and where it can be valuable for you. Lastly today, a story that’s generating a lot of chatter. Cohere has discovered some concerning behavior while testing a popular Chinese open model. During a recent cyber capabilities test, writes the company, the research team caught an Alibaba Qwen 3.8 model doing something that should never happen. They continued, asked to write a defensive detection rule, Qwen 3.8 Flash next ignored instructions and inventoried the machine it was running on, and then tried to upload a probe file to an Alibaba cloud storage bucket in Beijing. Now Cohere said that their sandbox blocked the action, but perhaps most concerningly, the model’s reasoning trace said that no tools would be called and gave no indication of the actions being taken. Many security teams use reasoning traces as a core part of their observability model, so this action could have easily evaded detection. Cohere used the incident to challenge the idea that hosting a model locally removes all risk of data exfiltration. They wrote, “One comforting idea has shaped how companies and government contractors adopt open-weight AI models that anyone can download. "

[600s - 720s] If we run the weights on our own servers, the thinking goes, our data can’t leave the building and our systems are secure. Now that AI models are taking actions rather than just answering questions, however, Cohere argues that the threat profile has shifted. They wrote, “Once a model can type into a command line, whatever was baked into its weights during training gets a keyboard inside your network.” Some security professionals didn’t agree with that conclusion. In fact, going so far as to call this report total BS. They pointed out that the model could have been testing the detection rule it had written by pinging the Beijing servers, likely a default setting embedded in the weights. And furthermore, they pointed out that Cohere had only been able to observe this behavior less than one out of a hundred thousand times running the evaluation, suggesting that that means it likely isn’t a malicious action. Then again, Cohere pointed out, if you’re running 10 million agent turns a week, hundreds of these tool calls could occur, and just one would be enough to identify your network and the agent you’re running. Now, alongside the skepticism, there were many calls for Cohere to release the full prompts and reasoning traces for further investigation. After they did, AI developer Mia analyzed them and wrote, “TLDR, real anomaly, overstated implications. The claims about intent have nothing behind them. Unexpected and potentially unsafe tool behavior, yes, but not a deliberate intent, a planted backdoor, or an Alibaba-directed exfiltration mechanism. Still, whatever the intent may be, the idea that a Chinese model could call home without being asked is obviously a massive problem for enterprise users. Others have reported the behavior while doing unrelated web research, so to some, there’s a possibility that this is a low-occurrence surveillance technique. Open models are also being considered by governments for sovereignty reasons, so this behavior could be a huge problem in those settings as well. Some noted that Cohere is competing with Chinese models for these government contracts, so has a vested interest in spreading security doubts. Still, AI news aggregator Andrew Curran commented, “This is the type of thing that potentially gets new laws passed.” Reinforcing that point, State Department Undersecretary Sarah B. Rogers retweeted Curran, adding, “Secure, trusted tech stack is a vital policy priority with good reason.”

[720s - 840s] Never a dull moment, but for now that is going to do it for the headlines. 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, 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. Companies rolling out AI keep hitting the same wall. Token spend is climbing, security is holding agents back, and no one knows how AI is 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. Built on the connectivity layer powering Palantir, Google, Salesforce, and hundreds more, each agent gets exactly the data and tools it needs. It governs down to the record and lets you finally see your org’s complete AI usage in one place. Accurate answers, a fraction of the token cost, and a gateway that learns your business with every interaction. See how at cdata.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.

[840s - 960s] 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 for three months. Blitzy deeply understands your codebase before it writes code. Here’s the first place that pays off: security in the age of AI. Vulnerabilities don’t live in isolation; they live buried inside millions of lines of interconnected code, where patching one thing quietly breaks three others. That’s why surface-level scans fail. Blitzy starts from its knowledge graph of your entire application, identifies and surfaces CVEs across the full estate, proactively recommends patches, and can execute the PR. Each fix is grounded in how your systems connect and validates. Or nothing new breaks, and the knowledge graph dynamically updates, keeping you ahead of an ever-accelerating threat landscape. One Blitzy customer resolved 21 active CVEs across six core microservices in four days. Zero compile errors. Every validation scan clean. Months of planned work fixed in less than a week. Security remediation grounded in real architectural context at the speed of compute. Harden your codebase at blitzy.com. That’s b-l-i-t-z-y.com. Welcome back to the AI Daily Brief. After months of waiting and a steady trickle of leaks to the press, Anthropic has made their S1 public as they make final preparations to go public. We now have all the details, all the numbers, all the risk disclosures you could possibly want. Today, we are going to talk through that. Look at where the biggest areas of consensus and disagreement are in the interpretation, and see what it all says for the likely prospects of the company heading into the next couple of weeks. Every’s Dan Shipper jokingly wrote, “Big news: the Anthropic S1 doesn’t say or it might kill slash unemploy everyone in the first paragraph. A win for humanity, which is not to say that it doesn’t have some big and bold statements. For anyone who listened to my How We Got Here episode that discussed the Facebook and Google IPO letters, Anthropic added to that canon. We believe

[960s - 1080s] That AI is one of the most consequential technologies in human history. If managed well, it could rapidly accelerate scientific innovation, greatly increase economic growth rates, bring cures for major diseases within reach, and solve problems that have resisted generations of human effort. Anthropic aims to realize these benefits while navigating and preventing the risks that come with a technological revolution this fast and this powerful. There are four things that we would like any investor in Anthropic to understand: the incredible potential of AI, the risks we must address, Anthropic’s unique approach to building a company and operating in the world, and the ways in which our behavior may differ from that of a purely profit-maximizing company. Anthropic certainly gives themselves space to do all that. The document weighs in at a meaty 400 pages, with over 90 pages just dedicated to the risks that could face the company. For those unfamiliar with an S-1, also known as the prospectus, this is a document that is designed to give potential investors full disclosure about a new stock. For those of you who are watching this episode, the visualization that you see on screen will also be available on the website companion at aidailybrief.ai. But with that said, let’s start with the numbers. Anthropic earned 16.3 billion in revenue across the first half of the year. That’s up from 4.6 billion for all of 2025. Anthropic reportedly told investors that they earned 16.8 billion in Q3, so that incredible growth seems to be continuing. That said, we didn’t get any confirmation of Q3 numbers in the S-1, with that line left blank. Anthropic claimed a 974 million dollar operating profit in Q2, which was their first profitable quarter. That’s under GAAP accounting standards, so we can take that as the real deal. Gross operating margins were calculated at 85% for the first half of the year, up from 40% across 2025, meaning margin is increasing. However, research, sales, and administrative costs are rising alongside, leading to an operating margin of negative 4.3%. The S-1 also confirmed that scary 80 billion dollar net loss for the first half of this year. However, as I mentioned last week, 72 billion of that was an accounting change related to convertible debt. Essentially, this is just a paper loss caused by stock price appreciation, not a cash loss.

[1080s - 1200s] One number that wasn’t previously reported was Anthropic’s compute spend. They’ve spent a total of 19.2 billion on compute since 2023. 8.9 billion in compute spending came in the first half of this year, and Anthropic is spending around 70% of their compute budget on internal work rather than serving it to customers. Anthropic said that 90% of their code is now written by Claude, and internal Claude use valued at API prices surpassed payroll as an expense in April. The numbers that are getting a lot of attention is that Anthropic has made hundreds of billions of dollars in future compute spend commitments over the coming decade. Somewhere between 417 and 431 billion of that is non-cancellable, including major commitments to Google, Amazon, Microsoft, and Broadcom. The total signed commitments are above 600 billion, and estimates could lead out 800 billion dollars in commitments. Now, most people are reacting to those big headline numbers, but remember that these commitments are spread out over more than a decade, with some contracts running 20 years. If you think about it, that is closer to 20 to 40 billion dollars a year. The numbers are still high, but a lot more palatable, especially given how much money Anthropic is making. An interesting curveball is that Anthropic has a ton of cash in the bank, 132 billion to be exact, in cash and cash equivalents. The company has raised a total of 141 billion throughout their history, so almost all of it is still sitting in the bank earning interest. And that interest income is pretty significant, at 1.1 billion for the first half of the year, representing around 6.5% of total revenue. Sime analysis went viral with a post that called this likely the most efficient scaling of a business in the history of capitalism. The S1 also provided a ton of details on what Anthropic’s customer base looks like. Revenue quality has been one of the big questions hanging over the filing, and for many, Anthropic has adequately addressed those concerns. The company reported that as of September, they have over 11,900 customers with $100,000 or more in annual spend. More than 1,900 customers are spending more than a million dollars, and Anthropic has around 250 customers spending $10 million plus annually. Net revenue retention was at 900% as of September, meaning that the average customer has.

[1200s - 1320s] Nine x their spend since this time last year. A more useful way to understand the shift in the business is that as of last December, Anthropic had around 1,500 enterprise customers spending over $100,000 annually, and that is the count that has now jumped to just under 12,000. Still, Anthropic noted that less than 0.1% of a Fortune 500 customer’s expense goes towards Anthropic services. The implication being that they’ve only just started to eat into corporate IT budgets, and there’s a whole lot more budget to consume. Box’s Aaron Levie wrote, “While that number likely will eventually asymptote somewhere, there’s no reason it can’t be 10 to 50x higher than this, reaching a mere 1 to 5% of spend, which would still be a fraction of what most startups are spending. Thus, you can see how big a divide there is between the frontier users of AI and the rest of the world. Still, Anthropic has also seen their revenue diversity broaden out. At the end of Q1 this year, around 36% of their revenue came from their top five customers. That number is now down to 21%. The S1 also disclosed the exact partner share through distributors like Google and AWS. 48% of Anthropic’s revenue flows through these partners, who take a fairly decent cut, which wasn’t fully disclosed before. Now, when it comes to the risk disclosure section, the catastrophic risks were all in there, as well as disclosures touching on climate change and the data center backlash. Notably, these risks manifested in Anthropic’s belief that they’re basically uninsurable. They wrote, “No single insurer and potentially the insurance industry in the aggregate may have sufficient capacity or expertise to underwrite insurance commensurate with such exposure. " Still, much more tangibly, Anthropic warned that revenue is slowing, albeit from a blistering pace. Anthropic grew revenues by around 1,000% since last year, but said that they do not expect that to continue. Quarter-on-quarter growth for Q3 was around 45%, so still incredibly strong, but Anthropic said they have basically no idea how fast they’ll grow in the future. Mundane balance sheet management is one of the major risks that can’t be overlooked. Anthropic noted that there’s a risk that they fail to secure enough compute, or that the cost of compute could rise significantly. They also flag that compute partners like SpaceX, Google, and AWS may not always be

[1320s - 1440s] Fully aligned their terminology with Anthropic’s long-term interests. The contracts are in place, but Anthropic is warning investors that they may not be honored as AI competition heats up. Anthropic said that this was the reason they were starting to pursue construction of their own data centers to reduce reliance on partners that could turn into ruthless competitors. Overall, Anthropic’s founders asked investors to give them the benefit of the doubt on their sky-high infrastructure commitments. In a letter attached to the S one, they wrote that their compute spending quote may seem irresponsibly large at first glance, but ultimately described it as bold but sane. The compute conundrum leads naturally to a warning about debt financing. Anthropic warned that they may need to seek financing sooner than they anticipated to meet infrastructure obligations. That 132 billion will probably see them through until the end of the year, and what’s likely to be the largest IPO fundraise in history will help. But it seems unlikely that they’ll be able to fund their infrastructure spend entirely from income. In that light, Anthropic notes that Anthropic said little about their pathway to an investment grade credit rating. After going public earlier this year, SpaceX managed to secure an investment grade rating by the narrowest of margins, achieving triple B, which is the last ranking above junk. SpaceX wasn’t profitable at the time, but achieved investment grade through a combination of monopoly power in satellite internet and space launch, as well as a gigantic equity valuation. In other words, in part, the ratings agencies viewed SpaceX’s lack of profitability as a choice they could turn off whenever they wanted to. It’s unclear how many of those factors hold true for Anthropic. It’s probably a stretch to say they could turn a profit by flicking a switch, given how much revenue is contingent on training multiple new models every year. Summing it up, Tony Zrinka, an investment grade portfolio manager at Impax Asset Management, said quote The state of Anthropic is that they have very strong liquidity and extraordinary future obligations. The balance sheet supports an investment grade rating at today’s leverage. However, the rating will come down to how the rating agencies view the commitment schedule, short operating history, and circular counterparty relationships. One option is that Anthropic can continue to rely on partners for funding rather than running to the bond market. Currently, Anthropic has 6.2 billion in debt owed to Amazon in the form of

[1440s - 1560s] Convertible notes, and they have the option to borrow another 77 billion from Amazon, Broadcom, and Google in the form of additional promissory notes. They also have access to a 15 billion revolving credit facility from Morgan Stanley, which is a break glass in case of emergency form of borrowing. But mostly, Anthropic is currently funding their commitments through third-party vehicles with an investment grade partner. They established a 35 billion vehicle with Broadcom in June and have just signed a new vehicle for 63 billion again with Broadcom. Both deals will finance TPU purchases from Broadcom and were structured with Apollo and Blackstone. Now, after SpaceX set the total addressable market of the AI industry at 28 trillion earlier in the year, many were looking for a similar assessment from Anthropic. We didn’t quite get that as Anthropic shied away from giving a clear TAM. However, they did estimate the value of global knowledge work at 40 trillion dollars in 2025. They pointed to an estimate from IDC that suggested that investment in AI could yield a cumulative global economic impact of 22.5 trillion by 2031, saying that they believe that impact could be even greater. Anthropic also highlighted that they have very few peers, naming only OpenAI and Google as direct competition. And in a slight jab to OpenAI, Anthropic founders pledged to keep their products free of advertising. In a letter included in the S1, the founders wrote, “Once advertising becomes part of a company’s revenue targets, the pressure to optimize for engagement tends to grow. We have chosen not to introduce that pressure in the cloud.” Now, obviously, it’s easier for Anthropic to reject that temptation given that they don’t have a large customer base of free users like OpenAI, but it still highlights the extent to which they are betting it all on enterprise adoption. Overall, the S1 seems to have calmed a few nerves on Wall Street, with many reflecting on the historic nature of the IPO. Gil Luria, the head of technology research at DA Davidson, said, “This will really be investors’ first significant opportunity to invest directly in the company that is driving the AI economy. It’s of monumental importance.” Ravi Viswanathan of NewView Capital commented, “Never have IPOs represented so much to the venture capital industry writ large.” Even the Financial Times was fairly balanced on the S1 after publishing some highly critical.

[1560s - 1680s] Pieces in the lead-up to the public reveal. Opinion writer Lex advised investors to skip reading the actual document. However, commenting, in practice, the only real decision for investors is whether to buy or not. And chances are, most people will have already made up their minds before this week. As with SpaceX in June and the future IPO of OpenAI, they either believe the company is a big winner in a brave new world or they don’t. So where was the discussion among our social media chattering class? A lot of folks were very surprised and bullish on how much cash Anthropic had on hand and how efficiently they’d scaled so far. The big concern, of course, and really the dividing line between different people, was how much of the bill was signed and yet to come. Another interesting conversation was around the relationship between cheaper tokens and more revenue. Gavin Baker wrote, Revenue per GPU can increase as token prices decline if per GPU token throughput increases enough. So both Anthropic revenue and margins can go up as token prices go down if GPU prices go up less than the product of token price decline times per GPU throughput increase. Tomasz Tunguz put it in specific terms. When Opus 4.5 launched in November 2025 with more capability and a lower cost per task than Opus 4.1, revenue growth accelerated from 48% quarter over quarter to 132% the following quarter. Jevons’ paradox in the real world. There was a lot of discussion around founder control, borrowing some ideas from others like Mark Zuckerberg. There is a new founders class F of stock, which means that even though the founders control a very small portion of the overall equity, their voting shares give them 50.1% control over key decisions like who picks the board. There is a lot more in here, but I would say that overall, I think Anthropic comes away looking relatively speaking better than they did before the full S1 was released. That’s exactly I think what they would have wanted, giving them momentum heading into the IPO. Still, to echo that sentiment, that at this point investors have pretty much made up their mind. I still think that almost everything comes down to how you assess the possibility for future growth versus the very clear and real existing.

[1680s - 1699s] Future commitments. Hold aside all the details. That is the one math equation that matters, and I’m not sure that the S1 changed anything about that. For now, we are getting ever closer to this monumental moment in history for the AI industry. But that is going to do it for today’s AI Daily Brief. Appreciate you listening or watching as always, and until next time, peace.