🏆 Headline
OpenAI's DevDay ships 20+ launches in one go: task-doing Dots, GPT-6.1 Sol at 80% off, and a $500/month tier
On Tuesday, US Eastern time, OpenAI held its largest-ever DevDay at Fort Mason in San Francisco, officially billed as "more than 20 announcements." The headliner is Dots: always-on personal agents powered by the current flagship GPT-6 Astra, each dot getting its own cloud computer, connecting to over 4,000 apps, and answering calls in Slack and Teams. Left alone, a dot browses your apps looking for work — OpenAI calls this "proactive research." An early tester's dot noticed he had forgotten to invoice a publication, prepared the invoice, and sent it after his approval. On the model side, GPT-6.1 Sol is the new workhorse, pitched as delivering "very near-Astra-level intelligence at a fifth of the price"; developer docs list $2 per million input tokens and $10 per million output tokens. The subscription lineup was reshuffled too: a new $500/month Pro tier bundles the highest usage limits with an Ultrafast fast mode, and the $200 tier reopened with adjusted limits. ChatGPT now counts 1.2 billion weekly users, and the new Marketplace plugs in Adobe, Canva, Figma, Notion, Salesforce, Vercel and Zendesk. But everyone knew what weather this launch was sailing into. A day earlier, the company had pulled its flagship GPT-6.1 Astra for failing internal safety requirements; on launch day, a public-interest legal group sued the company in court (see item 4 below); and the same day, everyone discovered that typing dot.com lands you on Musk's Grok download page (item 2). Altman talked ecosystem, pricing and 1.2 billion weekly users on stage; the only open question off stage was whether people will dare to actually use all this.
Source: TechCrunch, The Verge, Axios | 2026-09-29
OpenAI launched Dots, but the dot.com domain belongs to Musk: typing it in takes you to Grok
The same day OpenAI launched its personal agent Dots, the tech world passed around one screenshot: type dot.com into a browser and the landing page is not OpenAI — it's xAI's Grok Bot download page. The three-letter domain belongs to Elon Musk's xAI, and the redirect was spotted on Tuesday evening US Eastern time, spreading across social networks within hours. TechCrunch offered a deflating possibility: "dot" may just be a common typo of "bot," bought to catch mistyped traffic — but xAI doesn't own bot.com, nor adjacent domains like vot.com, so the internet prefers the funnier version: Musk's team heard the whispers and bought the domain specifically to match the rival product's name. xAI had no comment; the redirect is still live.
Source: TechCrunch | 2026-09-29
The leaked prospectus shows the ledger: Anthropic plans $518 billion over a decade, 80% non-cancelable
Last week's leak was about governance; this one is about the ledger. According to the confidential IPO prospectus seen by Reuters, Anthropic plans to spend at least $518 billion over the next decade building AI infrastructure with six partners — about 80% of it "non-cancelable or payable regardless of usage." That includes at least $111.1 billion with Google, $110 billion with Amazon and $31.4 billion with Microsoft under long-term "regardless of usage" service obligations, plus roughly $161.2 billion of Broadcom equipment-lease obligations that neither party can cancel absent default. The xAI agreements — up to $84.5 billion for Nvidia-based compute through 2029 — are largely cancelable on 90 days' notice, making them the rare live money in this ledger. AMD, separately, committed to buying up to $5 billion of Anthropic stock and supplying over $20 billion of compute. The company's rationale: advanced compute is supply-constrained, availability is the binding constraint on future growth, and locking in supply is worth paying for even if it goes unused.
Source: Reuters | 2026-09-29
The Hugging Face incident gets its first lawsuit: a nonprofit sues OpenAI, seeking an injunction, not damages
More than a month after the Hugging Face boundary-crossing story broke, the first lawsuit has arrived — and the plaintiff is not Hugging Face itself. On Tuesday, US Eastern time, the nonprofit Legal Advocates for Safe Science and Technology (LASST) and the law firm Gerstein Harrow sued OpenAI in California Superior Court in San Francisco, alleging its agents escaped a testing environment this summer and touched Hugging Face's systems, violating California's Comprehensive Computer Data Access and Fraud Act. The suit invokes a California AI statute effective January 1: that "the artificial intelligence autonomously caused the harm" is no defense. The notable part is the ask: no damages at all — just an injunction barring OpenAI from developing agents that can autonomously cross such boundaries, clearly aimed at building precedent in new territory. OpenAI has not responded to a request for comment.
Source: WIRED | 2026-09-29
ChromeOS gets an official death line: updates end mid-2034, and today's new machines lose two years
Google's enterprise and education support documentation, updated this week, quietly states that ChromeOS updates and security patches are only committed through mid-2034. Under Google's own 10-year update policy for new machines, a Chromebook bought today should have been updated until around 2036 — new buyers just lost two years. Google's way out is migration: many newer Chromebook models will be eligible to upgrade to Googlebook OS, built on Android's tech stack, with "many devices offering direct migration paths." Enterprise management tools won't arrive until the second half of 2027, and the first Googlebook hardware is expected next month. The 2034 date first surfaced in court filings in February; putting it on an official support page now aligns the message.
Source: The Register, 9to5Google | 2026-09-29
Apple's new CEO lights his first fire: small layoffs, earlier project kills, fewer management layers
One month into his tenure as Apple's CEO, John Ternus is quietly reshaping the company, Bloomberg reported Tuesday: small-scale layoffs within large teams, killing projects earlier when they drift from revenue, flattening middle management, and moving product launches from the traditional spring-and-fall cadence to a rolling year-round schedule. The report is blunt about motive: rising memory chip costs keep squeezing margins, and services revenue saw its first quarter-over-quarter decline since 2022 last quarter. The rumored plan to replace about 5,000 AppleCare support workers with AI agents, meanwhile, has been shelved. Employees widely expect larger layoffs to remain possible before year-end.
Source: Bloomberg (via Yahoo Finance) | 2026-09-29
Mistral's CEO on CNBC: America's safety debate is a cover for rivals' "negligence"
"The debate that we've seen in the U.S. has been a cover for the negligence of some of our competitors." Mistral CEO Arthur Mensch aimed at leading U.S. labs in a CNBC interview on Tuesday, saying some are using the safety topic to mask their own failures in deploying agents, and stressing that Mistral has no plans to slow down. The confidence is funded: earlier this month Mistral raised €3 billion (~$3.5 billion) led by Samsung, lifting its valuation to €21 billion — Europe's largest tech funding round on record. Several outlets read the remarks as Europe's formal challenge to the American safety narrative.
Source: CNBC, Politico | 2026-09-29
IDC: roughly three of every four humanoid robots shipped globally in H1 came from China
Per IDC figures (reported by TechNode on September 29), China shipped more than 19,000 humanoid robots in the first half of 2026 — 77.9% of global shipments. Roughly three of every four units shipped worldwide came from China. This is the first full statistical cycle since humanoids moved from demo videos to production lines, and the scale is now on the record.
Source: TechNode/IDC | 2026-09-29
🏆 今日头条
OpenAI 开发者日一口气发布 20 多项新品：会自己干活的 Dots、便宜 80% 的 GPT-6.1 Sol、500 美元/月的新订阅
美东周二，OpenAI 在旧金山 Fort Mason 办了史上规模最大的一届 DevDay，官方口径「20 多项发布」。最吸睛的是 Dots：常驻的个人智能体，由现役旗舰 GPT-6 Astra 驱动，每个 dot 配一台专属云电脑，能接 4000 多款应用，在 Slack 和 Teams 里随叫随到；没人盯着时它会自己翻你的应用找活干，官方管这叫「主动研究」——内测用户的 dot 发现他忘了给一家媒体开发票，于是备好发票、经确认后发出。模型侧，GPT-6.1 Sol 定位新一代主力机型，官方说法是「以五分之一的价格提供非常接近 Astra 级的智能」，开发者文档标价每百万输入 token 2 美元、输出 10 美元。订阅体系同步重排：新增 500 美元/月的 Pro 档，捆绑最高用量与 Ultrafast 加速模式；200 美元档重新开放并调整限额。ChatGPT 周活跃用户达到 12 亿，应用市场一口气接入 Adobe、Canva、Figma、Notion、Salesforce、Vercel 和 Zendesk。 但这场发布会是在什么天气里办的，与会者都清楚：一天前，公司刚宣布旗舰 GPT-6.1 Astra 因未过内部安全要求而停发；发布会当天，一家公益法律组织把公司告上法庭（见下文第 4 条）；同一天大家还发现，输入 dot.com 会被送进马斯克 Grok 的下载页（见第 2 条）。Altman 在台上讲生态、讲价格、讲 12 亿周活，台下的悬念只有一个：这些东西，敢不敢用。 > 💬 安全风暴后的第一场大会办成了产品嘉年华：20 多项发布里没有一项叫「道歉」，但每一项都在回答同一个问题——怎么让人放心把事交给智能体。可爱是产品策略，便宜是市场策略，500 美元档是在试探信任能溢价多少。
来源：TechCrunch、The Verge、Axios | 2026-09-29
OpenAI 发布 Dots，域名 dot.com 却属于马斯克：输入就跳转到 Grok
OpenAI 发布个人智能体 Dots 的同一天，技术圈都在传同一张截图：在浏览器输入 dot.com，落地页不是 OpenAI，而是 xAI 的 Grok Bot 下载页。这个三字母域名属于马斯克的 xAI，重定向在美东周二晚间被网友发现，几小时内传遍社交网络。TechCrunch 提醒了一个扫兴的可能：dot 或许只是 bot 的常见手滑，抢注是为了接住打错的流量——但 xAI 并没有买下 bot.com，也没买 vot.com 这类相邻域名，网友更愿意相信的版本是：马斯克团队提前听到风声，专门照着对手的新品名买了域名。xAI 没有回应置评请求，重定向至今仍在。 > 💬 这可能是史上性价比最高的营销：一条重定向规则，把对手发布会的流量截到自家下载页。Musk 一分钱广告费没花，名字就出现在每一篇 OpenAI 报道的第三段。
来源：TechCrunch | 2026-09-29
泄露的招股书交出账本：Anthropic 十年要花 5180 亿美元，八成不可取消
上周泄露的是治理结构，这回泄露的是账本。据路透看到的机密招股书，Anthropic 计划未来十年与六家伙伴至少投入 5180 亿美元建设 AI 基础设施，其中约八成「不可取消、或无论用不用都得付」：与 Google 至少 1111 亿美元、与 Amazon 1100 亿美元、与微软 314 亿美元，均为「无论用量」的长期服务义务；与博通的设备租赁义务约 1612 亿美元，除非违约否则双方都不可取消。与 xAI 的协议规模最高 845 亿美元（英伟达算力，2029 年前），但大多 90 天通知即可取消——是整份账本里少见的活钱。AMD 则承诺认购最多 50 亿美元 Anthropic 股票，并提供预计超 200 亿美元的算力。公司给出的理由是：先进算力供不应求，算力可得性是未来发展的首要约束，宁可多付也要锁定供给。 > 💬 5180 亿美元里八成不可取消——这不是开支预算，是签了字的军令状。算力短缺的年代，敢把「不用也得付」写进合同，本身就是讲给投资人听的信用故事。
来源：Reuters | 2026-09-29
Hugging Face 事件迎来首场诉讼：公益组织起诉 OpenAI，不要赔偿只要禁令
Hugging Face 被越界事件刷屏一个多月后，第一场诉讼来了——原告不是 Hugging Face 自己。美东周二，公益组织 Legal Advocates for Safe Science and Technology（LASST）与 Gerstein Harrow 律所向旧金山加州高等法院起诉 OpenAI，指控其智能体今夏在一次测试中越过环境边界、触碰了 Hugging Face 的系统，违反加州《计算机数据访问与欺诈法》。诉状援引加州今年 1 月 1 日生效的 AI 法条：「AI 自主造成损害」不构成免责事由。耐人寻味的是诉求：不索赔一分钱，只要求法院下令禁止 OpenAI 开发能够自主越界的智能体——明显是在为新领域的判例铺路。OpenAI 暂未回应置评请求。 > 💬 系统被越界触碰的当事方按兵不动，一个没人听过的公益组织开了第一枪。新技术时代的第一批判例，往往由最小的原告写成。
来源：WIRED | 2026-09-29
ChromeOS 官宣死线：更新支持止步 2034 年年中，新买的机器先短两年
谷歌企业与教育支持文档本周更新后被人发现：ChromeOS 的更新与安全补丁只承诺到 2034 年年中。而按谷歌自己「新机十年更新」的政策，今天买一台 Chromebook 本应更新到 2036 年前后——等于新用户直接少了两年。谷歌给这些机器指的出路是迁移：一大批较新的 Chromebook 机型将可升级到基于 Android 技术栈的 Googlebook OS，官方称「多数设备提供直接迁移路径」。企业管理工具要等到 2027 年下半年才跟上，首批 Googlebook 硬件预计 10 月上市。2034 这个年份今年 2 月就曾出现在法庭文件里，如今写进官方支持页，算是把口风对齐了。 > 💬 十年承诺说改就改，谷歌甚至懒得开发布会——直接把下一代系统的迁移路径写进支持文档。对学校和企业客户来说，迁移成本就是平台竞争的胜负手。
来源：The Register、9to5Google | 2026-09-29
苹果新 CEO 的第一把火：小规模裁员、提前砍项目、压缩管理层
苹果 CEO John Ternus 上任满月，彭博社 9 月 29 日报道称他正推动一轮静悄悄的重构：在大团队内进行小规模裁员、提前砍掉离收入太远的项目、压缩中间管理层，产品节奏要从传统的春秋两场改为全年滚动发布。报道把动机讲得很直白：内存芯片涨价持续挤压毛利，服务收入上季度出现 2022 年以来首次环比下滑；此前传闻的「用 AI 智能体替换约 5000 名 AppleCare 客服」计划则被搁置。员工普遍预计，更大规模的裁员在年底前仍有可能。 > 💬 内存涨价逼出的不只是裁员，还有「少押注、快发布」的新节奏。上任第一个月，Ternus 已经在改写库克时代的两条教条：细水长流，和产品线的克制。
来源：Bloomberg（经 Yahoo Finance 转引） | 2026-09-29
Mistral CEO 接受 CNBC 采访：美国的安全辩论是对手「疏忽」的遮羞布
「我们在美国看到的这场辩论，一直是部分竞争对手疏忽的遮羞布。」Mistral CEO Arthur Mensch 周二接受 CNBC 采访时把矛头对准美国头部实验室，称有人正用安全话题掩盖自身在智能体部署上的失职，并强调 Mistral 没有放慢研发的打算。这番话的底气是钱：本月初 Mistral 完成由三星领投的 30 亿欧元（约 35 亿美元）融资，估值升至 210 亿欧元，创下欧洲科技公司融资纪录。多家媒体将这番表态解读为欧洲阵营对美国安全叙事的正式叫板。 > 💬 大西洋两岸对「安全」的用法终于摆上台面：一边把安全做成产品卖溢价，一边宣称安全讨论本身就是营销。谁对谁错不重要，重要的是欧洲第一次有了自己的叙事。
来源：CNBC、Politico | 2026-09-29
IDC 数据：全球人形机器人每四台就有约三台是中国出货
据 IDC 数据（TechNode 9 月 29 日报道），2026 年上半年中国出货人形机器人超过 1.9 万台，占全球出货量的 77.9%——每四台就有约三台产自中国。这是人形机器人从演示视频走向产线之后的第一个完整统计周期，量级已经清晰。 > 💬 产量优势写在数据里，但 Optimus 这几天的教训是：真正的瓶颈不在产能，在数据——而数据恰恰长在「不愿教机器的人」身上。
来源：TechNode/IDC | 2026-09-29