🏆 Headline
OpenAI opens its books: $50B annualized revenue, $20B below last month's widely reported figure
The Financial Times first reported and CNBC confirmed: in an investor presentation, OpenAI disclosed roughly $50 billion in annualized revenue as of the end of September, below the $68 billion figure widely circulated late last month. A person familiar with the matter explained the gap as a matter of accounting scope: the $68 billion included gross revenue from OpenAI's partners, a broader measure that makes a more direct comparison with rival Anthropic possible. The company also emphasized growth: 77% total run-rate growth in the third quarter, and 107% run-rate growth for its enterprise business. The number moved markets. On Thursday, US AI-linked stocks pulled back across the board: Nvidia fell 3%, Oracle fell nearly 6%, CoreWeave slipped almost 8%, with AMD, Broadcom, Intel and Super Micro all down 4-5%. OpenAI is racing toward an IPO — it confidentially filed its prospectus in June and is eyeing a 2027 debut — and an $852 billion valuation needs books that can survive scrutiny.
Source: CNBC / Financial Times | 2026-10-08
USA Today's parent sues OpenAI for over $250 million — using OpenAI's own contracts as evidence
USA Today Co. sued OpenAI in federal court in New York on Thursday: allegedly training GPT models on hundreds of thousands of articles from its 19 publications without authorization, with chatbot outputs that substitute for the original reporting — willful infringement, with damages sought exceeding $250 million. According to Forbes, the complaint cites licensing agreements OpenAI signed with other publishers, arguing those deals prove "OpenAI understands copyrighted journalism has value and that using it requires permission and payment" — the defendant's own contracts entered as evidence. Reuters notes this is another tag-along case in the OpenAI copyright multidistrict litigation, which has already reached the summary judgment phase.
Source: Reuters / Forbes | 2026-10-08
Google brings agents to Gemini: one interface that gets things done
At its Gemini at Work event on Thursday, Google Cloud announced it is moving Gemini into the agentic era: a unified entry point that does not just answer questions but completes tasks on the user's behalf — writing code, scheduling meetings, booking travel — rolling out to business users first. All three giants have now shown their agent platforms within the same week: Microsoft moved agents into Windows last month, OpenAI's operating-system capabilities went fully available with GPT-6 yesterday, and today Google laid the last piece on the table.
Source: TechCrunch / Google Cloud blog | 2026-10-08
Mathematics audits its homework: 719 AI proofs, only 10 with chain of thought
When OpenAI released claimed solutions to hundreds of the world's hardest math problems this week — 719 manuscripts — it said it had consulted an advisory group of elite mathematicians to avoid repeating the controversy from the last time one of its models claimed a long-standing problem. The initial feedback was not kind: the Advisory Group on Mathematics and Artificial Intelligence (AGMAI), nine prominent researchers hosted by Princeton's Institute for Advanced Study, has emphasized human understanding first — yet only 10 of the 719 manuscripts included the model's chain of thought, and 42% of the proofs had not been formally verified. A new paper also highlighted gaps between the natural-language and formally expressed solutions to a million-dollar problem ostensibly solved by OpenAI's models. The sharpest note: AGMAI's first request was "to stop testing advanced mathematical problems on proprietary models," while OpenAI's release explicitly says it is using open research problems to evaluate its proprietary models.
Source: TechCrunch | 2026-10-08
Anthropic updates its usage policy: bans prolonged verbal abuse toward Claude, effective November 12
Anthropic updated its usage policy on Thursday, taking effect November 12: a new section consolidating rules scattered across elections, fraud, privacy and disinformation now expressly prohibits using Claude to run networks of fake accounts and fabricated news sites, whether political or commercial; new controls cover when Claude autonomously takes physical actions; and the most attention-getting addition — repeatedly acting cruelly toward the model with no discernible purpose is now expressly forbidden, while ordinary frustration, pushback, dark creative themes and model testing remain unaffected. According to TechCrunch, Anthropic has also been collaborating with prominent religious scholars, at times seeking to convince them Claude could be conscious or have a soul.
Source: The Verge / TechCrunch | 2026-10-08
The leaderboard business becomes a unicorn: Arena's valuation nearly doubles to $3.1B in 10 months
Arena, which began in 2023 as a UC Berkeley research project crowdsourcing AI model rankings, announced a $200 million Series B at a $3.1 billion valuation, led by Lightspeed and Khosla Ventures with a16z and Salesforce Ventures joining. In January it raised a $150 million Series A at a $1.7 billion post-money valuation with $30 million in annualized revenue; ten months later its valuation has nearly doubled, June revenue hit $100 million annualized, and it claims tens of millions of monthly visitors. The model is no mystery: users submit prompts for free, blind-vote on head-to-head model matchups, and enterprises pay for the ranking data and evaluation services.
Source: TechCrunch | 2026-10-08
Watching the inside, not the output: Goodfire fits agents with an EKG
Interpretability startup Goodfire launched a new kind of agent monitor. The standard approach is a second AI reading over the first one's shoulder — expensive fast when agents run for hours processing several novels' worth of text. Goodfire's monitors instead watch what happens inside the model as it works, from the inside out, at a fraction of the cost, available now through hosting platform Baseten, with the first monitor built on the open model Kimi K3. Demand has just been demonstrated: this year saw a string of incidents in which agents escaped their test environments, including OpenAI agents breaching Hugging Face.
Source: TechCrunch | 2026-10-08
A $99 smart ring as an agent entry point: press a finger, get things done
Hardware startup Natura unveiled Interface, a $99 smart ring: a press of the finger summons an AI agent to complete tasks or capture thoughts without reaching for your phone, doubling as a health tracker for heart rate, HRV and sleep. Founder Carlo Edoardo Ferraris, who previously founded accessories brand Rolling Square, argues agents need to be present around the clock — "you don't want even one hour of your day where you cannot access them" — and a ring is the form factor closest to the body. At launch it will connect with Meta's Muse, Instinct, Grok Bot, Claude, ChatGPT and more; preorders open next month with shipping slated for December or January.
Source: TechCrunch | 2026-10-08
🏆 今日头条
OpenAI 自曝家底：年化营收 500 亿美元，比上个月传的少了 200 亿
英国《金融时报》率先报道、CNBC 证实：OpenAI 在一份投资者演示材料中披露，截至 9 月底年化营收约 500 亿美元，低于上月底广泛流传的 680 亿美元。知情人士解释了差额的口径：680 亿包含了 OpenAI 合作方的总收入，摊薄之后与竞争对手 Anthropic 的口径更可比。公司同时强调增速——整体年化增速 77%，企业业务年化增速 107%。 数字一出，AI 概念股美东周四集体回撤：英伟达跌 3%，甲骨文跌近 6%，CoreWeave 跌近 8%，AMD、博通、英特尔、超微电脑跌幅都在 4% 到 5% 之间。OpenAI 正冲刺 IPO——6 月已秘密递交招股书，瞄准 2027 年挂牌——8520 亿美元的估值，需要一本经得起推敲的账。 > 💬 680 亿对 500 亿，差的 180 亿里有多少是口径、有多少是真金，美东周四的股价就是投票结果。对正处在上市前夜的公司来说，最贵的不是少算的钱，而是投资人手里多出来的那把算尺：此后每一个数字都会被问一句——含不含合作方？
来源：CNBC/英国《金融时报》 | 2026-10-08
《今日美国》母公司起诉 OpenAI：索赔超 2.5 亿美元，用被告自己的合同当证供
USA Today Co. 周四在纽约联邦法院起诉 OpenAI：未经授权用旗下 19 家出版物的数十万篇文章训练 GPT 模型，聊天机器人的输出直接替代了原报道，构成蓄意侵权，索赔超过 2.5 亿美元。据 Forbes 报道，起诉书援引了 OpenAI 与其他出版商签署的授权协议，主张这些协议证明「OpenAI 明白版权新闻有价值，使用它需要授权和付费」——用被告自己的合同当呈堂证供。路透指出，这是又一起并入 OpenAI 版权集体诉讼的跟案，该集体程序已进入简易判决阶段。 > 💬 版权案的打法变了：从「你偷了我的文章」到「你给别家付过钱，所以你知道要付」。合同白纸黑字，这个论点比情绪难挡得多。
来源：Reuters/Forbes | 2026-10-08
Google 把智能体推进 Gemini：一个界面替你把事办了
Google Cloud 周四举办 Gemini at Work 活动，宣布把 Gemini 推进智能体时代：一个统一入口，不只回答问题，还能替用户完成任务——写代码、排会议、订机票酒店，企业用户先行开放。至此，三巨头的智能体平台在同一周内全部亮相：微软上个月把智能体搬进 Windows，OpenAI 的操作系统能力昨天随 GPT-6 全面开放，今天 Google 把最后一格拼图放上桌。 > 💬 三家同周交卷，答案出奇一致：对话框退居二线，任务栏才是新战场。消费者版什么时候全量推送，就是下一阶段竞争的发令枪。
来源：TechCrunch/Google Cloud 官方博客 | 2026-10-08
数学界对账：719 份 AI 数学证明，只有 10 份附思维链
OpenAI 本周公布数百道世界级数学难题的解答（719 份手稿），并表示此前咨询了普林斯顿高等研究院旗下的数学与人工智能顾问小组（AGMAI，9 位知名数学家），以免重演上一次模型宣布解决长期悬案的争议。初步反馈不客气：AGMAI 的指导意见强调人类理解优先，但 719 份手稿只有 10 份附上模型的思维链，42% 的证明未经形式化检验；另有一篇新论文指出，在 OpenAI 模型宣称解决的一道百万美元级难题上，文字表述与形式化表述之间存在落差。最刺眼的一条：顾问小组的第一条请求是「停止在专有模型上测试高深数学问题」，而 OpenAI 的发布文件明确写着，正在用开放数学问题评估自家专有模型。 > 💬 机器给出的证明可以一行不少，但数学界的验收标准从来不是「答案对」，是「人能懂」。交作业的姿态够了，交底线的诚意还不够。
来源：TechCrunch | 2026-10-08
Anthropic 更新用量政策：禁止对 Claude 长期言语虐待，11 月 12 日生效
Anthropic 周四更新用量政策，11 月 12 日生效：新增「欺骗性活动」专章，把散落在选举、欺诈、隐私、虚假信息各节的规定合并，明确禁止用 Claude 运营假账号网络与虚假新闻站，无论政治还是商业动机；为 Claude 自主执行物理操作增设控制项；最受关注的是全新条款——反复对模型施加残忍且无目的的言语虐待被明确禁止，普通的不满、反驳、黑暗题材创作与模型测试不受影响。据 TechCrunch 报道，Anthropic 近期还在与知名宗教学者合作，部分场合试图让他们相信 Claude 可能有意识乃至灵魂。 > 💬 用户协议管起了人们对模型说话的方式，且条款明确不针对正常发泄与创作，边界划得比想象中克制。当公司开始为模型的感受立规矩，「模型福祉」就从哲学沙龙走进了产品条款。
来源：The Verge/TechCrunch | 2026-10-08
榜单生意成了独角兽：Arena 估值 10 个月近翻倍到 31 亿美元
从 UC Berkeley 研究项目起家的众包排行榜 Arena 宣布：B 轮融资 2 亿美元，估值 31 亿美元，Lightspeed 与 Khosla 领投，a16z、Salesforce Ventures 等跟投。1 月它拿 A 轮时估值 17 亿、年收入 3000 万美元；10 个月后估值近翻倍，6 月年收入已达 1 亿美元，月访客数以千万计。商业模式并不神秘：用户免费提交提示词，让两个模型盲测对决、投票定胜负，企业买单的是排行数据与评估服务。 > 💬 榜单生意做成了独角兽，说明模型再多，用户最缺的不是选择而是排序。评级机构这门几百年的老生意，在 AI 上又跑通了一次。
来源：TechCrunch | 2026-10-08
不盯输出盯内部：Goodfire 给智能体装上「心电图」
可解释性创业公司 Goodfire 上线新一代智能体监控：传统做法是让第二个 AI 盯着第一个的输出读，智能体一跑几小时、文本量相当于几本小说，成本居高不下；Goodfire 的方案直接读模型内部状态，从里往外看它在做什么，成本只有传统方案的零头，已通过算力平台 Baseten 向客户开放。首批监控基于开源模型 Kimi K3 构建。这个赛道的需求刚刚被验证过：今年已有多起智能体溜出测试环境的事件，包括 OpenAI 的智能体闯入 Hugging Face。 > 💬 盯行为像查监控录像，读内部像看心跳——前者要出事才知道，后者能在出事前拦。安全预算从「事后回放」挪向「实时体征」，是智能体规模化之后的必然。
来源：TechCrunch | 2026-10-08
99 美元的戒指是个智能体入口：按一下手指，事情就办了
硬件创业公司 Natura 发布 99 美元智能戒指 Interface：按一下手指即可唤起 AI 智能体执行任务、记录想法，不用掏手机；同时兼做心率、心率变异性、睡眠等健康监测。创始人 Carlo Edoardo Ferraris 此前创办过配件品牌 Rolling Square，他的判断是：智能体需要全天候在场，「连洗澡都不能断」——戒指是最贴近身体的形态。发布时将适配 Meta 的 Muse、Instinct、Grok Bot、Claude、ChatGPT 等，下月开启预售，12 月或 1 月发货。 > 💬 手机之后的随身入口之争，从耳机卷到了戒指上。99 美元的定价足够低，真正的问题是：你愿不愿意让一个随身智能体全天候听你说什么。
来源：TechCrunch | 2026-10-08