Daily ·
Anthropic prices IPO at US$83; Gemini 4 Neon ships
Anthropic prices its record IPO at US$83 a share, Google ships Gemini 4 Neon, Microsoft swaps in seven in-house models, and banks put agent-driven losses on the customer.
Ten stories from the last twenty-four hours: the largest listing in market history gets its price, two Google model launches land inside three hours of each other, and the banks quietly finish rewriting who pays when an agent gets it wrong.
- Anthropic prices its IPO at US$83 a share, above range — Reported pricing comes in above the US$70 to US$78 range the filing implied. If Anthropic sells all its shares at that price it raises more than US$68 billion, and with US$26 billion of private placements the offering tops US$94 billion — past SpaceX's US$86 billion listing in June. Shares start trading Thursday on Nasdaq under the ticker "ANT", against a US$1.4 trillion valuation implied by last week's filing. SiliconANGLE · Nov 18
- Google ships Gemini 4 Neon, its fourth new Gemini in a month — Positioned for complex software engineering, knowledge work and science workflows. It scores 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level — third overall, behind larger and slower models — and ranks second on the Vals Index at 70.9%. Priced at US$3 per million input tokens (US$0.15 cached, a 95% discount) and US$15 output, with a 1 million-token context window. Google · Nov 18
- Microsoft swaps in seven in-house models for always-on agents — Microsoft AI unveiled the batch at Ignite across code, security, images, speech and routing, positioned on price-to-performance against its own partner's lineup. Microsoft says MAI Code 2 averages 98% of GPT-6.1 Sol's performance across four benchmarks at 40% of the cost, and 96% of Claude Opus 5.5 at 5% of the cost. In GitHub Copilot it runs US$0.50 per million input tokens and US$2.50 output at the promotional price through May 17, 2027, then US$1 and US$5. The Decoder · Nov 18
- Hugging Face routes model requests through OpenRouter — Inference Providers now picks a provider per call on price, speed and reliability, with automatic failover; a ":fastest" suffix picks highest throughput and ":cheapest" the lowest price. Bring-your-own API keys stop working for language models, though they remain for image and video. OpenRouter reports capacity errors down 46%, time-to-first-token up 16.9% and throughput up 16.3%. The Decoder · Nov 18
- Snapchat lets AI coding assistants build AR Lenses — Easy Lens CLI lets creators describe a Lens in plain language and have Claude Code, Cursor or Codex build, test and publish it, drawing on more than 100 AR building blocks and 50 AI asset generators. Snap says the tool can take a Web Lens game from idea to published in a single day, and that it paid US$1.8 million to Lens+ community creators last month. TechCrunch · Nov 18
- Banks put agent-driven account losses on the customer — Citi updated its Online User Agreement this week: customers who grant account access to third-party AI agents are responsible for what those agents do — including, per the terms, losses if the tool leaks credentials or is hacked. JPMorgan and Bank of America set similar policies last month, PNC changed terms today, and Citizens emailed customers this week with the same warning. TechCrunch · Nov 18
- Project Genie opens Genie 4 with real-time audio and remixing — Google AI Ultra subscribers get the new world model: prompt mid-session to shift weather, drop in a character or redirect the action, with native audio that adapts as the scene unfolds. Genie 3 stays in the model picker for longer runs and panorama-seeded spaces, and a community gallery lets users share and remix worlds. AI Pro members can join a waitlist. Google · Nov 18
- Reson8 releases Rumble-2-Flash as free open weights — A 500-million-parameter transcription model covering more than 40 languages under Apache 2.0, aimed squarely at non-English accuracy. Reson8 reports a 7.09% average word error rate across seven languages (~5% in English) and 17.6% on ten lower-resource languages, against 39.82% for Whisper Large V3 on the same set. The API is priced from €0.01 per hour of audio. SiliconANGLE · Nov 18
- Trump pushes a Pax Silica chip hub in the Philippines — The US-led pact on chips and critical minerals has at least two dozen signatories, and its flagship proposal is a 1,620-hectare (4,000-acre) zone for chip factories, mineral processing and data centres, with a framework agreement expected by the end of the year. Indigenous Aeta communities and protesters who met Trump's plane argue the zone displaces residents and cuts them off from water. The Verge · Nov 18
- House panel demands a Microsoft briefing on a reported ByteDance deal — The Select Committee on China asked Microsoft on Wednesday to brief lawmakers on a report that it is nearing a US$500 million deal letting ByteDance use OpenAI models through Azure. Chair John Moolenaar called it unacceptable for a Pentagon contractor; Microsoft says it takes misuse of its AI services seriously and does not comment on confidential negotiations. CNA · Nov 19
Try This Next
- 🔧 Download Rumble-2-Flash and test it on a non-English recording — 500 million parameters, Apache 2.0, no API key needed. SiliconANGLE
- 🎥 Open Project Genie and try re-prompting a world mid-session, watching how the audio tracks the edit. Google
- 📝 Write the agent-liability clause into your product docs before shipping anything that touches a bank account — the banks have already decided where the loss lands. TechCrunch
每日简报:Anthropic IPO 定价每股 83 美元;Gemini 4 Neon 发布
Anthropic 以每股 83 美元为创纪录的 IPO 定价,谷歌发布 Gemini 4 Neon,微软换上七款自研模型,银行把智能体造成的损失算在客户头上。
过去二十四小时的十条要闻:史上最大规模上市敲定价格,谷歌在三个小时内连发两款模型,而银行悄然改写了智能体出错时由谁买单。
- Anthropic IPO 定价每股 83 美元,高于区间 — 据报道,定价高于招股文件暗示的 70 至 78 美元区间。若全部股份按此价售出,Anthropic 将募资逾 680 亿美元;加上 260 亿美元的私募配售,整体发行规模超过 940 亿美元,超过 SpaceX 六月 860 亿美元的上市。股票周四在纳斯达克以「ANT」为代码开始交易,对应上周文件暗示的 1.4 万亿美元估值。 SiliconANGLE · 11月18日
- 谷歌发布 Gemini 4 Neon,一个月内第四款新 Gemini — 面向复杂软件工程、知识工作与科研流程。在其 xhigh 思考档位上,Artificial Analysis 智能指数得分为 61,总体第三,落后于更大更慢的模型;Vals 指数以 70.9% 排名第二。定价为每百万输入 token 3 美元(缓存输入 0.15 美元,折扣 95%),输出 15 美元,上下文窗口 100 万 token。 Google · 11月18日
- 微软以七款自研模型替换外部模型,瞄准常驻智能体 — 微软 AI 在 Ignite 上一口气发布覆盖代码、安全、图像、语音与路由的模型,以性价比对标其合作伙伴的产品线。微软称 MAI Code 2 在四项基准上平均达到 GPT-6.1 Sol 的 98% 性能,成本为其 40%;对 Claude Opus 5.5 达到 96% 的性能,成本仅 5%。在 GitHub Copilot 中,促销价至 2027 年 5 月 17 日为每百万输入 token 0.50 美元、输出 2.50 美元,之后为 1 美元与 5 美元。 The Decoder · 11月18日
- Hugging Face 将模型请求改由 OpenRouter 路由 — Inference Providers 现在按价格、速度与可靠性逐次选择供应商,并支持故障自动切换;「:fastest」后缀选吞吐最高者,「:cheapest」选价格最低者。语言模型的自带 API 密钥停止使用,图像与视频仍可自带。OpenRouter 称容量错误下降 46%,首 token 时间改善 16.9%,吞吐提升 16.3%。 The Decoder · 11月18日
- Snapchat 允许用 AI 编程助手制作 AR 滤镜 — Easy Lens CLI 让创作者用自然语言描述滤镜,由 Claude Code、Cursor 或 Codex 完成构建、测试与发布,可调用 100 多个 AR 构建模块与 50 个 AI 素材生成器。Snap 称该工具能让一款 Web Lens 游戏在一天内从想法走到上线,并称上月向 Lens+ 社区创作者支付了 180 万美元。 TechCrunch · 11月18日
- 银行把智能体造成的账户损失算在客户头上 — 花旗本周更新在线用户协议:授权第三方 AI 智能体访问账户的客户,需为智能体的行为负责——按条款,包括工具泄露凭证或被黑造成的损失。摩根大通与美国银行上月已出台类似政策,PNC 今日修改条款,Citizens 本周也向客户发出同样警告。 TechCrunch · 11月18日
- Project Genie 开放 Genie 4,支持实时音频与再创作 — Google AI Ultra 订阅用户可用上新的世界模型:在体验过程中随时输入提示,改变天气、加入角色或改变剧情走向,并配有随场景演变的原生音频。Genie 3 仍保留在模型选择器中,用于更长时长与全景种子空间;新的社区画廊可分享与再创作世界。AI Pro 会员可加入候补名单。 Google · 11月18日
- Reson8 以开放权重免费发布 Rumble-2-Flash — 这是一个 5 亿参数的转写模型,支持 40 多种语言,以 Apache 2.0 许可发布,直指非英语准确率。Reson8 称在七种语言上平均词错率 7.09%(英语约 5%),在十种低资源语言上为 17.6%,而 Whisper Large V3 在同一测试上为 39.82%。API 价格自每音频小时 0.01 欧元起。 SiliconANGLE · 11月18日
- 特朗普在菲律宾推动 Pax Silica 芯片枢纽 — 这一美国主导的芯片与关键矿产协定已有至少二十多个签署国,其旗舰提案是占地 1,620 公顷(4,000 英亩)的园区,涵盖芯片工厂、矿产加工与数据中心,框架协议预计年底前签署。原住民艾塔族社区与在机场迎接特朗普的抗议者认为,该园区将迫使居民迁离并切断其水源。 The Verge · 11月18日
- 美国众议院委员会要求微软就字节跳动交易作出说明 — 众议院中国问题特别委员会周三要求微软就一项报道向议员作简报:微软即将达成 5 亿美元交易,让字节跳动通过 Azure 使用 OpenAI 模型。委员会主席 John Moolenaar 称,对一家五角大楼承包商而言这不可接受;微软表示严肃对待其 AI 服务被滥用的风险,且不评论保密客户谈判。 CNA · 11月19日
接下来可以试试
- 🔧 下载 Rumble-2-Flash,用一段非英语录音试试——5 亿参数,Apache 2.0,无需 API 密钥。 SiliconANGLE
- 🎥 打开 Project Genie,在体验中途重新输入提示,观察音频如何跟随改动。 Google
- 📝 在发布任何接触银行账户的产品之前,先把智能体责任条款写进产品文档——银行已经决定了损失落在谁头上。 TechCrunch
More Dailies
Content on Anagnorisis is summarized, paraphrased, and editorialized from publicly available sources for length and clarity. Original sources are linked where available. All trademarks belong to their respective owners.
