周末新闻少,今天三条,都是这周中发布、我们前几期没写到的:Google的新中档模型,OpenAI自己公布的芯片设计成绩,ChatGPT dots往Plus用户放。每条都对过公司原文和独立媒体报道,对不上或没查到的地方文里写出来。
先补一个昨天的缺口。昨天我们写Anthropic周五的收盘价没查到,CNBC后来的报道写的是:“The shares sold off on Friday, closing at $88.30”,离83美元的发行价不远。同一篇还写到Nvidia:“Nvidia shares fell for a seventh straight weekly, losing 46% of their value over that stretch.”
没放进来的两件事。安全公司Dragos的研究负责人HD Moore发文说,他们向AI服务发了约两百万条埋有追踪凭证的提示词,发现“data from at least three AI providers was exposed”,但文中没点名是哪几家,我们也没找到能打开的独立报道,先不写。SEC对Nvidia展开民事调查的消息,源头是《华尔街日报》11月17日的报道,SEC发言人只说“does not comment on the existence or nonexistence of a possible investigation”,Nvidia的回应我们没查到,等有正式文件再写。
01 Gemini 4 Neon上线,每百万token 3/15美元
Google 11月18日发布Gemini 4 Neon。博客原文:“Gemini 4 Neon is rolling out today in the Gemini API in Google AI Studio”,另外还有Gemini for business,Antigravity里给“paid Google AI Pro and Ultra subscribers”,Gemini App里给Pro和Ultra订阅用户。免费用户这次没份。
价格是这条最值得看的地方。原文:“The model is priced at $3 per million input tokens”,后半句是“$15 per million output tokens, with a 1 million-token context window.” 缓存输入每百万0.15美元。9to5Google做了对比:“Neon is priced at $3/1M input tokens and $15/1M output tokens, or twice as expensive as Hydrogen.” 而Argon和Carbon现在是促销价,9to5Google写的是:“Argon and Carbon cost $2/1M input and $10/1M output until November 30.” 促销结束后:“it will be $4/1M input and $20/1M output, with Neon 25% cheaper as a result.”
定位上,Google的说法是Carbon管最难的长任务,Hydrogen管日常快速执行,“Neon balances depth and efficiency.” 成本的说法更具体:“At its default high thinking level, it pairs a high Artificial Analysis Intelligence Index score with a 32% lower cost per task than Carbon.”
跑分Google列了一长串,挑几个:Artificial Analysis Intelligence Index“#3 overall with a score of 61”,Artificial Analysis Coding Agent Index“#1 overall”(58.6),Terminal-Bench 4.1是65.2%。这些都是Google在自家博客里引用的数字,9to5Google转述的也是同一组,没有额外的独立测试。
还有个面向安全团队的版本:“Gemini 4 Neon for cyber is rolling out to trusted defenders in the coming days as part of the Fairwind Program.” 要申请得找Google Cloud客户团队。
我的判断:这次真正的变化是价格结构。12月起,如果你在用Carbon,同样的API调用换成Neon,单价直接便宜四分之一;Google还声称每个任务的成本低32%,但这个数字依赖它在什么任务上测、思考档位怎么设。别只看榜单排名,拿自己线上真实的10到20个请求,在Neon和Carbon上各跑一遍,比较结果质量和总token消耗,再决定要不要切。地区限制两个来源都没写,澳洲开发者用API之前先在AI Studio里确认模型列表。
来源:Google博客;9to5Google。
02 OpenAI模型改写AMD GPU电路,PPA最多+14%
OpenAI 11月18日发了一篇讲芯片设计的博文,最受关注的是和AMD的合作。原文:“We have partnered with AMD on PPA improvements and seen up to 14% improvement relative to heavily optimized baselines on the flagship GPU IP.” PPA是芯片设计里的功耗、性能、面积三项指标。时间上,OpenAI说“This represents roughly a year of progress achieved in less than 5 months”。
做法是改代码,不是改架构:“Our model identified further improvements by rewriting the code that describes the chip’s circuits (SystemVerilog RTL).” 每次改写都要过验证:“It used Synopsys tools to verify that each rewrite preserved the block’s behavior”,通过验证的方案再比较功耗、性能、面积。
落地程度:“AMD has already included some of these changes in a recent design it sent for manufacturing”,另一部分进了未来版本的生产设计文件。CPU那边也有:“AMD also verified single-digit percentage gains in PPA across key blocks of its upcoming next-generation flagship CPU”。注意这里是个位数,14%只对应GPU。
同一篇博文还提到OpenAI自己的推理芯片Jalapeño,原文是“These optimizations were silicon proven during Jalapeño’s silicon bring-up.” AlphaSignal的说法是Jalapeño是“OpenAI's Broadcom-built inference chip”。也就是说,真正在成品芯片上测过的是Jalapeño,不是AMD的GPU。
要打折扣的地方,AlphaSignal列得很清楚:“OpenAI has not published the block-level baselines”;14%是未具名GPU模块上的最好结果,不等于整颗GPU快14%;而且数字出自一家“which has commercial and financial interests tied to AMD”的公司。Analytics India Magazine也提到AMD此前给过OpenAI最多1.6亿股的认股权证。
我的判断:芯片RTL优化过去靠资深工程师一轮轮手调,周期以年计。这件事说明,只要有严格的形式化验证兜底,模型就能在这种「改错代价极高」的代码上干活。对写软件的人,可以直接借用这个模式:让模型大胆改,用测试和验证工具卡住正确性,只采纳通过验证的改动。至于14%,等AMD那颗芯片出来、有第三方实测再下结论。
来源:OpenAI博客;AlphaSignal;Analytics India Magazine。
03 ChatGPT dots开始给Plus用户测试
dots是OpenAI 9月推出的常驻Agent。帮助中心的定义:“Your dot is an always-on agent in ChatGPT that can take on ongoing responsibility”,能在对话之间继续推进任务。之前只给高价套餐,产品页写的是“Access is available now across Pro, Business Premium, and Enterprise plans”。
11月19日9to5Mac报道,OpenAI工程负责人在X上说开始向“a small number of ChatGPT Plus users”开放测试。帮助中心的资格写得更细:“selected Plus users aged 18 and older in markets excluding the European Economic Area, Switzerland, and the UK.” 澳洲不在排除名单里,但这是灰度,帮助中心说得直白:“OpenAI Support is unable to enable it for you, and restarting the app won't give you access.”
团队版也在放:“Dots are rolling out to eligible ChatGPT Business Standard members in the United States.” 有用量限制,原文就一句“Usage limits apply.” 企业版的时间表:“Starting November 18, 2026, dots become available by default unless your admin opts out.” 免费期到明年1月13日,之后“Starting January 14, 2027, billing begins for all enabled dots.”
使用入口:iOS、Android的ChatGPT App、桌面App和桌面网页都能创建,“Dots are not supported on mobile web.” 还有几条限制:从Pro降到Plus会失去dot;“Your dot cannot initiate calls to you at launch.”;暂时不能给dot单独的邮箱地址。
Plus用户能用多少,目前没公开。9to5Mac的原话:“It's not yet clear whether ChatGPT Plus and Business Standard users will have the same Dots usage allowances”。
我的判断:常驻Agent从每月几百美元的套餐往20美元档放,说明OpenAI对它的算力成本心里有底了,也想尽快拿到更多真实使用数据。如果你在澳洲用Plus、刚好被选中,先拿一件低风险的长期任务试,比如每周整理某个主题的资料,不要一上来就交给它涉及付款或对外发消息的事。企业管理员要注意:dots已经默认开启,1月14日开始计费,不想用的话现在就去后台关掉。