AI周刊第543期:谷歌成为最新一家追赶TypeSafe成立8周的Jev的科技巨头

内容来源:https://aiweekly.co/issues/google-is-the-latest-tech-giant-to-chase-typesafes-8-week
内容总结:
决策模型赛道升温:谷歌、OpenAI、微软相继入局,中国团队表现亮眼
本报讯 一种不生成文本、只输出最优答案及概率的新型AI模型——“决策模型”正快速崛起,成为继聊天机器人之后大模型厂商争夺的新战场。9月15日,TypeSafe AI率先发布首个决策模型Jev,谷歌于本周三推出自研版本,OpenAI与微软则已于10月先后入场。三家巨头的定价高度一致:每百万输入token收费不超过0.10美元,输出端完全免费。
中国团队已在市场中占据一席之地。 上海人工智能实验室称,其40亿参数的Intern-Decision 4B模型在论文测试中以90.02%的平均得分超越Jev v1.13.0的88.74%,单张RTX 4090显卡上处理三问请求仅需44毫秒。微软的决策模型则基于阿里巴巴Qwen3.5-9B进行后训练,输入价格为每百万token 0.042美元。此外,Nace AI以开源方式发布四款Drex决策模型,公开全部权重与推理代码,其中Drex Max在Decision Index 0.4.1上得分67.52,领先Jev的60.70。
资本与开发者生态同步升温。 TypeSafe AI在发布数周后即完成由Andreessen Horowitz领投、红杉与DCVC参与的8.7亿美元融资,估值达75亿美元,CEO Diogo Almeida曾任OpenAI研究员。开发者社区方面,Jev成为Vercel有史以来采用最快的模型——上线两周内,Vercel AI网关付费团队中使用率达35%,约为第二名Sonnet 5.5的近两倍;Vercel自身也已用Jev替换GPT-6 Luna进行有害指令分类。另有开发者发布可在本地运行的Jev-Style 2B v4,部分训练数据来自Jev公开API。
生产环境中的分工正在形成。 Vercel月度报告显示,其AI网关月处理token超100万亿,生产级AI正分化为“生成”与“判断”两类任务。10月网关上的决策模型增至8家厂商15款,占全部请求的26%,但仅占支出的0.53%,单次请求成本约为语言模型的七十分之一。近三分之一付费团队已使用决策模型,团队规模越大使用率越高:月用量超100亿token的团队中七成已采用。不过收入仍集中在语言模型一侧——团队每在Jev上花费1美元,在Anthropic模型上花费64美元。
高难度判断仍需语言模型。 Artificial Analysis决策指数v2显示,Claude Haiku 5.5以77分、GPT-6 Luna以76分领先最佳决策模型Drex Max v1.1的73分,但成本高达15倍、耗时悬殊:GPT-6 Luna每百万任务需1530美元、11.34秒,Drex Max仅需102美元、275毫秒。为防基准测试被“刷分”,该指数中已有45%使用私有数据,高于此前的25%。
业内建议,若代码仅需从固定列表中选答案,可优先尝试决策模型;最困难的判断仍应保留给语言模型。值得注意的是,约11%的Vercel付费团队只使用决策模型、完全未调用文本生成模型。
中文翻译:
本期通讯聚焦决策模型——一类不生成文本的新型AI模型。你给它一个输入和一份固定的答案列表,它便在毫秒间返回最佳答案及其概率。TypeSafe AI于9月15日发布了首个此类模型Jev。周三,谷歌也推出了自己的决策模型,此前OpenAI和微软已于10月相继入局,三家均按每百万输入token收费0.10美元或更低,输出不收费。上海AI实验室称其4B模型在自家测试中击败了Jev,而微软的模型则是基于阿里巴巴的Qwen构建的。在最难的决策上,聊天机器人得分仍略高,但成本是前者的15倍。居中的是:在Vercel的网关上,10月份决策模型承接了四分之一的AI请求。
赞助商
Arca为你的AI会话付费
手头堆着几百个Claude Code、Codex或Cursor会话闲置没用?不如拿来换钱。行业动态
当下AI圈什么最火——从开发者信息流到模型排行榜。
- Jev是Vercel有史以来监测到采用速度最快的模型。据TechCrunch报道,上线两周内,Vercel AI网关上35%的付费团队已在用Jev,覆盖面几乎是第二受欢迎模型Sonnet 5.5的两倍。Vercel自己也将GPT-6 Luna换成了Jev,用于对有害指令进行分类。
- TypeSafe的发布让开发者开始讨论“信任一个数字”这件事。Earendil首席技术官Armin Ronacher对TechCrunch表示,50%的答案“可能就是抛硬币”,他会忽略;而95%的答案则是他可以据此行动的。他还认为决策模型可以用来决定由哪个模型来处理某项任务。
- 一位开发者发布了一款可在本地机器上运行的Jev风格模型。据其维护者闫朝亮介绍,Jev-Style 2B v4的部分训练数据来自Jev公开API收集的答案。他与TypeSafe没有任何关联。
- InfoQ为工程师们详细解读了这一新格式。你发送一个状态和一组带类型的问题,Jev便一次性返回所有答案——一个选择、一个分数或一个是/否概率。TypeSafe称之为“系统一”模型。
- 免费开源的决策模型如今已在公开排行榜上超越Jev。Nace AI发布了四个Drex决策模型,附带全部权重和推理代码,同时也通过OpenRouter提供服务。Drex Max在决策指数0.4.1上得分67.52,而Jev 1.14.0为60.70。快讯
实验室角斗士时代 - 谷歌周三发布了自己的决策模型。Gemini Decision 1和全新的Decisions API于11月11日上线,支持对文本、图像、音频、视频和PDF进行是/否、选择和评分标准决策。其返回的是经过校准的决策结果,输出token为零,因此没有文本需要解析。
- OpenAI向所有客户开放了其版本,基于加速版GPT-6 Luna构建。据TechCrunch报道,OpenAI于10月22日全面开放Decisions API。产品负责人Nikunj Handa表示,答案生成时间已从1到3秒缩短至100到180毫秒。AI语言辅导应用Speak将操作间延迟从1.4秒降至0.13秒。
- 开创这一品类的初创公司在发布数周后融资8.7亿美元。Jev的开发商TypeSafe AI在由Andreessen Horowitz领投、Sequoia和DCVC参投的一轮融资后估值达75亿美元。CEO Diogo Almeida此前曾在OpenAI担任研究员。DeepSeek的悄然接管
- 上海AI实验室的4B模型在论文测试中击败了Jev。Intern-Decision 4B在七个测试套件上平均得分90.02%,而Jev v1.13.0为88.74%。在单块RTX 4090上,它回答一个三问题请求仅需44毫秒,更小版本约需33毫秒。
- 微软的决策模型基于阿里巴巴的Qwen构建。微软对Qwen3.5-9B进行了后训练,制成Microsoft-Decision-1,并表示很快将把模型迁移到其他底座上,包括自家的MAI模型和OpenAI的模型。输入每百万token收费0.042美元,输出免费。全面自动模式
- Liquid AI的开源决策模型在游戏级GPU上8毫秒即可给出答案。Liquid称,d1-3B在决策指数上得分48.57,领先所有参数量低于10B的模型。在RTX 4090上耗时8毫秒,在Jetson Orin Nano上耗时50毫秒,权重可免费下载。
- 聊天机器人在决策上得分仍然更高,但成本是前者的15倍。在Artificial Analysis的决策指数v2上,Claude Haiku 5.5得分77,GPT-6 Luna得分76,比最佳决策模型Drex Max v1.1的73分高出约3分。GPT-6 Luna每百万任务成本1,530美元,而Drex Max为102美元;每任务耗时11.34秒,而Drex Max为275毫秒。该指数目前有45%使用私有数据(此前为25%),以防基准测试被针对性优化。生产级AI正分化为两类工作
Vercel每月通过其AI网关路由超过100万亿token,其11月报告指出,生产级AI正在分化为两类工作:生成与判断。
决策模型于9月中旬登陆该网关。到10月,已有来自八家厂商的15个决策模型,承接了全部请求的26%,但仅占支出的0.53%。按每次请求计算,其成本约为语言模型的七十分之一。
10月份近三分之一的付费团队使用了决策模型。团队越大,使用比例越高:月用量低于1000万token的团队中略超五分之一使用了决策模型,而月用量超过100亿token的团队中这一比例高达十分之七。
这些流量几乎全部流向了Jev,其单独获得的请求量接近所有Anthropic模型总和的四倍。钱则流向了另一个方向:团队每在Jev上花1美元,就在Anthropic上花64美元。
这正是各家实验室当前争夺的分化格局。语言模型仍然赚走大部分收入,而软件内部廉价、快速的调用正转向另一类模型。谷歌对Gemini Decision 1每百万输入token收费0.10美元,输出不收费。OpenAI的Decisions API同样收费0.10美元,不收取输出、缓存读取或缓存写入费用。微软每百万输入token收费0.042美元,输出免费。核心要点 - 如果你的代码需要聊天机器人从列表中做选择,试试决策模型。谷歌、OpenAI和微软现在都在卖,每百万输入token收费0.10美元或更低,输出不收费。
- 中国力量已进入这一市场。上海AI实验室的Intern-Decision 4B在论文测试中击败了Jev,微软的模型也基于阿里巴巴的Qwen构建。
- 最难的调用仍留给语言模型。Claude Haiku 5.5和GPT-6 Luna仍在Artificial Analysis决策指数上领先约3分,但成本高15倍,耗时也长得多。值得一读
- MMDM,一个涵盖28个决策模型的基准测试:最佳模型在有多个正确答案的难题上仅得36.81%,Perplexity的Decider从世界知识题的78.75%骤降至约束规划题的47.95%。
- Jeva,一个从零设计的决策模型:作者认为决策模型在重新排列选项时应给出相同答案,并且应该能够回答“以上都不是”。
- The Decoder对Jev的初探:TypeSafe如何推销一个对选项打分而非回答客户的模型,以及它的局限在哪里。等等,什么?
- Jev以那位说“更便宜的东西会被用得更多”的经济学家命名。据TechCrunch报道,威廉·斯坦利·杰文斯描述了价格下降如何使人们越来越多地使用某种商品,而TypeSafe预计廉价智能终将无处不在。
- 约11%的Vercel付费团队只使用决策模型,完全不用文本模型。Vercel表示,10月份这些团队没有向任何生成文本的模型发送过一个请求。本周投票
你已经在用决策模型了吗?
上周有413位读者投票:
本周哪个结果最能告诉你AI的走向?
你已经在用决策模型了吗?
下次见,
Alexis
英文来源:
This issue is about decision models, a new kind of AI model that does not write text. You give it an input and a fixed list of answers, and it returns the best answer with a probability, in milliseconds. TypeSafe AI released the first one, Jev, on September 15. On Wednesday Google released its own, after OpenAI and Microsoft did in October, and all three charge $0.10 or less per million input tokens and nothing for output. Shanghai AI Lab says its 4B model beat Jev on its tests, and Microsoft built its model on Alibaba's Qwen. Chatbots still score a little higher on the hardest decisions, at 15 times the cost. In the middle: on Vercel's gateway, decision models took one in four AI requests in October.
Sponsor
Arca pays for your AI sessions
Got hundreds of Claude Code, Codex or Cursor sessions just sitting there? Might as well get paid for them.In the Wild
What's trending in AI right now, from the developer feeds to the model charts.
- Jev was the fastest-adopted model Vercel has ever measured. Within two weeks of launch, 35% of paid teams on Vercel's AI Gateway were using it, nearly double the reach of the second most popular model, Sonnet 5.5, TechCrunch reported. Vercel itself swapped Jev in for GPT-6 Luna to classify harmful commands.
- TypeSafe's launch had developers talking about trusting a number. Armin Ronacher, CTO of Earendil, told TechCrunch that a 50% answer is "maybe a coin toss" he would ignore, while a 95% answer is one he can act on. He also sees it picking which model should handle a job.
- A developer released a Jev-style model you can run on your own machine. Jev-Style 2B v4 was trained in part on answers collected from Jev's public API, its maintainer Chaoliang Yan says. He has no tie to TypeSafe.
- InfoQ walked engineers through the new format. You send a state and a set of typed questions, and Jev answers them all in one pass with a choice, a score or a yes/no probability. TypeSafe calls these System One models.
- Free, open decision models now beat Jev on a public index. Nace AI released four Drex decision models with all the weights and inference code, and they are also served through OpenRouter. Drex Max scores 67.52 on Decision Index 0.4.1, against 60.70 for Jev 1.14.0.
Quick Hits
The Lab Gladiator Era - Google released its own decision model on Wednesday. Gemini Decision 1 and a new Decisions API went live on November 11 for yes/no, choice and rubric decisions across text, images, audio, video and PDFs. Its answers are calibrated decisions with zero output tokens, so there is no text to parse.
- OpenAI opened its version to every customer, built on a sped-up GPT-6 Luna. OpenAI made its Decisions API widely available on October 22, TechCrunch reported. Product lead Nikunj Handa said answers now take 100 to 180 milliseconds instead of one to three seconds. Speak, which makes AI language tutors, cut the time between actions from 1.4 seconds to 0.13 seconds.
- The startup that started the category raised $870 million weeks after launch. TypeSafe AI, maker of Jev, is now valued at $7.5 billion after a round led by Andreessen Horowitz, with Sequoia and DCVC. CEO Diogo Almeida previously worked as a researcher at OpenAI.
DeepSeek's Quiet Takeover - Shanghai AI Lab's 4B model beat Jev on the paper's tests. Intern-Decision 4B averaged 90.02% across seven test suites, against 88.74% for Jev v1.13.0. It answers a three-question request in 44 milliseconds on one RTX 4090, and the smaller versions take about 33 milliseconds.
- Microsoft's decision model is built on Alibaba's Qwen. Microsoft post-trained Qwen3.5-9B to make Microsoft-Decision-1 and says it will soon rebase the model on others, including its own MAI models and OpenAI's. Input costs $0.042 per million tokens and output is free.
Auto Mode Everything - Liquid AI's open decision model answers in 8 milliseconds on a gaming GPU. d1-3B scores 48.57 on the Decision Index, ahead of every model under 10B parameters, Liquid said. It takes 8 ms on an RTX 4090 and 50 ms on a Jetson Orin Nano, and the weights are free to download.
- Chatbots still score higher on decisions, at 15 times the cost. On Artificial Analysis' Decision Index v2, Claude Haiku 5.5 scores 77 and GPT-6 Luna 76, about 3 points ahead of the best decision model, Drex Max v1.1, at 73. GPT-6 Luna costs $1,530 per million tasks against $102 for Drex Max, and takes 11.34 seconds per task against 275 milliseconds. 45% of the index now uses private data, up from 25%, to guard against benchmark gaming.
Production AI just split into two jobs
Vercel routes more than 100 trillion tokens a month through its AI Gateway, and its November report says production AI is splitting into two jobs: generation and judgment.
Decision models arrived on the gateway in mid-September. By October there were 15 of them from eight vendors, and they took 26% of all requests while accounting for 0.53% of spend. Per request, they cost about a seventieth as much as a language model.
Nearly a third of paid teams used one in October. The bigger the team, the more likely: just over one in five teams running fewer than 10 million tokens a month used a decision model, and seven in ten teams running more than 10 billion did.
Almost all of that volume went to Jev, which alone got nearly four times as many requests as all Anthropic models. The money still went the other way: for every $1 teams spent on Jev, they spent $64 on Anthropic.
That is the split the labs are now fighting over. Language models still earn most of the money, and the cheap, fast calls inside software now go to a different kind of model. Google charges $0.10 per million input tokens for Gemini Decision 1, with no charge for output. OpenAI charges the same $0.10 for its Decisions API, with no output, cache-read or cache-write charges. Microsoft charges $0.042 per million input tokens and nothing for output.
Key Takeaways - If your code asks a chatbot to pick from a list, try a decision model. Google, OpenAI and Microsoft all sell one now, for $0.10 or less per million input tokens and nothing for output.
- Chinese work is already inside this market. Shanghai AI Lab's Intern-Decision 4B beat Jev on its paper's tests, and Microsoft built its model on Alibaba's Qwen.
- Keep a language model for the hardest calls. Claude Haiku 5.5 and GPT-6 Luna still lead Artificial Analysis' decision index by about 3 points, but cost 15 times as much and take far longer.
Worth Reading - MMDM, a benchmark of 28 decision models: the best model gets only 36.81% on hard questions with several right answers, and Perplexity's Decider drops from 78.75% on world knowledge to 47.95% on constraint planning.
- Jeva, a decision model designed from scratch: the authors argue a decision model should give the same answer when you reorder the options and should be able to say "none of these."
- The Decoder's first look at Jev: how TypeSafe pitched a model that scores options instead of answering customers, and where its limits are.
Wait, What? - Jev is named after the economist who said cheaper things get used more. William Stanley Jevons described how a falling price can make people use a commodity more and more, TechCrunch reported, and TypeSafe expects cheap intelligence to end up everywhere.
- About 11% of Vercel's paid teams used a decision model and no text model at all. In October those teams did not send a single request to a model that writes text, Vercel said.
This week's poll
Are you using a decision model yet?
Last week, 413 of you voted:
Which result this week will tell you the most about AI?
Are you using a decision model yet?
Until next time,
Alexis
文章标题:AI周刊第543期:谷歌成为最新一家追赶TypeSafe成立8周的Jev的科技巨头
文章链接:https://news.qimuai.cn/?post=5566
本站文章均为原创,未经授权请勿用于任何商业用途