谷歌又发布了新的版本 而这次与上一代Gemini 更新的时间相差无几 甚至有些反常 当人们看到这个内容的时候 他们的第一反应大概就是 谷歌又要开始堆版本的事情了 但是你如果仔细观察Gemini 这次更新的方向的话 就会发现事情并没有那么简单 由于谷歌此次想要解决的问题并非是 如何再次创造出一个跑分第一的模型 而是更加实际的一个问题 人工智能到底要怎么做到真正的 进入到人们的日常工作当中去呢 是开发人员每天都可以愉快的使用它 并且能够让企业长久的支付费用呢 因此Gemini这次并没有把所有的注意力 集中在哪个模型更聪明的问题上 而是在逐步加强代码开发、工具使用 和复杂的任务处理等方面的工作 agent场景下稳定性、效率方面的表现 也就是说 谷歌想把Gemini从一个你 偶尔发几个问题给它的AI中 变成一个能够长久的 为你服务的主力模型 这次的变化看上去只是 对软件的一个版本升级 但是它所蕴含的意义却 远远大于参数提升本身 因为在人工智能代理真正 的投入到工作中去之后 一个任务往往要调用 模型十几遍到几十遍不等 此时人们比较的已经 不是哪一个模型最厉害呢 而会变成哪一个模型可以以较低 的成本稳定的性能完成工作呢 这样就会影响到 人工智能行业的未来竞争模式 因此今天我们这期的内容 就不再只是单纯的去聊 一下Gemini新版到底有多牛了 而是从这次更新的背 后来拆开来思考3个问题 为什么谷歌会忽然如此 重视一个不一定要追求第一 但是非常擅长做事的模型呢 在agent时代的背景下 人工智能的价格是否可以不再仅仅 以每百万个TOKEN多少钱来衡量呢 另外谷歌、Openai、Deepseek等公司 接下来又将会怎样进行竞争呢 如果能解决好这三方面的问题的话 那么你就能够发现 谷歌此次发布的新模型 并非只是单纯的创新 在于它已经提前预判到了人工智能 下一阶段最大应用场景的可能性 大家好我叫财经翻译官 本期我们从Gemini这次 的新版本发布说起 把谷歌想要抢走的东西 一层层的拆开来给你看 再说一下这次更新中 最不符合常理的部分 过去的几年里 在人工智能行业的发布会上 大家都在比的是谁能在最 困难的基准测试中拿到第一名 或者解出其他模型无法 解答的数学题和推理题 好像模型只要足够聪明 就可以一统江湖了 但是这次Gemini3.7Flash并没有 想要争夺最强大脑的称号 谷歌给自己定的位置是Workhorse model 即主力干活的模型 它并不追求在极难的问题上获得胜利 它的目标就是解决开发人员每天 都会频繁使用的各种真实的场景问题 编写业务代码 读取很长的文件 调用各种API 自动浏览网页查找资料 处理日常工作中的各项事务 这些事情并不酷炫 也不能登上科技新闻头条 但是他们是开发人员 每天都要做的80%的工作 定位的变化要比性能提高更加重要 这说明整个行业对 考核的标准发生了变化 以前人们比的是谁的回答最聪明 而现在要开始比的是 谁能用更便宜的价格 在更多的步骤中遇到错误时 能够自己纠正 重新规划路线 就如公司在招聘的时候 最聪明的大学毕业生 当然是最好的人选了 但是真正能够把日常 工作做的既快又好 不需要经常有人来收拾烂摊子的人才 才是团队中最不可或缺的核心力量 我跟很多做人工智能 应用的朋友交流过 他们之前遇到的最大 难题就是这样一个问题 很多模型跑分看起来很厉害 但是真正放到生产环境中使用 就会出现各种各样的问题 让其写出一个完整的功能 它可以帮到你遗漏了依赖包 让其修改现有的项目代码 它可以把其他正常的功能搞砸 对于需求不明确的地方 它不会提问 全部依靠自己的猜测来完成工作 并且所产出的结果也无法使用 最后开发人员还要花费更多 的精力来处理这个问题 从而降低工作效率 此次Gemini3.7Flash的出现 显然是为了应对上述问题而设计的 我们可以看一下一些 公开的基准测试数据 就可以知道它所要 提高的方向非常明确 按照Frontier Code 1.1代码基准来计算的话 在3.6版本中得分为34.4% 而在43.6% 在更加接近真实的工程 项目中进行的DEPSE 1.1测试上 从49%直接上升到65.3% WebDevArena 面向网页开发的比赛 已经上升到1,588分 用终端命令行来测试 TerminalBench 2.1的时候 得到的成绩是85.8% 最极端的情况就是Oswald 2.0 即测试人工智能控制电脑桌面的能力 该分数由原来的3.8%猛增到现在的47.9% 之前用的模型在电脑上 操作的基本都是点错按钮 开错软件的水平 现在可以独立完成 将近一半的桌面端工作 例如整理文件、使用软件、填写表格等 对于agent来说 这是实实在在的进步 但是谷歌并没有掩饰自己的缺点 在更难的TerminalBench 3.0测试中 谷歌只获得了14.9%的成绩 小于GPT-5.6terror的20.8%的情况下 在极端复杂工程任务上 并不是行业的顶尖水平 但是对于大多数 开发者日常的需求而言 它的能力已经足够使用了 并且更加懂得怎样把事情做好、做对 针对性的加强了一些在 实际开发中经常用到的能力 严格按照字面意思来执行 并且不会随意添加 一些没有被要求的功能 对于需求不明确的 地方会主动去询问 并不是自己瞎猜 当发现路径走不通 或者错误很多时 就会及时止损 不会一条路走到黑 浪费大量的时间与计算资源 这些特点在跑分中是无法体现出来的 但是真正使用的时候 体验上的差别就会很明显 讲完了性能之后 我们就来聊聊这次更新中最为 核心的一点agent时代经济学 也就是人工智能模型成本账 在算法上已经发生了变化 人们比较模型的价格时 只关心每百个TOKEN多少钱 哪个更便宜就用哪个 在之前的单轮对话中是正确的 当你提出一个问题的时候 模型就会回答一次 并且计算的非常准确 但是到了agent代理的时代之后 该算法就完全不起作用了 agent不是一次回答一个问题就结束了 而是一直的去完成一系列的任务 首先要弄清楚目标是什么 然后调动相应的工具来实现这个目标 接着读取相关的文件 编写代码并进行测试 一旦发现问题就会及时修正和重试 整个过程中可能会使用 到十几乃至几十个模型 此时衡量成本的主要指标不再 是每百万个TOKEN的价格 而变成了完成一个完整的 任务所需要的所有费用 有些模型的价格看上去很低廉 但是第一回做不正确的话 就要反复尝试几次 期间还要调用很多没有用到过的工具 最后算总账的时候 费用反而要比单个模型的价格高一些 但是只要有一次能够做到 的事情的模型就值钱多了 就比如你要请个装修工人 一个工钱便宜 但是干活粗糙 装完了还要拆掉重新安装三次 最后花费的钱和时间都比 找一个工钱稍微高一点 但是只用一次就可以 完成的工作的人要多得多 Gemini3.7Flash本次优化的重点是 提高第一次就做对的概率 降低不必要的工具使用 以及无意义的重复尝试 它不一定是最便宜的型号 但是只要能减少重试次数 那么完成同一项工作 的总成本就会降低 这才是企业用户真正关心的账单 毕竟企业在使用人工智能的时候 是为了降低成本 提高效率 并不是为了购买最低价的代币 按照这样的成本考量来分析 它所采取的价格策略的话 就可以很容易的理解 到谷歌的战略意图了 官方公布的售价为 在2026年底之前是推广价 输入TOKEN每百万0.75美元 输出TOKEN每百万3.75美元 自2027年1月1日开始实行正式价格 在上涨至1.5美元的时候 输出上涨到7.5美元 也就是翻倍 从价格上来看 它要比很多开源模型要贵一些 但是结合其性能以及 任务完成率来考量的话 其实它的性价比是非常高的 谷歌给出这样的价格 并不是依靠该模型来短期内获利 而是有两重明确的战略目的 第一项任务就是夺取 开发者默认的选择权 现在的AI开发迁移成本其实要 比人们想象中的要高一些 当开发者习惯于某 一个模型回答的方式 并且把提示词库工具链工 作流都建立在该模型之上时 就不会轻易更换其他模型了 更换一次就需要重新调整和测试 费用很高 目前谷歌采用的是接近 成本的价格进行推广 目的是为了让开发者以较 低的成本使用到产品中去 在开发者习惯之后再逐步 绑定到Gemini生态系统上来 即使到了2027年再提价 如果总的费用仍然低于竞争对手的话 大多数人是不会选择离开的 第二个目的就是夺取 企业的模型路由的主导权 稍微有点规模的企业并 不是只会使用一种模型 而是会进行分层路由 大约有5%的最复杂的最难的 任务会被交由最昂贵且能力 最强的顶级模型来处理 剩下的高频重复难度较低的 日程工作都交给性价比高 速度快、稳定性强的主要模型来完成 95%的任务中 调用量是顶级模型的几十上百倍 这才是人工智能市场的最大基础盘 Gemini3.7Flash打向的是这个市场 不会和顶级模型争夺那5%的高光场景 而是吃掉剩下的95%的高频调用 看上去单价不高 但是架不住使用次数多 细水长流的收入要比依靠顶尖的 人工智能来赚钱更有实际意义 除了模型本身之外 谷歌还拥有其他厂商所不具备的优势 那就是它自己的生态系统 而此次同步推出的新品Spark 全天候代理则是用这个牌 即使是第三方的人工智能做的最好了 也免不了有一个绕不过去的问题 它拿不到你个人上下文的信息 要让其帮助你处理邮件的话 就需要你自己把邮件的 内容导出来再上传上去 让其为你安排日程 需要自己把日程表发送给他 让其修改文档 上传文件等事情也是 一次又一次的进行 十分不便 但是Spark是由谷歌直接 开发出来的代理软件 它可以非常深入的与Gmail Google 日历、Google文档等几乎每一个 职场人使用的软件进行连接 它可以自动读取你的 邮件日程文档等信息 并且不需要你不断的 进行数据同步操作 可以直接帮助你回复邮件 安排会议、修改文档、整理资料等等 这样的用户体验是 第三方人工智能很难做到的 但是权限越高 所承担的风险也就越大 如果人工智能可以随意读取和 修改你的电子邮件文档等信息 那么一次错误的操作就会 导致重要的文件被删除 或者给客户提供错误的内容 所造成的影响要比普通 的AI胡言乱语要大得多 因此谷歌也需要做好三件事情 极高的操作透明度 让使用者知道 人工智能每一步要做什么 重要的操作要人工二次确认 例如发送电子邮件删除 文件之前都要得到你的同意 还有一整套的审计日志 所有的操作都可以被倒退回去 另外它也加入了Antigravity开发环境 把人工智能的操作放在一个可以回滚 受到控制的边界之中 即使人工智能操作出错 或者是遇到网络不稳定 验证码等现实生活中的意外情况 也可以通过一键回滚 功能回到之前的状况 并不会把你的本地环境弄垮 虽然这些细节并不酷炫 但是他们是agent真正的去面对 和解决工作中遇到的各种 问题所必须跨越的一道坎 把目光放远一些的话 目前人工智能界的三大 巨头们所走的道路已经十分明确 并且这三条道路根本就不是一回事 谷歌的Gemini3.7Flash主要卖点为 性价比、工程完成率、工具链绑定 以及全套生态入口 目的是成为整个行业的主要工作模式 并且占领最大的日常使用市场 OpenAi所走的道路是性能上限 与极速路线相结合的方式 它所推出的新一代推理服务Ultrafast 在最高峰的时候可以达到 每秒钟750个输出TOKEN的速度 就是为了消除agent 连续工作时延的问题 以前agent做任务的时候要等 它一步一步的反应出来 卡顿感很强 极速推理可以让整个 过程像真人操作一样流畅 主要针对的是高端场景以及实时交互 例如实时语音对话 实时设备控制、复杂的推理等等 对于延迟和能力上限 有较高要求的场景 而Deepseek走的是差异化的路线 V4Pro版本反而提高了售价 并且开始追求更高的毛利额 之前是以廉价替代品的身份出现的 现在则要向高附加值的产品发展了 不再进行价格战 利用一定的场景来获得收益 三条不同的道路 实际上表明了 人工智能市场的分化趋势 并且不会形成一家独大的局面 高端场景、大众市场、性价比场景等 未来的各个场景都会出现相应的玩家 而竞争也不再是简单的 谁跑分高谁赢的问题了 但是我们也应该清醒地认识到 谷歌此次的战略部署 并不是完全没有风险 在一些方面 如果没有做好 可能会对它实现目标产生影响 第一个问题就是发布的频率过快 每隔三周进行一次大的版本更新 对于普通的用户来说是一件好事 但是对于企业用户而言 兼容性测试以及维护的成本将会很高 企业把主要业务都放在API上 每次模型更新都需要 重新测试所有的功能 害怕新的版本改变了输出格式 或者功能发生变化 从而影响到线上的业务 迭代速度过快会使得保守 企业无法将关键业务流程上线 第二个问题就是跑分与 实际业务之间的差异 公开基准测试是标准的问题 但是企业的实际业务环境千变万化 有的企业有着几十年的老系统 有的企业则存在复杂的内部后端 还有一些非常奇特的边缘场景 跑分高并不意味着在这样的 环境中可以保持稳定的运行状态 在实际的企业用户的使用过程中 是否能够持久的为企业用户提供服务 则要根据实际情况来判断 第三个问题是优惠 期满之后的价格翻倍 目前低价为推广期的价格 在2,027年正式售价时会直接翻倍 到那时很多小微企业和个人 开发者的成本账都要重新算了 到时候竞品推出 性价比更高的产品的话 就会有一部分价格敏感型用户流失 第四问是关于隐私以及越权的风险 agent获得了邮箱、文档 日程等高权限之后 在一次误读误写中所造成的后果 要比普通的AI幻觉要严重得多 普通的幻觉就是说错了某句话 而高权限agent所出现的幻觉 有可能会带来实际的损失 如果这个问题处理不好的话 企业和个人用户都不敢 真的把高权限交给别人 第五个问题是 谷歌的产品线非常分散 目前Gemini有Pro Flash Ultra等版本 并且还推出了Spark代理和Antigravity 开发环境等一系列名称繁多的产品 普通的用户根本无法 分辨出哪些是哪一种产品 自己选择的话就选哪一个吧 产品的线太多 反而会使用户的挑选成本增加 说到这的时候 可能会有人联想到之前谷歌 的核心高管离职创业的事情 在那个时候 很多人都认为谷歌的 人工智能要不行了 核心的人才走了之后 后续的迭代也一定跟不上 结果才过了几个月的时间 Gemini三周发布一次新版本 迭代的速度一点都没有减缓过 反而越来越快了 这件事情实际上也是 给我们的一个警示 目前的人工智能竞争并不是 依靠少数顶尖的研究人员 单打独斗就可以做到的了 现在已经变成了一种 工业化流水线式的竞争 依靠的是算力规模、数据飞轮 自动化评测体系 产品分发渠道组成的一个完整的系统 顶尖的人才固然很重要 但是系统的迭代能力才是 目前大厂的核心竞争力 流水线运转的情况下 个别人员的调动并不会 改变整个系统迭代的速度 对普通人来说 这同样有着非常直接的启示 以后人类的核心价值观并 不是跟AI比谁写代码更快 谁做执行更强 这些问题模型都会做的很快 而且很便宜 人的真正不可替代之处 在于确定正确的目标 控制最后结果 对决策负责 而模型只起到高效的执行作用 倒不如先学会如何利用好这个执行层 来为模型定方向 查结果的人 最后谷歌这次能不能成功 其实不需要看发布会吹的多么好 只需要关注4个主要的数据就可以了 第二是开发者选择的标准 它是否真的会成为大家 生产环境中默认使用的模型 而不会只是用来测试和玩耍 并且编写一些简单的脚本而已 第二步就是对整个 项目的成本进行计算 包括了重试、人工修复等费用之后 该项目真正完成的任务总成本 到底是不是真的比竞争对手低呢 第三是用户的粘性 Spark这样的代理产品能否使 用户形成高频使用的好习惯 而不能只是新鲜感 一下就忘了 第四是生态治理方面 谷歌能否把杂乱无章的 产品线整理的井然有序 并且能够有效的降低用户的 决策难度以及使用成本呢 因此回头来看 谷歌这次又发布了Gemini新版本 真正值得我们关注的是 它多出了多少个参数 并不是哪一个排行榜上超越了谁 更主要的是 谷歌正把Gemini的功能由原来的 回答问题的AI逐渐转变为现在 的可以帮你做事情的工作型AI 而其背后所代表的是 人工智能行业的一个重要变革 以前人们比拼的是模型有多聪明 接下来要拼的是模型是否能完成任务 谁能使得agent更加稳定的使用工具 并且减少错误的发生 也不再反复尝试同样的工作 而又能把每次任务的成本降到最低 那么他就有可能真正的 进入到企业的工作当中去 因此Gemini这次更新 我认为更值得关注的是3个方面 开发人员是否真的会把 它们放到生产环境中去使用 企业在计算了全部的成本之后 是否还会愿意长久的使用 还有谷歌是否能够利用Gmail Docs Calendar等整个生态系统 并且把Gemini变成用户每天 都要用到的应用程序 如果上述的事情最后 都能够运行起来的话 那么谷歌此次不断更新Gemini 并且频繁的进行更新 就不只是普通的版本升级了 在某种程度上来说 它已经预示着人工智能 代理时代到来的可能性 但是现在下结论为时尚早 人工智能行业的变化很快 今天处于领先地位的企业并不 代表明天仍然会保持领先地位 值得我们长期关注的是 这些模型是否可以从发布会 中发布的参数、跑分等信息中 真正走入到普通人的工作流程 以及企业的实际业务当中去 这也是我一直想要做的 在于告诉人们谁发布过什么 并不是说只把新闻 本身讲清楚就可以了 而要把新闻背后所涉及的产业 逻辑也一并拆开来给大家看 如果你想继续学习人工智能 科技巨头以及美股市场的深层次原理 请关注本频道 最后还想问一下你 你觉得Gemini将来最大的优势是什么 是模型本身的实力 还是背后谷歌整个生态系统的力量 欢迎大家在评论区发表自己的观点 本期内容仅为产业 信息分享与观点交流 不构成任何投资建议或交易依据 AI行业变化非常快 任何判断都存在不确定性 请大家独立思考 理性看待市场变化 我们下一期见 Google has released a new version And this time with the previous generation Gemini The update time is almost the same. Even a little abnormal When people see this content Their first reaction was probably Google is about to start the heap version thing again But if you look closely at Gemini, The direction of this update You will find that things are not so simple Since the problem Google wants to solve this time is not How to create a benchmark-first model again It's a more practical question. How does artificial intelligence really work? Get into people's daily work? Yes, developers can happily use it every day. And can the company pay for it for a long time? So Gemini didn't pay all the attention this time Focusing on the question of which model is smarter Instead, it is gradually strengthening code development and tool use And complex task processing, etc Performance of stability and efficiency in agent scenarios That is to say Google wants to replace Gemini with you Occasionally send a few questions to its AI Become a long-lasting The main model that serves you This change seems to be just Upgrade a version of the software But the significance it contains Far greater than the parameter promotion itself Because in artificial intelligence agents really After putting in the work A task is often called Models range from a dozen to dozens of times At this point, people are already comparing Which model is not the most powerful? And which model can be turned into a lower What about cost-stable performance to get the job done? This will affect The future competitive model of the artificial intelligence industry Therefore, the content of today's issue No longer just talk How awesome is the new Gemini version? But from the back of this update Then I took it apart and thought about 3 questions Why did Google suddenly do this? Valuing one doesn't have to be the first But what about models that are really good at doing things? In the context of the agent era Can the price of artificial intelligence be no longer just How much is it per million tokens? In addition, companies such as Google, Openai, and Deepseek How are we going to compete? If you can solve these three problems Then you can find out Google's new model released this time It's not just about innovation. In anticipating artificial intelligence Possibilities for the next phase of maximum use cases Hello everyone, my name is Financial Translator. In this issue, we will start with Gemini this time. The new version is released Take what Google wants to take away Take it apart layer by layer to show you Again, in this update The most irrational part Over the past few years At the artificial intelligence industry conference Everyone is competing to see who can be the best First place in a difficult benchmark Or solve that other models cannot Solved math and reasoning problems It seems that as long as the model is smart enough You can rule the rivers and lakes But this time Gemini3.7Flash did not Want to compete for the title of the strongest brain Google's position for itself is the Workhorse model It's the workhorse model. It's not about winning the hardest problems. Its goal is to solve the problem for developers every day All kinds of real-world scenarios that are frequently used Writing business code Read long files Call various APIs Automatically browse a web to find information Handle various tasks in daily work These things are not cool It doesn't make the tech headlines. But they are developers 80% of the work you do every day Changes in positioning are more important than performance improvements This shows that the entire industry is The standards of assessment have changed People used to compare who had the smartest answers And now it's time to start comparing Who can afford a cheaper price? When you encounter an error in more steps Able to correct itself Re-route Just like when a company is hiring The brightest college graduates Of course the best candidate But it can really bring everyday The work is done quickly and well People who don't always need someone to clean up the mess It is the most indispensable core strength in the team I work with a lot of artificial intelligence Friends of the app have communicated The largest they have encountered before The hard problem is one such question A lot of the model scores look pretty good. But really put it in the production environment All kinds of problems arise Let it write a complete function It can help you leave out the dependent packages. Let it modify the existing project code It can mess up other normal functions Where requirements are not clear It doesn't ask questions. It's all about guesswork. And the output is also unusable. In the end, developers have to spend more The energy to deal with this problem Thereby reducing work efficiency The appearance of Gemini3.7 Flash Obviously designed to deal with the above problems We can take a look at some Open benchmark data It knows what it wants. The direction of improvement is very clear Calculated according to the Frontier Code 1.1 code baseline It scored 34.4% in version 3.6. And at 43.6% Closer to real engineering On the DEPSE 1.1 test conducted in the project From 49% to 65.3% directly WebDevArena Web development-oriented competitions Has risen to 1,588 points Test with the end point command line When TerminalBench 2.1 The result is 85.8%. The most extreme case is Oswald 2.0. That is to test the ability of artificial intelligence to control computer desktops The score has skyrocketed from 3.8% to 47.9% now. The model I used before is on the computer The operation is basically clicking the wrong button. The level of opening the wrong software It can now be done independently. Nearly half of all desktop tasks Such as organizing documents, using software, filling out forms, etc For the agent That's real progress. But Google doesn't hide its shortcomings In the more difficult TerminalBench 3.0 test Google got 14.9 percent. Less than 20.8% of GPT-5.6 terror On extremely complex engineering tasks It is not at the top level of the industry However, for the majority For the daily needs of developers It's powerful enough to use And better understand how to do things well and correctly Targeted strengthening of some Capabilities that are often used in practical development Strictly follow the literal meaning And will not be added at will. Some features that were not requested Unclear requirements The place will take the initiative to ask. I'm not guessing When you find that the path is blocked Or a lot of mistakes It will stop the loss in time. It won't go dark one way Waste a lot of time and computing resources These characteristics cannot be reflected in the running points But when it is actually used The difference in experience will be obvious After talking about the performance Let's talk about the latest updates in this update The core point of agent-age economics That is, the artificial intelligence model cost account The algorithms have changed. When people compare the prices of models Only care how much per hundred TOKEN Use whichever is cheaper It was correct in the previous one-round conversation. When you ask a question The model will answer once And the calculations are very accurate But in the age of agents It doesn't work at all. Agents don't answer one question at a time It's always a series of tasks. The first thing to do is to figure out what the goal is Then mobilize the appropriate tools to achieve this goal Then read the relevant files Write code and test it Once a problem is found, it will be corrected and retried in time. It may be used throughout the process To a dozen or even dozens of models At this point, the main metric for measuring costs is no longer available Is the price per million TOKEN And it becomes a complete All expenses required for the task Some of the models look really cheap. But if you do it incorrectly the first time. You have to try it over and over again. During this period, many unused tools will also be called At the end of the general ledger The cost is actually higher than the price of a single model But as long as you can do it once The model of the thing is much more valuable Let's say you want to hire a decorator. A cheap wage But the work is rough After installing, you have to remove it and reinstall it three times. In the end, it costs more money and time than Find a salary that is slightly higher But you can only use it once. A lot more people get the work done. The focus of this optimization of Gemini3.7Flash is Increase the probability of getting it right the first time Reduce the use of unnecessary tools And pointless repetition It doesn't have to be the cheapest model But as long as the number of retries can be reduced Then complete the same task The total cost will decrease This is the bill that business users really care about After all, when companies use artificial intelligence It's to cut costs. Increase efficiency It's not about buying the lowest-priced token. Analyze according to such cost considerations The price strategy it adopts It is easy to understand To Google's strategic intentions The official selling price is The promotion price is before the end of 2026. Enter TOKEN $0.75 per million Output TOKEN $3.75 per million The official price will be implemented from January 1, 2027 When it rose to $1.50 Output rose to $7.50 That's doubled. From the price point of view It's a bit more expensive than many open source models. However, considering its performance and Considering the task completion rate In fact, its cost performance is very high. Google offers such a price It's not about short-term profits. It has two distinct strategic objectives. The first task is to seize Developer default options The current cost of AI development migration is actually A little bit higher than people think. When developers become accustomed to a certain The way a model answers And put the prompt thesaurus tool chain worker When the workflow is based on the model It will not be easy to replace other models. Once replaced, it needs to be readjusted and tested. The cost is high Google is currently using something close to The price of the cost to promote The purpose is to enable developers to compare Low cost to use in the product After developers get used to it, gradually Bind to the Gemini ecosystem Even if the price increases again in 2027 If the total cost is still lower than the competition Most people are not going to leave. The second objective is to seize The dominance of enterprise model routing A slightly larger company does not Not only one model will be used Instead, it will perform hierarchical routing It's about five percent of the most complex, the most difficult. The task will be entrusted to the most expensive and capable The most powerful top-level model to handle The rest of the high-frequency repetitions are less difficult The schedule work is handed over to cost-effective Fast and stable main model to complete 95% of the tasks The call volume is dozens or hundreds of times that of the top model This is the biggest foundation for the artificial intelligence market Gemini3.7 Flash is targeting this market Won't compete with top models for that 5% highlight scene Instead, eat the remaining 95% of high-frequency calls The unit price doesn't seem to be high. But I can't stand it, and I use it a lot. A steady stream of income is better than relying on the top Artificial intelligence to make money is more practical Except for the model itself Google also has an advantage over other vendors It's its own ecosystem. And the new product Spark launched simultaneously this time All-weather agents use this card Even third-party artificial intelligence does the best There is inevitably a problem that cannot be bypassed It can't get your context. If you want it to help you deal with emails You need to send the email yourself. Export the content and upload it again. Let it organize your schedule for you You need to send him the schedule yourself. Let it modify the document Uploading files and other things Do it again and again Very inconvenient But Spark is directly operated by Google Developed proxy software It can be very deep with Gmail Google. Calendars, Google Docs, and almost every Workplace users use software to connect It can automatically read your Email schedule documents and other information And you don't need to keep Perform data synchronization operations I can directly help you reply to emails. Schedule meetings, modify documents, organize materials, etc This user experience is Third-party artificial intelligence is difficult to do But the higher the authority The greater the risk If artificial intelligence can read and Modify your email documents and other information Then a wrong operation will Causing important files to be deleted Or provide customers with the wrong content The impact is more than ordinary The AI gibberish is much bigger So Google also needs to do three things well Extremely high operational transparency Let users know What to do every step of artificial intelligence Important operations need to be confirmed manually. For example, sending an email to delete Your consent is required before the document There is also a complete set of audit logs All operations can be reversed It also adds Antigravity development environment Put the operation of artificial intelligence in a rollback Within the boundaries of control Even if the artificial intelligence operation goes wrong Or encounter network instability CAPTCHAs and other unexpected situations in real life It can also be rolled back with one click. The function returns to the previous situation It's not going to bring down your local environment. Although these details are not cool But they're real agents to face And solve various problems encountered in the work A hurdle that the problem must cross If you look a little farther The current three major artificial intelligence The path taken by the giants is very clear And they're not the same thing. The main selling points of Google's Gemini3.7 Flash are Cost performance, project completion rate, tool chain binding And a full set of ecological entrances The goal is to become the main mode of operation for the entire industry And capture the largest market for daily use The path taken by OpenAi is the upper limit of performance In combination with the extreme speed route It launched a new generation of inference service, Ultrafast. It can be reached at the peak. 750 output TOKEN speed per second To eliminate the agent The problem of continuous working delay In the past, when agents did tasks, they had to wait It's taking things one step at a time. The card feels very strong Extremely fast reasoning allows the entire The process is as smooth as a human operation Mainly targeting high-end scenarios and real-time interaction Like real-time voice conversations. Real-time device control, complex reasoning, and more For latency and capacity limits Scenarios with higher requirements And Deepseek takes a differentiated route The V4Pro version has actually increased the price. And they began to pursue higher gross margins. It came as a cheap alternative. Now we need to move towards higher value-added products No more price wars Make use of certain scenarios to generate revenue Three different paths In fact, it shows that The differentiation trend of artificial intelligence market And it's not going to be monolithic. High-end scenarios, mass market, cost-effective scenarios, etc There will be corresponding players in each scenario in the future And competition is no longer simple It's a question of who has the highest score and who wins But we should also be soberly aware Google's strategic deployment this time It's not risk-free. In some ways If not done well Could have an impact on how it achieves its goals The first problem is that the frequency of releases is too fast A major version update every three weeks This is a good thing for the average user However, for enterprise users The cost of compatibility testing and maintenance will be high The enterprise focuses its main business on APIs Every model update requires Re-test all functions Afraid that the new version will change the output format Or the function changes. And that's going to impact business online. Iterating too fast can make it conservative Businesses cannot bring critical business processes online The second issue is the relationship between the benchmark and Differences between actual businesses Public benchmarks are a matter of standards However, the actual business environment for enterprises is ever-changing Some companies have decades-old systems Some companies have complex internal backends There are also some very peculiar fringe scenes A high score does not mean that you are in such a situation The environment can maintain a stable operating state In the actual use process of enterprise users Can we provide sustainable services to enterprise users? It should be judged according to the actual situation The third issue is discounts After the expiration date, the price will double The current low price is the price during the promotion period. It will directly double the official price in 2,027 By then, many small businesses and individuals The developer's cost account has to be recalculated At that time, competing products will be launched. More cost-effective products A portion of price-sensitive users churn The fourth question is about privacy and the risk of overreach Agent obtained email, document After high authority such as schedule The consequences of a misreading or writing Much more serious than ordinary AI hallucinations A common hallucination is saying something wrong. And the hallucinations of high-powered agents There may be actual losses If this problem is not handled well Businesses and individual users are hesitant Really delegate high authority to others The fifth question is Google's product line is very fragmented Gemini currently has Pro Flash Ultra and other versions And also launched Spark proxy and Antigravity A series of products with various names such as development environment Ordinary users simply cannot Identify which products are which Choose whichever one you choose. There are too many product lines. Instead, it will increase the user's selection cost Speaking of which Some people may think of Google before The core executives' departure to start a business At that time Many people believe that Google's Artificial intelligence is dying After the core talents leave Subsequent iterations will definitely not keep up It's only been a few months Gemini releases a new version every three weeks It's not slowing down at all. It's getting faster This thing is actually A warning for us The current artificial intelligence competition is not Relying on a few top researchers You can do it alone It has now become a Industrial assembly line competition It relies on computing power scale and data flywheel. Automated evaluation system A complete system composed of product distribution channels Top talent is important But the iterative ability of the system is The current core competitiveness of large factories When the assembly line is running The transfer of individual personnel will not Altering the rate at which the entire system iterates For ordinary people And this has a very direct implication as well. Future human core values and It's not about who writes code faster than AI Who executes better? And these problems the model does very quickly. And it's cheap. The true irreplaceability of human beings It's about setting the right target. Control the final outcome Be responsible for decision-making The model only plays an efficient execution role It's better to learn how to make good use of this execution layer first To orient the model The person who checked the results And will Google do it this time? In fact, you don't need to see how well the press conference blows You only need to focus on 4 main data The second is the criteria for developers to choose Will it really become a reality for everyone? Default model used in the production environment It's not just for testing and playing. And write some simple scripts The second step is to evaluate the entire The cost of the project is calculated After including retries, manual repairs, etc. The total cost of the task actually completed by the project Is it really lower than the competition? The third is user stickiness Can a proxy product like Spark make Users form good habits of high-frequency use It can't just be novelty. I forgot about it. The fourth is ecological governance Can Google put the messy The product line is organized in an orderly manner And it can effectively reduce the user's What about the difficulty of decision-making and the cost of use? So looking back Google has released a new version of Gemini this time What really deserves our attention is How many more parameters does it have? It's not about which leaderboard surpasses whom. More importantly Google is transforming Gemini's functionality from the original The AI that answers questions is gradually transforming into the present A working AI that can help you do things. And what it represents is An important change in the artificial intelligence industry People used to compete about how smart models were. The next thing to fight is whether the model can complete the task Who can make agents use tools more stably? And reduce the occurrence of errors You stop doing the same thing over and over again. And minimize the cost of each task Then he may be truly Enter the work of the company Therefore, Gemini this update I think there are three areas that deserve more attention Will developers really put They are used in the production environment After the company calculates the full cost Will you still be willing to use it for a long time? And will Google be able to leverage Gmail? The entire ecosystem, including Docs Calendar And turn Gemini into a user every day Applications that are required If the above things finally If it can be run Then Google keeps updating Gemini this time. And updated frequently It's not just an ordinary version upgrade. To some extent It already heralds artificial intelligence The possibility of the age of agency But it's too early to tell. The artificial intelligence industry is changing rapidly The leading companies today are not Representatives will still maintain their lead tomorrow It is worth our long-term attention that Are these models available from the press conference? In the parameters, scores and other information published in Really step into the workflow of ordinary people And the actual business of the enterprise This is what I've always wanted to do It's about telling people who posted what. It's not just the news. Just make it clear. And to put the industry behind the news The logic is also taken apart for everyone to see If you want to continue learning artificial intelligence Tech giants and the deep principles of the US stock market Please follow this channel Finally, I would like to ask you What do you think will be Gemini's greatest advantage in the future? It's the power of the model. Or is it the power behind Google's entire ecosystem? Everyone is welcome to express their opinions in the comment section This issue is for industry only Information sharing and exchange of ideas Does not constitute any investment advice or trading basis The AI industry is changing rapidly There is uncertainty in any judgment Please think for yourself A rational view of market changes See you next time.