2026 年 10 月，AI 圈像被打翻了的多米诺骨牌：谷歌端出能「运行世界」的 Genie 4，OpenAI 用 GPT-6.1 Terra 把电脑操作成本打到了人力以下，月之暗面把 3 万亿参数的开源大模型摆上桌，Anthropic 一边递交 IPO 一边推出个人智能体 Claude Duet。这一个月，大模型从「会说话」集体转向「会干活」，开源与闭源、通用与专用、世界模型与智能体三条赛道同时开火。
本文不聊理想主义，只盘十件此刻最该知道的事，每件都给出「是什么、为什么重要、对你意味着什么」。
📦 配套资源：本文为资讯盘点类，无源码包、配置清单或图表源文件，配套资源项为空，跳过开篇资源位。
目录
- 谷歌 Genie 4：从「生成视频」到「运行世界」
- OpenAI GPT-6.1 Terra：电脑操作成本首次低于人力
- Kimi K3.1 Pro：3 万亿参数开源模型登顶开源榜
- Anthropic：递交 IPO 与个人智能体 Claude Duet
- DeepSeek V4.1 Flash 与 MiniMax M3.1：开源侧的双响炮
- 谷歌 Gemini 4 Argon 全球开放
- OpenAI 千智能体 32 分钟完成数月脑细胞分析
- 八部门「十五五」规划：智能体普及率目标超 90%
- 十件大事一览对照表
- 写在后面：这波浪潮意味着什么
1. 谷歌 Genie 4：从「生成视频」到「运行世界」
北京时间 10 月 28 日，谷歌 DeepMind 发布了新一代通用世界模型 Genie 4，官网口径是「迄今最逼真、最具交互性的世界模型」。它最颠覆的一点，是把生成式 AI 从「生成视频」硬生生推到了「运行世界」：仅凭一句提示词，模型就能实时生成一个可交互的动态虚拟世界，场景随你的输入持续调整，人物能自主响应、环境会即时演化。
它到底新在哪？ 最直观的是两点——世界变真实了，而且第一次有了声音。Genie 4 是谷歌首个能原生生成音频的实时世界模型，环境声与人物对话随世界同步产出；同时支持实时物理模拟，人物动作、互动对话、场景修改都能在一句话指令下即时响应。用户不只是「观看」AI 生成的世界，而是能「进入其中」与环境互动。
谷歌 CEO 皮查伊把这一步称为开启「言出法随」的「绿洲」时代。有行业人士感叹，两年前 AI 连静态图像里的手都渲染不好，如今已能即时生成带实时物理与现场音频的可玩 3D 世界。
为什么重要？ 世界模型（world model）一直被视作机器人训练与 agent 进化缺的那块拼图——它提供的不只是画面，而是可控、可交互、带物理规则的训练环境。Genie 4 把「能看」推进到「能听、能交互」，等于给具身 AI 与生成式智能的融合铺了路。谷歌正探索其在游戏设计、影视制作、教育、机器人训练等领域的落地。
要泼的冷水：它目前仍是研究预览，只向受信任测试者开放，普通用户还摸不到；且官方给所有音视频都打了 SynthID 水印。
Genie 4 参考资料：
- 谷歌重磅发布 Genie 4：这才是真正的「世界模型」 -- 新智元
- 谷歌发布 Genie 4 大模型 文本指令可实时生成可交互虚拟环境 -- 第一财经
- DeepMind 发布实时交互世界模型 Genie 4 -- AI工具宝箱
- 谷歌 Genie 4 来了，「一句话生成世界」更近了 -- 腾讯新闻
2. OpenAI GPT-6.1 Terra：电脑操作成本首次低于人力
10 月 26 日，OpenAI 推出 GPT-6.1 Terra，官方把它定位为「速度快、效率高、专为高强度专业工作」的模型，尤其擅长视觉推理、电脑操作（computer use）与清晰自然的写作。它同日在 ChatGPT Work、Codex 与 API 中开放，覆盖 Plus、Pro、Business、Enterprise、Edu 各档。
真正的看点在成本结构。Terra 定价每百万 token 输入 \(1、输出\)5，较上一代直降 50%（价格砍半）；最高强度下单任务成本仅为 Claude Opus 5.5 的九分之一。在 Terminal-Bench 4.0 上，Terra 成功率 50.3%，对比 Claude Opus 5.5 的 21.5%；事实错误率从 53.6% 降到 16.7%。
为什么重要？ 这组数字意味着电脑操作类自动化第一次进入了「按任务算比人力便宜」的区间。过去我们谈 token 成本，现在业内开始谈「单任务成本」——cost per completed task 而非 cost per token，成了路由决策的新标尺。当完成一个任务的成本真的低于一段人工操作，agent 大规模接管电脑操作的经济账才算第一次算得过来。
GPT-6.1 Terra 参考资料：
- 推出 GPT-6.1 Terra -- OpenAI
- Addendum to GPT-6 Astra System Card: GPT-6.1 Terra -- OpenAI
- OpenAI 发布 GPT-6.1 Terra，价格砍半 -- AI工具宝箱
- Release notes -- OpenAI
3. Kimi K3.1 Pro：3 万亿参数开源模型登顶开源榜
10 月 23 日，月之暗面（Moonshot AI）发布 Kimi K3.1 Pro，这是其 K3.1 系列第一款模型。它总参数 2.8 万亿（一说 3 万亿）、激活 1040 亿，是当前已公开的最大开源权重模型；上下文 100 万 token，定价输入 \(2、输出\)10（较前代降 33%），API 与 Kimi.ai 当天即用。
它的能力不止于体量：在 Artificial Analysis Intelligence Index 上拿到 55 分，为开源模型历史最高，位列全体第四；Terminal-Bench 4.0 达到 65.7%。权重计划 11 月 4 日公开。
为什么重要？ 这是开源侧对闭源旗舰的一次正面回应。3 万亿参数却能在 55 分这个档位，还比上一代便宜三分之一，说明中国 AI 厂商在「参数规模 × 成本 × 能力」三者的平衡上确实上了一个台阶。对中文开发者与企业来说，K3.1 Pro 意味着拿到一个可本地部署、能打前沿任务的开源底座。它在游戏开发（Unreal/Unity/Godot）、Blender/CAD 3D 建模等长尾专业任务上表现突出，瞄准的正是通用大模型覆盖不到的垂直深水区。
Kimi K3.1 Pro 参考资料：
- Kimi K3.1 Pro 发布 -- Moonshot AI
- October 2026 AI Model Updates -- Local AI Zone
- AI 官方発表まとめ 先週 -- TECH NOISY
4. Anthropic：递交 IPO 与个人智能体 Claude Duet
Anthropic 是本月的「双料主角」。
一是 IPO。10 月 27 日，Anthropic 递交 IPO 文件，首次全面公开财务：2026 年第二季度营收 116 亿美元，同比暴增 1368%，同期营业利润 9.74 亿美元。这是前沿大模型厂商首次以 IPO 申报材料形式披露完整财务，被视为检验大模型商业模式能否持续盈利的关键样本。这场发行可能让公司估值达到约 1.5 万亿 美元，但亚洲大部分投资者想直接从配售中分份额「几乎不可能」，拿不到份额的资金正在改道买入其已上市股东、加密衍生品与半导体个股。
二是产品。10 月 21 日，Anthropic 发布个人 AI 智能体 Claude Duet，进入个人 Agent 赛道，可主动监控应用并执行会议安排等任务。这已是近两个月内第七、八款形态相近的同类产品——8 月 xAI 推出 Grok Bot、9 月 Meta 发布 Muse 并带动股价上涨 27%。文章认为，个人智能体赛道已进入同质化竞争阶段，厂商更多在交互形态与入口上做文章，差异化能力尚未显现。
为什么重要？ IPO 揭开了大模型公司「到底赚不赚钱」的盖子，而 Claude Duet 说明前沿厂商正集体把战场从「模型能力」转向「个人 Agent 入口」。一个管钱的账本、一个抢入口的产品，两条线都指向同一点：AI 从技术叙事进入商业兑现期。
Anthropic IPO 与 Claude Duet 参考资料：
- Anthropic 递交 IPO 文件 -- PHP中文网
- Anthropic 发布 Claude Duet -- 腾讯新闻
- Anthropic IPO 亚洲配售受阻 -- Bloomberg
- Anthropic Newsroom -- Anthropic
5. DeepSeek V4.1 Flash 与 MiniMax M3.1：开源侧的双响炮
如果说 Kimi 是把开源的天花板抬高，那 DeepSeek 与 MiniMax 这一对，则是把开源侧「单位任务的成本」打到了地板——这十天里，开源国产模型在「低价打高频」这件事上卷到了极致。
DeepSeek V4.1 Flash：10 月 17 日升级到 V4.1-Flash-1017，这是其全新模型结构系列中的最小尺寸成员，带原生多模态视觉理解能力，主打「能力上限更高、推理更快、吞吐更大、可扩展」。官方称这次升级是一次「性能–费用的帕累托改进」：相比上一版，HLE（w/o tools）冲到 42.6、Terminal-Bench 4.0 冲到 42.3、DeepSWE v1.1 到 77.0，而在提升性能的同时还缩短了任务运行时间与输出长度，让人能以更低成本、更快速度跑完同样的活。它在 Artificial Analysis 上拿到智力指数 45，并针对 DeepSeek Harness、腾讯 CodeBuddy、WorkBuddy 等平台做了专门优化。
MiniMax M3.1 Flash：10 月 19 日发布，10 月 23 日开源权重。它最出圈的不是参数，而是一个「马甲」——那个在 OpenCode 上匿名霸榜 12 天的 Space Bunny，就是 M3.1 Flash 的早期版本。它是个 MoE（Mixture-of-Experts，混合专家）模型，总参数 4280 亿、每个 token 只激活 230 亿，100 万上下文，原生吃文本/图像/视频。价格输入每百万 token \(0.15**、输出 **\)0.6、缓存读取 $0.015。
为什么重要？ 这两家把「Agent 任务单价」打到了一个新量级。MiniMax 的定价和 DeepSeek 低峰价一模一样，但 DeepSeek 高峰时段要翻倍到 \(0.3/\)1.2，M3.1 Flash 却是全天不变——所以它不是「更便宜」，是「全天候都便宜」。网友实测最直观：做一个硬件发布落地页，M3.1 Flash 花 2.3 万 token、\(0.014，而 Claude Opus 5.5 花 6 万 token、\)1.22，设计水准却在一个档次，账单差了近 88 倍。
要泼的冷水：便宜不等于全能。在 DeepSWE v1.1 这类硬核工程任务上，M3.1 Flash 是 73.7，DeepSeek 是 77；OSWorld 2.0 上 M3.1 Flash 61.7，也明显落后 Claude Opus 5 的 70.6。它的幻觉率偏高，做物理约束类任务也会翻车。所以更准确的说法是：高频跑量的活交给它，难啃的重构留给更强的模型。
DeepSeek V4.1 Flash 与 MiniMax M3.1 参考资料：
6. 谷歌 Gemini 4 Argon 全球开放
10 月 29 日，谷歌把 Gemini 4 Argon 推向全球：开发者与企业可通过 Gemini API 使用，Google AI 订阅用户能在 Antigravity、Gemini app 与搜索的 AI Mode 里直接用。
它强在哪？ 官方定位是「为最难的问题而生」，擅长编码、金融、法律、网络安全、科学等领域的复杂多步任务。关键是它在 Vals Index 上拿了第一——这个指数专门衡量模型在金融、编码、法律、税务等关键经济领域的表现，Argon 是当前跨领域领先的旗舰。谷歌甚至用它自己完成了 Chrome 跨数百文件的架构重构，把数月的工作压缩到了几周。
定价：入门价输入每百万 token \(2**、输出 **\)10（缓存输入打 95 折）；需要注意的是这只是促销价，11 月 30 日后输入涨到 \(4、输出从 12 月 1 日起涨到\)20。
为什么重要？ 本月谷歌其实有两张牌：前面那张是「运行世界」的 Genie 4，这张是「干深活」的 Argon。一个负责造世界，一个负责处理现实里最繁琐的专业任务，两条线合起来，说明谷歌不再单纯拼「模型会多少」，而是拼「模型能不能替你扛下高价值、高难度的垂直工作」。对开发者来说，Argon 是全球通用（而非仅美国）的旗舰入口。
Gemini 4 Argon 参考资料：
- Gemini 4 Argon 正式向所有人开放 -- Google
- Gemini 4 Argon: Built to solve your hardest problems -- The Keyword
- Gemini 4 Argon 的新功能 -- Google AI for Developers
- Gemini Developer API 定价 -- Google AI for Developers
7. OpenAI 千智能体 32 分钟完成数月脑细胞分析
10 月，OpenAI 展示了「agent 规模化」最有冲击力的一幕：它和 PerturbAI、Scripps Research 一起，用数千个 AI agent（由 GPT-Rosalind 驱动）并行工作，在 32 分钟里，从数百万个小鼠脑细胞的数据中，识别出值得进一步研究的生物学线索。
它到底干了啥？ 传统上，这类大规模实验（比如基因扰动实验）产生的数据远超科学家手动能分析的上限。OpenAI 让几千个 agent 并行地去「读实验数据 + 检索科学文献」，把基因扰动引起的效应与文献中的既有发现关联起来，从而发现一些研究者原本没注意到的、意想不到的生物学联系。这些联系能帮科学家理解疾病、找到新的治疗研究方向。
为什么重要？ 这一仗打的是「科学发现能不能被 agent 规模化」。单看 32 分钟这个数，它把过去可能需要数月、靠组内一小撮人反复推导的工作，压缩成了「几千个 agent 并行跑一段代码」。更关键的是，它不是「让 AI 胡说八道」，而是让 agent 去「接住」那些被实验室数据中淹没、靠人眼已看不到的线索。当 AI 开始替科学家决定「下一个值得做的实验是什么」，agent 就不再只是工具，而成了科研流程里的一环。
OpenAI 千智能体脑细胞分析参考资料：
8. 八部门「十五五」规划：智能体普及率目标超 90%
如果说前面几件是厂商侧的「点」，那这一件就是国家侧的「面」。10 月底，中央网络安全和信息化委员会印发 《国家信息化「十五五」规划》，为未来五年的信息化画了一张任务书、时间表、路线图。
最值得留意的一个数：在《规划》列出的 12 项主要指标里，提升幅度最大的是「新一代智能终端、智能体等应用普及率」——从 2025 年的 40% 提到 2030 年的 90% 以上，五年间翻一番以上。换句话说，国家层面已经把「智能体」摆到了和 5G、IPv6、数字经济核心产业同一个量级的位置上，并给出了明确的量化目标。
《规划》的其他信号也很有指向性：互联网普及率从 80.1% 提到 85%，每万人 5G（含 5G-A）基站从 34 个提到 50 个，数字经济核心产业增加值占 GDP 比重从 10.5% 提到 12.5%，并「适时启动第六代移动通信（6G）商用」。它还明确提出加快量子科技、脑机接口、具身智能、6G 等技术产业化应用。
为什么重要？ 这是智能体第一次以「国家普及率指标」的形式出现，意味着它不再只是技术圈的热词，而进入了产业政策的硬约束。对从业者而言，这是一个明确的「确定性信号」：往智能体方向投入，未来五年有明确的政策托底；对企业而言，「智能体普及率 90%」不是一个口号，而会实打实地转化为对各行业智能化改造的考核与扶持。
八部门「十五五」规划参考资料：
9. 十件大事一览对照表
把这一月的几条主线塞进一张表，你会更清楚地看到它们在往哪走——从「会说话」，到「会干活」，再到「进场干活」。
三句话读懂这波浪潮：一是开源与闭源同时提速，开源在参数规模与单位成本两端同时逼近闭源；二是世界模型与智能体并进，一个造「能玩的世界」，一个抢「能干的入口」；三是国家与资本同时入场，智能体第一次进入政策指标，IPO 揭开了商业兑现的盖子。
10. 写在后面：这波浪潮意味着什么
回头看这一个月，AI 圈的节奏已经从「每个月出一个更强的模型」变成了「每个月同时开几条新赛道」。世界模型、开源万亿参数、个人智能体、Agent 规模化、国家政策——它们不是孤立的事件，而是在同一个拐点上结出的果。
什么时候该关注这件事：如果你做 agent、做具身智能、做企业智能化改造，这个月是「该认真看」的信号月——无论从成本（Terra、M3.1）、能力（Genie 4、K3.1 Pro）还是政策（「十五五」规划）看，往智能体方向投入都有了更明确的依据。
什么时候不该慌：如果你只是围观，那这些发布大多还是「研究预览」或「入口争夺」，离大规模落地仍有距离。别被「单任务成本九分之一」「32 分钟替代数月」这类数字冲昏头——它们描述的是特定场景下的最优解，不代表通用替代已经发生。
本专栏《LLM 大模型技术专栏》从 Transformer 注意力机制一路讲到 LangChain、RAG 与 MoE 落地，每篇都配一段可直接运行的代码，每周更新一篇，聚焦真实可用的模型与 agent 实践。想按顺序读完，可在专栏页订阅。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
In October 2026, the AI circle was smashed to the domino placard: Google brought out Genie 4, Openai, who was able to "run the world" with GPT-6.1 Terra to lower the operating costs of computers, and the dark side of the month put the opening source LLM of 3 trillion parameters at the table and Anthropic presented the IPO while launching a personal AI agent Claude Duet. This month, LLM moved from "talking" to "working" with open-source and closed-source, general and exclusive, world models and AI agent tracks.
Instead of talking about idealism, this is just about 10 things that should be known at this moment, each of them says, "What is it, why is it important, what it means to you?"
Complementary resources: This is an information inventory category, an open-source package, a configuration list or a chart source file, a resource package that is empty and skips the opening resource slot.
Contents
Google Genie 4: From "Generating Video" to "Run the World"
- Open AI GPT-6.1 Terra: Computer operation costs are lower than manpower for the first time
- Kimi K3.1 Pro: 3 trillion argument open source model to top open source list
- Anthropic: Hand over IPO and individual AI agent Claude Duet
- DeepSeek V4.1 Flash and Mini Max M3.1: The Double Sound of Open Source Gun!
Google Gemini 4 Argon Global Open
- Open AI thousands of AI agent 32 minutes to brain cell analysis for months
- Eight Sector 15: AI agent coverage is over 90%.
- Ten major events.
- In the back: What does this wave mean?
Google Genee 4: From "Generating Video" to "Run the World"
On October 28th, Beijing time, Google DeepMind publish developed a new generation of universal world models, Genie 4, the most realistic and interactive model of the world to date. One of its most subversive things is that it pushes the generation of AI from "generating video" to "run the world": with a single prompt, models can produce an interactive, dynamic virtual world in real time, adjusted as you enter it, and people can respond on their own.
What's new is two things -- the world is real, and for the first time there's a voice. Genee 4 is the first real-time world model of Google to generate indigenous audio, with environmental sound and character conversations synchronized with world output; and supports real-time physical simulations, with people acting, interactive dialogues, and scene modification responding immediately under one sentence. Users are not just "watch" the world created by AI, but are able to "enter" the environment.
Google CEO Pichai calls this step the beginning of the "Oasis of Words". The industry lamented that, two years ago, AI had a bad hand rendering in static images and that it was now able to produce a 3D world with real-time physics and live audio.
The world model has been seen as a puzzle of robotic training and the evolution of an individual -- It provides not just a picture, but a controlled, interactive and physical training environment. Genee 4 moves "seeable" to "hearable, interactive" and paves the way for a combination of physical AI and generating intelligence. Google is exploring its landing in areas such as game design, video production, education, robotics training.
Cold water to be poured: it is still a research preview, open only to trusted testers, not yet touched by ordinary users; and all audio and video recordings are officially synthID watermarked.
Genie 4 References:
Google weighs publish Genie 4: That's the real "world model" -- the new brain.
- Google publish Genie 4 LLM text commands can generate interactive virtual environments in real time -- first finance
- DeepMind publish real-time interactive world model Genie 4 -- AI Toolbox
- Google Genie 4 is here.
2. Open AI GPT-6.1 Terra: Computer operation costs are lower than manpower for the first time
On October 26th, Open AI launched GPT-6.1 Terra, which is officially a model of "quick, efficient, and specialized in high-intensity professional work", especially in visual reasoning, computer operations and clear nature writing. It was opened the same day in ChatGPTWork, Codex and API, covering the Plus, Pro, Business, Enterprise, Edu.
The real focus is on the cost structure. Terra price per million token input \ (1, output \), which is 50% lower than the previous generation (50% price cut); the maximum strength of the individual task cost is only one ninth of Claude Opus 5.5. On Terminal-Bench 4.0, Terra success rate 50.3% compared with 21.5% of Claude Opus 5.5; the factual error rate dropped from 53.6% to 16.7%.
Why does it matter? This group of figures means that for the first time the computer operation type automation has entered a "mission-by-mission cheaper than manpower". We used to talk about token costs, and now business started talking about "single task costs" -- cost per completed task instead of cost per token, which became the new ruler for route decision-making. When the cost of completing a mission is really less than a manual operation, angent's economics to take over computer operations on a large scale will be calculated for the first time.
GPT-6.1 Terra Reference:
- Launch GPT-61 Terra -- Open AI
- Addendum to GPT-6 Astra System Card: GPT-6.1 Terra -- OpenAI
- Open AI publish GPT-61 Terra, half price -- AI Toolbox
- Release notes -- OpenAI
Kimi K3.1 Pro: 3 trillion argument open source model to top open source list
On 23 October, Moonshot AI publish Kimi K3.1 Pro, the first model of its K3.1 series. Its total parameter is 2.8 trillion (three trillion dollars), activated by 104 billion, the largest open source weight model currently available; context 1 million token, priced input \((2, output\)10 (33% lower than previous generations), API and Kimi.ai are used on the day.
Its capacity is beyond volume: 55 points on Artificial Analysis Information Index, ranked fourth in the history of open source models; Terminal-Bench 4.0 reached 65.7 per cent. The weighting plan was announced on November 4th.
Why does it matter? This is a positive response to the closed source flag. The 3 trillion-billion parameter, which is one third cheaper than the previous generation, shows that China ' s AI manufacturer has actually taken a step in the balance of "parameter size x cost x capacity." For Chinese developers and businesses, K3.1 Pro means getting an open source base that can be locally deployed and can play forward tasks. It has been prominent in long-tailed professional tasks such as game development (Unreal/Unity/Godot), Blender/CAD 3D modelling, aimed at vertical depths not covered by General LLM.
Kimi K3.1 Pro References:
Kimi K3.1 Pro publish-Moonshot AI
- October 2026 AI Model Updates -- Local AI Zone
- AI official escape watch week--
4. Anthropic: Submission of IPO and individual AI agent Claude Duet
Anthropic is this month's "two-factor."
The first was IPO. On October 27, Anthropic presented the IPO document for the first full disclosure of its finances: in the second quarter of 2026 it received $11.6 billion, an increase of 1368 per cent over the same period, and operating profits of $974 million over the same period. This was the first time that the front-line LLM manufacturer had disclosed a full financial disclosure in the form of an IPO declaration and was seen as a key test of the sustainability of the LLM business model. This release may have brought the company’s valuation to about $1.5 trillion, but most investors in Asia are re-routing their listed shareholders, encrypted derivatives, and semiconductor shares to “almost impossible” direct share of the distribution.
Two is the product. On 21 October, the individual Anthropic publish AI AI agent Claude Duet entered the personal Agent track to proactively monitor applications and carry out assignments, etc. This is a similar product in seven and eight sections in almost two months - in August, XAI launched Grok Bot, in September, Meta publish Muse, and pushed stock prices up 27%. According to the article, the individual AI agent track has entered the phase of homogenization competition, with more producers writing on interactive forms and entry points, and the ability to differentiate has not yet been demonstrated.
Why does it matter? IPO has unplugged LLM’s cap, and Claude Duet suggests that front companies are collectively shifting the battlefield from “model capabilities” to “individual Agent entrances”. A money book, a product that robs, both lines point to the same point: AI goes from a technical narrative to a commercial payoff.
Anthropic IPO and Claude Duet references:
- Anthropic Submit IPO File - PHP Chinese Network
- Anthropic publish Claude Duet-Tencent News
- Anthropic IPO Asian distribution blocked - Bloomberg
- Anthropic Newsroom -- Anthropic
5. DeepSeek V4.1 Flash and Mini Max M3.1: The Double Sound on Open Source Gun!
If Kimi is raising the open-source ceiling, DeepSeek and MiniMax are going to hit the open-source side of the "unit cost" to the floor. I'm sorry. In the last 10 days, the open source country model has been rolled to the point of "low-cost high-frequency."
DeepSeek V4.1 Flash: Upgraded to V4.1-Flash-1017 on October 17, which is the smallest member in its new model structure series, with original multi-modular visual understanding, leading to a "high upper limit of competence, faster reasoning, greater insulation, more scalable". Officially, this upgrade is a "performance-cost Paretto improvement": compared to the previous edition, HLE (w/o costs) burst to 42.6, Terminal-Bench 4.0 to 42.3, DeepSWE v. 1.1 to 77.0, while enhancing performance by reducing mission running time and output length, allowing people to run the same job at a much lower and faster rate. It obtained an intellectual index 45 on Artificial Analysis and made specific optimizations for platforms such as DeepSeek Harness, Tencent CodeBuddy and WorldBuddy.
Mini Max M3.1 Flash: 19 October publish, 23 October open source weight. It's not about parameters, it's about a "salary" -- the 12-day Space Bunny on OpenCode, the early version of M3.1 Flash. It is a model of MoE (Mixture-of-Experts, Mixture-of-Experts, Mixture-Experts) with a total parameter of 4280 billion, each token only activated 2.3 billion, 1 million context, raw text/image/video. Price input per million token \(0.15**, output **\) 0.6, cache reads 0.015.
Why does it matter? These two guys took the "Agent task unit price" to a new scale. Mini Max's pricing is exactly the same as DeepSeek's low price, but DeepSeek's peak time doubles to 1.2, and M3.1 Flash remains the same - so it's not "better" and "better" all day. Netizens measure the most intuitive: a hardware publish drop page, M3.1 Flash spends 23,000 token, \ (0.014, while Claude Opus 5.5 spends 60,000 token, \) 1.22, while the design level is in one slot and the bill is nearly 88 times lower.
Cold water to throw: Cheap isn't all-powerful. M3.1 Flash is 73.7, DeepSeek is 77; and M3.1 Flash 61.7 on OSWorld 2.0, also clearly behind the 70.6 of Claude Opus 5. It's hallucinating, and it flips when it's a physical binding type. So more precisely, the high-frequency running effort is left to it, and the hard-to-eat re-engineering is left to the stronger model.
DeepSeek V4.1 Flash and Mini Max M3.1
6. Google Gemini 4 Argon Global Open
On October 29, Google pushed Gemini 4 Argon to the global level: developers and businesses can use it through Gemini API, and Google AI subscribers can use it directly in Antigravity, Gemini app and the search for AI Mode.
Where is it? Official positioning is the "problem for the most difficult problems" with complex multi-step tasks in the fields of coding, finance, law, cybersecurity, science, etc. The key is that it takes first on Vals Index -- an index that specifically measures the performance of the model in key economic areas such as finance, coding, law, tax, etc., and Argon is currently the leading flagship across disciplines. Google even used it itself to reconstruct the Chrome architecture across hundreds of documents, compressing months of work to weeks.
Pricing: Enter entry price per million token \ (2**, output **) 10 (95 discount for cache input); it should be noted that this is only a promotion price, which increases to \ (4, output from 1 December) 20 after 30 November.
Why does it matter? Google this month actually has two cards: the one in front of it is Genie 4 who runs the world, and the one who works deep. The combination of two lines, one to build the world, one to deal with the most cumbersome professional task in reality, suggests that Google will no longer simply spell "how many models will be" but "can models carry high-value, difficult vertical jobs for you?" For developers, Argon is a global (not just the United States) flagship entry point.
Gemini 4 Argon reference:
- Gemini 4 Argon is officially open to all -- Google
- Gemini 4 Argon: Built to solve your hardest problems -- The Keyword
- Gemini 4 Argon 's new feature -- Google AI for Developments
Google AI for Developments
7. Open AI thousands of AI agent 32 minutes for brain cell analysis
In October, Open AI showcased the most powerful scene of "agent size": it worked in parallel with Perturbai, Scripps Research, with thousands of AI angents (driven by GPT-Rosalind) and in 32 minutes identified biological clues worth further study from data on millions of mice' brain cells.
Traditionally, the data generated by such large-scale experiments, such as genetic disturbance experiments, are far above the upper limit of manual analysis by scientists. Open AI allowed thousands of researchers to "read experimental data + retrieve scientific literature" in parallel, linking the effects of genetic disturbances to what is already found in the literature, and thus discovering unexpected biological links that some researchers had not noticed. These links help scientists understand the disease and find new treatment directions.
Why does it matter? The battle is "can scientific discovery be scaled up?" The 32 minutes alone, which has reduced the work that may have taken months to the past, which was repeatedly extrapolated by a small group of people, to "thousands of angents running a piece of the code." More importantly, it's not "let's let 'AI bullshit" but let's get angent to "catch" the clues that have been flooded with laboratory data and that are no longer visible. When AI started deciding for scientists what the next worthwhile experiment was, angent was no longer just a tool, but a part of the scientific process.
Open AI thousands of AI agent brain cell analysis references:
8. Eight-sector "Fifty-five" programme: AI agent coverage target exceeded 90%
If the previous items are "points" on the side of the factory, then this is "face" on the side of the country. At the end of October, the Central Committee for Cybersecurity and Informatization published the National Informatization " Fifteen Fives " Plan, which provides a mission, timetable and road map for the next five years.
The most noteworthy number: of the 12 major indicators listed in the Plan, the most significant increase was the "new generation smart terminals, AI agent applications, etc." From 40% in 2025, over 90% in 2030, more than doubled in five years. In other words, “AI agent” has been placed at the national level at the same level as 5G, IPv6, the core industry of the digital economy, with clear quantization targets.
Other signals of the Plan are also indicative: Internet penetration is mentioned from 80.1% to 85%, 5G (including 5G-A) base stations per 10,000 people is mentioned from 34 to 50, the value added of the core industry in the digital economy from 10.5% to 12.5% of GDP, and "starting sixth generation mobile communications (6G) commercials in due course". It also explicitly proposes accelerated industrial applications of quantum technology, brain interfaces, smartness, 6G etc.
This is the first time that AI agent has appeared in the form of a "national penetration indicator" that means that it is no longer just a technical community hot word, but that it has entered the hard grip of industrial policy. For practitioners, this is a clear "determinate signal": investing in AI agent, with a clear policy base for the next five years; for enterprises, 90 per cent of Z42Z penetration is not a slogan, but it translates into testing and support for the intellectual transformation of industries.
In the case of the "Fifty-five" project reference service:
9. Checklist of 10 major events
Put the main lines of this month into a table, and you will see where they're going -- from "talking" to "working" to "working."
Three sentences read the wave: first, open-source and closed-source, with open-source sources closing at both the size of the parameters and unit cost; second, the world model coming in hand-in-hand with AI agent, a "playable world" and a "powerful entrance"; and third, the country entering with capital, with AI agent entering the policy target for the first time, and the IPO opening up the commercial cover.
10. Write back: What does this wave mean?
Looking back at this month, the pace of the AI circle has changed from "a stronger model every month" to "a few new tracks each month." World models, open-source hundreds of billions of parameters, individual AI agent, Agent size, national policies — they are not isolated events, they are the result of the same nodes.
It is time to focus on this: if you do angent, do smart, do smart business, this month is the signal month of "need to look carefully" — whether in terms of costs (Terra, M3.1), abilities (Genie 4, K3.1 Pro) or policies (the "Fifty Five" plan) there is a clearer basis for investing in AI agent.
If you're just looking around, then most of these publishs are "research previews" or "entry competitions" still far from mass landings. Don't let the number of "one ninth of the cost of a single mission" "32 Minutes Instead of Months" go down in the head. They describe the best solution in a given scenario and do not mean that a generic replacement has occurred.
This column, " LLM Z43Z Technology column ", from Transformer ' s Attention to LangCain, RAG and MoE landed, each with a direct running code, updated weekly, focusing on real available models and angent practices. To read in order, you can subscribe to the column page.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.