第 041 期
2026 年 10 月 22 日 · …
Claude Duet：Anthropic 的主动型 agent：独立 app，基于 Opus 5.5，会主动找你、离开后继续干活、代你行动前先确认；先向美国 Pro/Max 用户开放 waitlist，和现有套餐共享额度。PHP2Shell：AI 找出影响约 70% 网站的漏洞：借助 OpenAI 模型的研究者和 Wiz 先后发现 PHP 引擎两个可远程执行代码的漏洞，PHP 5.4 起的版本都受影响；同一天还有 LiteLLM 投毒、Railway 环境变量泄露。OpenAI 模型和 Codex 进 Google Cloud：前沿模型和 Codex agent 工作流将进入 Google Cloud Model Garden，可用 GCP 承诺消费抵扣，之后还会进 Antigravity 和 Gemini Agent。
Claude DuetOpus 5.5PHPCodexLiteLLMRailwayGitLabNode.jsGoogle CloudModel GardenAntigravityDotsGPT-RosalindGrok BotPagesFlintPiOpenCodePDD KitChopinVS CodeGitHub CopilotGemini 4 ArgonT3 CodeVercel Token PassDeepSeek-V4.1-FlashMuse Wonder
70%
PHP 支撑的网站占比（估计），PHP2Shell 影响 PHP 5.4 以来的版本
6,000+
Claude Duet 可用的 Claude Marketplace 连接器与插件数
2.5B
Recursive Flint 每个 token 的激活参数（总参数 39.8B）
Issue 041
Oct 22, 2026 · …
Claude Duet: Anthropic's proactive agent: A standalone app built around Opus 5.5 that comes to you first, keeps working when you step away, and asks before acting on your behalf; waitlist for US Pro/Max users, sharing existing plan limits.PHP2Shell: AI finds a flaw affecting ~70% of websites: Researchers using OpenAI models, and Wiz independently, found two remote-code-execution bugs in the PHP engine affecting every version since 5.4; the same day brought a poisoned LiteLLM release and a Railway environment-variable breach.OpenAI models and Codex on Google Cloud: Frontier models and Codex agentic workflows are coming to Google Cloud Model Garden, payable from GCP spend commitments, and later to Antigravity and the Gemini Agent.
Claude DuetOpus 5.5PHPCodexLiteLLMRailwayGitLabNode.jsGoogle CloudModel GardenAntigravityDotsGPT-RosalindGrok BotPagesFlintPiOpenCodePDD KitChopinVS CodeGitHub CopilotGemini 4 ArgonT3 CodeVercel Token PassDeepSeek-V4.1-FlashMuse Wonder
70%
Estimated share of websites running on PHP; PHP2Shell affects every version since PHP 5.4
6,000+
Claude Marketplace connectors and plugins available to Claude Duet
2.5B
Active parameters per token in Recursive Flint (39.8B total)
第 040 期
2026 年 10 月 21 日 · …
DeepSeek 开源 DeepInfer：支撑 DeepSeek-V4.1-Flash 线上服务的推理框架以 Apache-2.0 开源，支持 NVIDIA SM100 和华为昇腾 950；KV cache 系统、PD 分离和负载均衡不在这次开源范围内。Codex：约 4 万账户额度算错：OpenAI 承认约 40,000 个账户实际套用的用量上限和所选套餐不一致，已修复并会补额度；同一天又给所有人重置了一次额度。tsrs：TypeScript 编译器的 Rust 移植：Cloudflare 和 VoidZero 联合推出 tsrs，与 TypeScript 团队合作而非分叉，同等配置下比 TS 7.0 快最多 8 倍、内存少约 70%，可跑在 WASM 里。
40,000 Codex 用量上限与套餐不一致、将获得补偿的账户数（约数）
DeepInferDeepSeek-V4.1-FlashCodexGPT-6 AstratsrsTypeScriptOxcOpenAI APIChatGPTRosalindClaudeClaude CodeOpus 5.5Grok Botdurable-harnessWorkflow SDKReason NetworkOpenClawSolid 2.0OpenCodeT3 CodeArtificial Analysis Intelligence Index
Issue 040
Oct 21, 2026 · …
DeepSeek open-sources DeepInfer: The inference framework behind DeepSeek-V4.1-Flash's production service is out under Apache-2.0, supporting NVIDIA SM100 and Huawei Ascend 950; the KV cache system, prefill/decode disaggregation and load balancing are not part of this release.Codex: limits wrong on ~40,000 accounts: OpenAI says about 40,000 accounts had usage limits that didn't match their selected plan; it's fixed and credits are coming. The same day, everyone got another limit reset.tsrs: a Rust port of the TypeScript compiler: Cloudflare and VoidZero launched tsrs, built with the TypeScript team rather than as a fork: up to 8x faster than TS 7.0 with ~70% less memory at the same worker count, and it runs in WASM.
40,000 Codex accounts (approx.) whose usage limits didn't match their plan and will be credited
DeepInferDeepSeek-V4.1-FlashCodexGPT-6 AstratsrsTypeScriptOxcOpenAI APIChatGPTRosalindClaudeClaude CodeOpus 5.5Grok Botdurable-harnessWorkflow SDKReason NetworkOpenClawSolid 2.0OpenCodeT3 CodeArtificial Analysis Intelligence Index
第 039 期
2026 年 10 月 20 日 · …
Space Bunny 原来是 MiniMax M3.1 Flash：跑了 200T+ token 的匿名模型揭晓身份：API 每百万 token 输入 $0.15、输出 $0.60，权重 10 月 23 日放出；Qoder、Cline、Vercel AI Gateway 首日接入，后两家限时免费。美国国土安全部：联邦民用机构移除 DeepSeek：DHS 依据 2018 年供应链安全法下令，联邦民用机构停止采购并移除 DeepSeek 产品，连「基于 DeepSeek 实现开发的第三方产品」也在范围内。OpenAI：低资源编程语言大幅进步：MultiLangBench 上 GPT-6 Astra 在 20 种语言上平均通过率 87.4%（GPT-5.6 Sol 为 26.8%）；COBOL 迁移评测 GPT-6.1 Sol 低推理档 86.6%。
$0.15 MiniMax M3.1 Flash 每百万输入 token 的 API 价格
MiniMax M3.1 FlashQoderClineVercel AI GatewayDeepSeek-V4.1-Flash-1017CodexGPT-6 AstraGPT-6.1 SolMultiLangBenchDaybreak BlueGrok BotDevin FusionDevin CoachClaude CodeLiteLLMT3 CodeCua Spaceslibfxfx0Tessl VerifiersNotion AIExa 3.0Gemini 4 ArgonWabi 2.0
Issue 039
Oct 20, 2026 · …
Space Bunny was MiniMax M3.1 Flash: The anonymous model that processed 200T+ tokens is revealed: API at $0.15 input / $0.60 output per 1M tokens, weights on Oct 23; Qoder, Cline and Vercel AI Gateway ship day-0 support, the latter two free for a limited time.DHS: federal civilian agencies must remove DeepSeek: Under a 2018 supply-chain security law, DHS ordered civilian agencies to stop buying and remove DeepSeek products, including third-party products developed using DeepSeek implementations.OpenAI: big gains in low-resource languages: On MultiLangBench, GPT-6 Astra averages an 87.4% test pass rate across 20 languages (GPT-5.6 Sol: 26.8%); on COBOL migration, GPT-6.1 Sol hits 86.6% at low reasoning effort.
$0.15 MiniMax M3.1 Flash API price per 1M input tokens
MiniMax M3.1 FlashQoderClineVercel AI GatewayDeepSeek-V4.1-Flash-1017CodexGPT-6 AstraGPT-6.1 SolMultiLangBenchDaybreak BlueGrok BotDevin FusionDevin CoachClaude CodeLiteLLMT3 CodeCua Spaceslibfxfx0Tessl VerifiersNotion AIExa 3.0Gemini 4 ArgonWabi 2.0
第 038 期
2026 年 10 月 19 日 · …
OpenCode 被连续攻击 48 小时：@thdxr 说攻击一直没停，防御用的前沿模型中途开始拒答，拿到 OpenAI 的 Daybreak Blue 后才好转；@paulg 的评价是「可预见的未来就是这样」。Anthropic Public Benefit Fund：5% 股份起步：估值每比基线高 50% 可再加至多 2%，连同联创和员工的捐赠，合计约 20% 股份用于公共利益；马斯克回复建议直接把股份给公众。DeepSeek 新 Flash 第三方成绩：便宜 19 倍：Browser Use 基准上 V4.1-Flash-1017 拿到 75/100，与 Opus 5.5 持平、单任务成本便宜 19 倍；Code Arena WebDev 榜上成为开源第一。
48 小时 OpenCode 遭到连续攻击的时长（@thdxr）
OpenCodeDaybreak BlueAnthropic Public Benefit FundClaudeGrok BotCodexdotChatGPT SitesDeepSeek-V4.1-Flash-1017T3 CodeADEMarkpstackaurora-alphavLLMQwen3.8Vercel AI Gateway
Issue 038
Oct 19, 2026 · …
OpenCode under attack for 48 hours: @thdxr says the attacks haven't let up; the frontier models used for defense started refusing mid-incident until OpenCode got access to OpenAI's Daybreak Blue. @paulg: "This is how things will be for the foreseeable future."Anthropic Public Benefit Fund starts with 5% of shares: Up to 2% more for every 50% rise in value over a baseline; with founder and employee pledges, roughly 20% of shares go to the public good. Elon Musk replied that the shares should go directly to the public.DeepSeek's new Flash, third-party results: 19x cheaper: V4.1-Flash-1017 scores 75/100 on the Browser Use benchmark, tying Opus 5.5 at 19x lower cost per task, and is now the #1 open model on Code Arena WebDev.
48 hours How long OpenCode has been under continuous attack (per @thdxr)
OpenCodeDaybreak BlueAnthropic Public Benefit FundClaudeGrok BotCodexdotChatGPT SitesDeepSeek-V4.1-Flash-1017T3 CodeADEMarkpstackaurora-alphavLLMQwen3.8Vercel AI Gateway
第 037 期
2026 年 10 月 18 日 · …
DeepSeek-V4.1-Flash-1017 开源：官方称 DeepSWE 77.0：MIT 协议放出权重，输出更短、完成更快；官方自测 DeepSWE、Automation-Bench 超过 Opus 5；Search / Fetch API 全面开放，每千次 0.75 美元；vLLM、Ollama、Cline、Qoder 当天接入。OpenCode 安全事件：1.4% Console 用户受影响：OpenCode 称与「大规模蒸馏攻击」有关，受影响的是绑定了外部模型提供方的用户；Anthropic 已吊销疑似泄露的 Claude key，Google 封禁了暴露的 Gemini key。OpenAI：Codex 额度再重置，Replit 默认换成 OpenAI：@thsottiaux 的「Day 13」再次重置 Codex 用量额度，1.3 万赞；TechCrunch 报道 Replit 把默认模型从 Anthropic 换成 OpenAI，Claude 仍可选。
77.0 DeepSeek-V4.1-Flash-1017 的 DeepSWE v1.1 得分（官方自测）
DeepSeek-V4.1-Flash-1017DeepSeek HarnessOpenCodeCodexChatGPTdotReplitClaudeOpus 5.5Claude Codetsc-tsT3 CodeCursorpstackGrok BotGemini 4 ArgonGitHub Copilotpi-durable
Issue 037
Oct 18, 2026 · …
DeepSeek-V4.1-Flash-1017 goes open: 77.0 on DeepSWE, per DeepSeek: Weights under MIT, shorter outputs and faster task completion; DeepSeek's own benchmarks put it above Opus 5 on DeepSWE and Automation-Bench; Search / Fetch APIs open to all at $0.75 per 1,000 calls; vLLM, Ollama, Cline and Qoder support it on day one.OpenCode security incident: 1.4% of Console users affected: OpenCode ties it to a "large scale distillation attack" hitting users who had linked external model providers; Anthropic revoked Claude keys believed affected, and Google blocked exposed Gemini keys.OpenAI: another Codex limit reset, Replit switches its default to OpenAI: @thsottiaux's "Day 13" reset Codex usage limits again, drawing 13k likes; TechCrunch reports Replit has made OpenAI its default model provider over Anthropic, with Claude still selectable.
77.0 DeepSeek-V4.1-Flash-1017 on DeepSWE v1.1 (DeepSeek's own benchmark)
DeepSeek-V4.1-Flash-1017DeepSeek HarnessOpenCodeCodexChatGPTdotReplitClaudeOpus 5.5Claude Codetsc-tsT3 CodeCursorpstackGrok BotGemini 4 ArgonGitHub Copilotpi-durable
第 036 期
2026 年 10 月 17 日 · …
Reflection 开放 Beam 权重：501B / 23B 激活，此前化名 cat-jam：Apache 2.0、1M 上下文，连训练用的 harness Mirror CLI 一起开源；Vercel 透露它此前以 stealth 模型 cat-jam 身份上线；Ollama 限时免费，可直接接 Claude Code、Codex、OpenCode。Grok Bot 一轮「提速 + 打磨」：电脑任务平均快 3 倍、其他回复耗时不到一半，卡片重新设计；新增 Robinhood、Coinbase 等连接器，团队可向 marketplace 投稿；官方 bot 模板一帖收藏 3.2k。OpenAI：GPT-6 Astra Pro 支持多 agent 并行：ChatGPT 里的 Astra Pro 更快、更会用插件和工具，还能把一个任务拆给多个 agent 并行做；OpenAI 工程师凌晨 3 点用 dot 排障，再接一个 Codex 线程在手机上指挥修复。
501B Reflection Beam 总参数（23B 激活）
Reflection BeamGrok BotChatGPTGPT-6 Astra ProCodexClaude CodeClaude Code ReviewCursorDevinFactoryClineQoderGemini 4 ArgonAntigravityPerplexityAnthropicClaude Mythos 5piOpenCodeGLM 5.3 Flash
Issue 036
Oct 17, 2026 · …
Reflection opens Beam: 501B / 23B active, formerly the stealth cat-jam: Apache 2.0, 1M context, shipped together with Mirror CLI, the harness it was trained on; Vercel says it previously ran in stealth as cat-jam; free on Ollama for a limited time and wired straight into Claude Code, Codex and OpenCode.Grok Bot gets a speed-and-polish pass: Computer tasks finish 3× sooner on average, other replies in under half the time, and chat cards are redesigned; new connectors like Robinhood and Coinbase, marketplace submissions open to teams; a roundup of official bot templates hit 3.2k bookmarks.OpenAI: GPT-6 Astra Pro runs multiple agents in parallel: Astra Pro in ChatGPT is faster and better with plugins and tools, and can split one task across parallel agents; an OpenAI engineer triaged a 3am incident with his dot, then steered a Codex thread from his phone.
501B Reflection Beam total parameters (23B active)
Reflection BeamGrok BotChatGPTGPT-6 Astra ProCodexClaude CodeClaude Code ReviewCursorDevinFactoryClineQoderGemini 4 ArgonAntigravityPerplexityAnthropicClaude Mythos 5piOpenCodeGLM 5.3 Flash
第 035 期
2026 年 10 月 16 日 · …
Muse Spark Open：Meta 的 816B 开放权重 coding 模型：816B 总参 / 42B 激活、1M 上下文、原生多模态，主打长程任务和硬核软件工程；Ollama 当天上云端，可直接接 Claude Code；Muse Wonder 同日进 Cursor，CursorBench 46.1%。Gemini 4 Argon 进 Antigravity，Pro 用户另给周额度：Argon 向付费 Google AI Pro 开放，额度与 3.8 Flash 分开算；但也有开发者实测写代码不如 Opus、幻觉多。Anthropic 对投资人称已盈利，Fable 5.5 传言再起：FT 报 Anthropic 告诉投资人公司已经盈利、在为 IPO 造势；同日开放 STEM / 概念推理两类 Fellowship；theo 一句「今天发 Fable 5.5 就结束了」6k+ 赞。
816B Muse Spark Open 总参数（42B 激活）
Muse Spark OpenMuse WonderCursorGemini 4 ArgonAntigravityGrok BotAnthropicOpenCodeDevinChatGPTCodexMagnitudeT3 CodeMayoipi-durable
Issue 035
Oct 16, 2026 · …
Muse Spark Open: Meta's 816B open-weight coding model: 816B total / 42B active, 1M context, natively multimodal, built for long-horizon tasks and hard software engineering; Ollama had it on cloud the same day, wired into Claude Code; Muse Wonder also landed in Cursor at 46.1% on CursorBench.Gemini 4 Argon arrives in Antigravity with a separate weekly Pro quota: Argon opens to paid Google AI Pro, with a quota separate from 3.8 Flash; but some developers report it codes worse than Opus and hallucinates a lot.Anthropic tells investors it's profitable; Fable 5.5 rumors return: FT reports Anthropic told investors it is profitable as it rallies support for an IPO; the same day it opened STEM and Conceptual Reasoning Fellowships; theo's “if they drop Fable 5.5 today it's over” got 6k+ likes.
816B Muse Spark Open total parameters (42B active)
Muse Spark OpenMuse WonderCursorGemini 4 ArgonAntigravityGrok BotAnthropicOpenCodeDevinChatGPTCodexMagnitudeT3 CodeMayoipi-durable
第 034 期
2026 年 10 月 15 日 · …
Claude Code 桌面端「v2」：多会话并行成主角：新会话屏、可存环境（文件夹 / SSH / 云）、并行任务自动开嵌套子会话；同日 Claude Code 加上 idle compaction，闲置约 1 小时前先压缩，回来少烧额度。Ultramarine 1 Flash：5 人、约 $300k，自称对标 Fable 5：Paper Instruments 称性能与 Fable 5 相近、便宜 60×；API 已开，权重承诺两周内放出——「开源」目前仍是预告。订阅不是营收大头：Anthropic Q3 $16.8bn：FT 报 Anthropic 第三季度营收 $16.8bn；theo 直言有公司每年在 Anthropic 花超 $5 亿，并晒出 106 万次请求数据：9 月 22 日起几乎全面从 Codex 迁到 Claude Code。
$16.8bn FT：Anthropic 第三季度营收
Claude CodeUltramarineAnthropicCodexGrok BotClaude PlatformMuse WonderFactoryGitLab
Issue 034
Oct 15, 2026 · …
Claude Code Desktop “v2”: parallel sessions take center stage: New session screen, saved environments (folders / SSH / cloud), nested sub-sessions for parallel tasks; Claude Code also gained idle compaction, compacting before the ~1h prompt cache expires to save usage.Ultramarine 1 Flash: 5 people, ~$300k, aimed at Fable 5: Paper Instruments claims Fable 5–class performance at 60× lower cost; the API is live and weights are promised within two weeks — “open source” is still a pledge for now.Subscriptions aren’t the revenue: Anthropic Q3 at $16.8bn: FT reports Anthropic Q3 revenue of $16.8bn; theo says some companies spend over $500M a year with Anthropic, and shows 1.06M requests revealing his switch from Codex to Claude Code on Sep 22.
$16.8bn FT: Anthropic third-quarter revenue
Claude CodeUltramarineAnthropicCodexGrok BotClaude PlatformMuse WonderFactoryGitLab
第 033 期
2026 年 10 月 14 日 · …
Gemini 4 Argon：一天铺进 Cursor / Copilot / OpenCode：DeepMind 全球上线；Cursor 报 CursorBench 52.8%（对标 Opus 5.5），价 $2/$10；GitHub Copilot、OpenCode、Antigravity 同日接通。Anthropic：日合并过 3 万，Git 扛不住自建源控：工程博客写清：5 月约 2,500 次/日 → 9 月逾 30,000；五个月从 git 迁完 monorepo。社区称系统为 Scion。GPT-5.6 Sol + XBOW：6 天拿下 Chrome full-chain：约 $40,000 算力、Google 2026 年首笔 full-chain 赏金 $250,000；gdb 转发强调防御侧也该用上这些工具。
30,000+ Anthropic 九月每日合并次数（五月约 2,500）
Gemini 4 ArgonCursorAnthropicCodexOpenCodeGrok Bot
Issue 033
Oct 14, 2026 · …
Gemini 4 Argon: into Cursor / Copilot / OpenCode in one day: DeepMind went global; Cursor reports 52.8% on CursorBench (on par with Opus 5.5) at $2/$10; GitHub Copilot, OpenCode, and Antigravity lit up the same day.Anthropic: 30k+ merges/day broke human-scale Git: Engineering post: ~2,500 merges/day in May → 30,000+ in September; five months to migrate the monorepo off git. Community calls the system Scion.GPT-5.6 Sol + XBOW: Chrome full-chain in 6 days: About $40,000 of compute and Google’s first 2026 full-chain bounty at $250,000; gdb amplified the defender-side takeaway.
30,000+ Anthropic daily merges in September (≈2,500 in May)
Gemini 4 ArgonCursorAnthropicCodexOpenCodeGrok Bot
第 032 期
2026 年 10 月 13 日 · …
Muse Wonder：Meta 正面冲进 agent 模型榜：Zuckerberg 与 Alexandr Wang 发布 Muse Wonder / Wonder Full；theo 把 Wonder Full 与 Opus 5.5 并列 S 档，并提醒 API 版与订阅 Full 版差距很大。The X Subscription：Grok + Cursor + X Premium 一本账：SpaceXSI / Cursor / poteto 同日宣布统一订阅，付费档共用额度池，Premium 以上含 Grok Bot，Grok 模型 coding 用量是对等 Cursor 套餐的两倍。OpenCode 2.0 + Bot 当编排器成主流用法：官方大改版上线后台 Service、Code Mode、后台子 agent；中文圈传开「用原生 API 调云端 Grok Bot」省视频/X 接口费；poteto 把验证环定为「正确代码」入口。
$8 The X Subscription 入门价（含 Grok / Cursor / X Premium）
Muse WonderCursorOpenCodeGrok BotClaude Code
Issue 032
Oct 13, 2026 · …
Muse Wonder: Meta storms the agent model board: Zuckerberg and Alexandr Wang launched Muse Wonder / Wonder Full; theo put Wonder Full next to Opus 5.5 in S-tier and warned that the API Wonder is far weaker than subscriber Full.The X Subscription: Grok + Cursor + X Premium, one bill: SpaceXSI / Cursor / poteto announced a unified plan with a shared usage pool; Premium+ includes Grok Bot; Grok coding usage is 2× the equivalent Cursor plan.OpenCode 2.0 + Bot-as-orchestrator goes mainstream: Official overhaul ships background Service, Code Mode, and background subagents; CN Twitter spreads calling cloud Grok Bot via native API; poteto pins verification as the entry to correct agent code.
$8 The X Subscription entry price (Grok / Cursor / X Premium)
Muse WonderCursorOpenCodeGrok BotClaude Code
第 031 期
2026 年 10 月 12 日 · …
Codex Day 7：上下文加长且不加价：Tibo 宣布 GPT-6.1 Sol 在 Codex 与 Work 里默认上下文从 272k 提到 500k（约 +80%），订阅用量不收 long context 附加费，长任务更少被迫 compaction。Grok Bot 从「能下单」变成「真帮忙」：有人让 Bot 用虚拟卡买到全球缺货的 iPhone 18 Pro Max；Elon 转发称把它当 24/7 超级员工；poteto 一篇「不知道用 Bot 干嘛」的帖几乎攒到三千收藏。agent 互联与「好代码」定义一起翻盘：Yobiru 开源让跨机 agent 经 Tailscale 互发消息并叫醒对方；Open Agent Connect 接入 Grok Bot；poteto 掀起「agent 时代还要不要 clean code」大讨论，DHH 与 Addy Osmani 跟进。
500k GPT-6.1 Sol 在 Codex / Work 的新默认上下文长度（原 272k）
CodexGrok BotClaude CodeCursorYobiru
Issue 031
Oct 12, 2026 · …
Codex Day 7: longer context, no surcharge: Tibo raised GPT-6.1 Sol’s default context in Codex and Work from 272k to 500k (~80%), and subscription usage no longer pays long-context fees — fewer forced compactions on long tasks.Grok Bot moves from demo to “it actually helped”: Someone had the Bot buy a sold-out iPhone 18 Pro Max with a virtual card; Elon quote-tweeted “this actually works” and called it a 24/7 super-employee; poteto’s “what do I use Bot for?” post neared 3K bookmarks.Agent mesh + a rewrite of “Good Code”: Yobiru open-sources cross-machine agent messaging and wake-ups over Tailscale; Open Agent Connect adds Grok Bot; poteto sparks a debate on clean code in the agent era, with DHH and Addy Osmani weighing in.
500k New default context length for GPT-6.1 Sol in Codex / Work (was 272k)
CodexGrok BotClaude CodeCursorYobiru
第 030 期
2026 年 10 月 11 日 · …
订阅与 agent 开始互通：Cognition 宣布任意个人 ChatGPT 套餐可接到 Devin，GPT 用量走你自己的配额；ChatGPT Dot 也能在 Codex 里开活、跟帖，并改 ChatGPT Work 的定时任务。Codex Day 6 是全球额度 Reset：Tibo「28 天」第 6 天没有上新功能，而是预告当日结束前全球重置 ChatGPT / Codex 用量；社区把「额度回血」当成这周最好用的功能之一。并行项目与开源侧继续卷：有人同时跑 16 个 Cursor Projects；OpenCode 把 OpenTunnel 迁到 10 个 Fly 区域；Browser Use 自称一年便宜 100 倍；非官方 DeepSeek Bot 插件正式介绍。
60%+ Vercel 上现已由 agent 发起的部署占比（今年 1 月约 3%）
DevinCodexGrok BotCursorOpenCode
Issue 030
Oct 11, 2026 · …
Subscriptions start crossing agent boundaries: Cognition lets any personal ChatGPT plan connect to Devin (GPT usage draws from your quota). ChatGPT Dots can start and follow Codex work and edit ChatGPT Work scheduled tasks.Codex Day 6 is a global usage reset: Day 6 of Tibo's 28-day run ships no new feature—just a promised end-of-day global ChatGPT / Codex usage reset that the community treated as the week's best 'feature'.Parallel projects and open tooling keep pushing: Someone is running 16 Cursor Projects at once. OpenCode moved OpenTunnel to 10 Fly regions. Browser Use claims a 100× cost drop in a year. An unofficial DeepSeek Bot plugin gets a formal intro.
60%+ Share of Vercel deployments that are now agentic (≈3% in Jan 2026)
DevinCodexGrok BotCursorOpenCode
第 029 期
2026 年 10 月 10 日 · …
Claude 把多智能体做成产品能力：Managed Agents dynamic workflows 公测（可编排最多 1000 个 agent，找 bug 实验 66/70）；Claude Code Projects 向所有 Pro / Max 等候名单用户开放。Codex 与 Grok Bot 各补一块交互：Codex Day 5 上线 Composer predictions（Pro 桌面端预测下一条指令）；Grok Bot 宣布自有邮箱，可用来注册服务、联系商家，Elon 还说要用 Grok Bots 管 Grokipedia。对齐报告与开源侧继续动：Anthropic 开始更频繁发布模型行为报告；OpenClaw 拿下 .claw 顶级域名；Cline 限时免费 Solar Mini 4；OpenCode 继续放 iOS TestFlight。
1000 Claude Managed Agents 单次 workflow 可编排的 agent 上限
ClaudeCodexGrok BotOpenCodeDevin
Issue 029
Oct 10, 2026 · …
Claude productizes multi-agent orchestration: Managed Agents dynamic workflows enter public beta (up to 1,000 agents per run; 66/70 bugs in their planted-bug demo). Every Pro / Max user on the Claude Code Projects waitlist is let in.Codex and Grok Bot each ship a new interaction surface: Codex Day 5 adds Composer predictions (Pro desktop suggests your next message). Grok Bot gets its own email for signups and outreach; Elon says Grok Bots will manage Grokipedia.Alignment reports and open-source keep moving: Anthropic starts publishing more frequent model-behavior reports. OpenClaw wins the .claw TLD. Cline makes Solar Mini 4 free for a limited time. OpenCode opens more iOS TestFlight slots.
1,000 Max agents Claude Managed Agents can orchestrate in one workflow
ClaudeCodexGrok BotOpenCodeDevin
第 028 期
2026 年 10 月 9 日 · …
Claude 往办公软件深处走：Claude Dashboards 与 Motion 公测（3.27 万赞、近 2 万收藏），Docs / Slides / Design 对所有套餐开放；Anthropic 同日推出 Cyber Mission，给开源项目免费做漏洞扫描。速度和入口同时开打：Codex Day 4 上线即时 steering 和 GPT-6.1 Sol Ultrafast（最高快 8 倍）；Google 发布通用工作 agent「Gemini」；Grok Bot 借 Claude Code / Codex 接入别家订阅，并接上 Shopify。开发者在争论「量」：一天合并几百个 PR 到底有没有意义；OpenCode 月活接近 2000 万；Qwen3.8-27B 成了开源压缩、本地推理和 RL 实验的共同底座。
8 倍 GPT-6.1 Sol Ultrafast 相比 Standard 的最高提速
ClaudeCodexGeminiGrok BotOpenCode
Issue 028
Oct 09, 2026 · …
Claude moves deeper into office work: Claude Dashboards and Motion enter beta (32.7K likes, nearly 20K bookmarks); Docs / Slides / Design open to every plan. The same day, Anthropic launches its Cyber Mission with free vulnerability scans for open-source projects.Speed and the front door, at the same time: Codex Day 4 ships instant steering and GPT-6.1 Sol Ultrafast (up to 8x faster); Google announces a universal work agent called 'Gemini'; Grok Bot taps other subscriptions via Claude Code / Codex and connects to Shopify.Developers argue about volume: Do hundreds of merged PRs a day mean anything? OpenCode nears 20M monthly actives; Qwen3.8-27B becomes the shared base for open compression, local inference and RL experiments.
8x GPT-6.1 Sol Ultrafast's max speedup over Standard
ClaudeCodexGeminiGrok BotOpenCode
第 027 期
2026 年 10 月 8 日 · …
小模型与便宜算力成主线：Haiku 5.5（3.3 万赞）10 万 token 内只要 Haiku 4.5 的十分之一，Sonnet 5.5 缓存读取减半，Max / Team 每月附送 API credits；微软把 137B 的 MAI-Code 搬到本地 PC。入口被重做：GPT-6 带 Intelligent UI 进入 ChatGPT，回答可以直接是可交互界面；Codex 与 ChatGPT Work 活跃用户到 4000 万；Grok Bot 改为按任务挑后端模型（9.5 万赞）。成本账被摊开：Theo 的 tsc-rs 烧了约 40 万美元 Codex token 没做成，换 Opus 约 2 万美元两周搞定；Cursor 重度用量被推算到一年 1.32 亿美元，Theo 逐条反驳。
1/10 Haiku 5.5 在 10 万 token 内的价格（对比 Haiku 4.5）
Haiku 5.5GPT-6CodexGrok BotCursor
Issue 027
Oct 08, 2026 · …
Small models and cheap compute set the agenda: Haiku 5.5 (33K likes) costs a tenth of Haiku 4.5 under 100K tokens; Sonnet 5.5 cache reads halved; Max / Team get monthly API credits. Microsoft brings 137B MAI-Code to local PCs.The front door gets rebuilt: GPT-6 with Intelligent UI comes to ChatGPT — answers can be interactive interfaces; Codex + ChatGPT Work hits 40M active users; Grok Bot now picks the best back-end model per task (95K likes).Agent costs get itemized: Theo's tsc-rs burned ~$400K of Codex tokens with no result, then ~$20K of Opus finished it in two weeks; Cursor heavy usage extrapolated to $132M/yr, Theo rebuts point by point.
1/10 Haiku 5.5 price under 100K tokens (vs Haiku 4.5)
Haiku 5.5GPT-6CodexGrok BotCursor