[{"data":1,"prerenderedAt":378},["ShallowReactive",2],{"header-counts":3,"footer-counts":6,"news-2026-ai-coding-july-2026":9},{"tools":4,"reviews":5},78,26,{"tools":4,"reviews":5,"playbooks":7,"news":8},22,24,{"id":10,"title":11,"body":12,"cover":366,"description":367,"extension":368,"lastVerified":369,"meta":370,"navigation":371,"path":372,"published":369,"seo":373,"sourceName":374,"sourceUrl":375,"stem":376,"__hash__":377},"news\u002Fnews\u002F2026\u002Fai-coding-july-2026.md","2026 年 7 月 AI 编程动态：GPT-5.6 登场、Cursor 换血，行业集体忙着装'护栏'",{"type":13,"value":14,"toc":353},"minimark",[15,19,58,65,69,80,100,107,116,120,131,150,154,161,183,187,194,246,253,257,291,295,307,310],[16,17,18],"h2",{"id":18},"要点",[20,21,22,30,36,46,52],"ul",{},[23,24,25,29],"li",{},[26,27,28],"strong",{},"GPT-5.6 登场（7-09）","：Sol \u002F Terra \u002F Luna 取代 GPT-5.5 时代，新增 Programmatic Tool Calling、多 agent beta、Ultra 模式（默认 4 并行、可配 16）",[23,31,32,35],{},[26,33,34],{},"Cursor Grok 4.5（7-08）","：跨桌面\u002FWeb\u002FiOS\u002FCLI\u002FSDK，号称用 Cursor 代码库交互数据训练；Cursor 3.0 用 Agents Window 取代旧布局",[23,37,38,41,42],{},[26,39,40],{},"Claude Code","：Sonnet 5 于 6-30 成默认模型，Fable 5 于 7-01 全球恢复；v2.1.206 默认开启云上 auto mode、新增 ",[43,44,45],"code",{},"\u002Fcommit-push-pr",[23,47,48,51],{},[26,49,50],{},"护栏周（7-06~11）","：Codex CLI 加 writes 审批、OpenHands 加 Budgets、Zed 加本地 llama.cpp——主题清一色\"治理\"",[23,53,54,57],{},[26,55,56],{},"格局变动","：Windsurf 被 Cognition 收编为 Devin Desktop；Amazon Q 向 Kiro 过渡；Copilot 6-01 起改用量计费；OpenAI 公开退出 SWE-bench Verified",[59,60,61],"blockquote",{},[62,63,64],"p",{},"本文为 AIHO 基于 Dreaming.press、Neodrop、SDD、AgentsIndex 等 2026-07 公开报道的整合整理。",[16,66,68],{"id":67},"一gpt-56重排-agent-技术栈","一、GPT-5.6：重排 agent 技术栈",[62,70,71,72,75,76,79],{},"7 月 9 日，OpenAI 发布 ",[26,73,74],{},"GPT-5.6"," 三款模型 ",[26,77,78],{},"Sol \u002F Terra \u002F Luna","，取代 GPT-5.5 时代。同步推出：",[20,81,82,88,94],{},[23,83,84,87],{},[26,85,86],{},"Programmatic Tool Calling","：结构化、可编排的工具调用；",[23,89,90,93],{},[26,91,92],{},"多 agent beta","：内置多 agent 协作；",[23,95,96,99],{},[26,97,98],{},"Ultra 模式","：默认 4 个并行 agent，可配置至 16 个。",[62,101,102,103,106],{},"Devin Desktop \u002F CLI 同日接入三款模型，GitHub Copilot 在 IDE \u002F CLI \u002F Web \u002F 云 agent \u002F 移动端全面暴露 GPT-5.6 Sol\u002FTerra\u002FLuna；GitHub 还把 ",[26,104,105],{},"Kimi K2.7 Code"," 加进 Business \u002F Enterprise，逼着管理员开始思考\"开源权重模型怎么进企业策略\"。",[62,108,109,110,115],{},"编程基准上，GPT-5.6 在 Terminal-Bench 系列保持领先（",[111,112,114],"a",{"href":113},"\u002Fmodels\u002Fgpt-5-6.html","模型卡"," 官方 2.1 口径 Sol 达 91.9%）；第三方 2026-07 实测另给 Terminal-Bench 2.0 口径 85.6%，并指其 SWE-bench Pro 仅 64.1%，落后同期 Claude Opus 4.8 的 69.2%。",[16,117,119],{"id":118},"二cursor换模型换界面","二、Cursor：换模型、换界面",[62,121,122,123,126,127,130],{},"Cursor 7 月 8 日上线 ",[26,124,125],{},"Grok 4.5","，覆盖桌面\u002FWeb\u002FiOS\u002FCLI\u002FSDK，官方称用\"数万亿 token 的 Cursor 代码库与开发者-agent 交互数据\"训练。有意思的是，Cursor 在 Grok 4.5 评估中",[26,128,129],{},"排除了 CursorBench","——因为早期 Cursor 代码库快照意外混入了训练数据。这给所有人提了个醒：接受任何厂商的 coding-agent 分数前，先问 benchmark 数据干净不干净。",[62,132,133,134,137,138,141,142,145,146,149],{},"界面侧，",[26,135,136],{},"Cursor 3.0"," 用 ",[26,139,140],{},"Agents Window"," 取代旧布局，云 agent 移出 Editor，Side Chats（",[43,143,144],{},"\u002Fside","、",[43,147,148],{},"\u002Fbtw","）可开持久 agent 会话并回 @ 主线程，还加了 Cmd+K 转录搜索、五个云 agent hooks。",[16,151,153],{"id":152},"三claude-code安静地变自主","三、Claude Code：安静地变\"自主\"",[62,155,156,157,160],{},"6 月 30 日，",[26,158,159],{},"Sonnet 5 悄然取代 Sonnet 4.8"," 成为 Claude Code 与各 Claude 套餐默认模型（无发布会，只有 changelog），原生 1M 上下文。Fable 5 因出口管制 6-12 下线、7-01 恢复。",[62,162,163,164,167,168,171,172,174,175,178,179,182],{},"v2.1.205–207 的变化更实质：auto mode 在 Bedrock \u002F Vertex \u002F Foundry 上",[26,165,166],{},"默认开启","（不想自主就显式 ",[43,169,170],{},"disableAutoMode","）；新增 ",[43,173,45],{}," 推送自动化、",[43,176,177],{},"\u002Fdoctor"," 可修剪 CLAUDE.md、",[43,180,181],{},"\u002Fcd"," 路径建议、外部 worktree 确认、升级后自动后台 agent 更新。默认假设已从\"人盯着\"切到\"云上自主运行\"。",[16,184,186],{"id":185},"四护栏周行业集体装刹车","四、\"护栏周\"：行业集体装刹车",[62,188,189,190,193],{},"7 月 6–11 日这一周，Codex、OpenHands、Claude Code、Zed 同一周发版，",[26,191,192],{},"几乎没有一条是关于\"写更好代码\"的","，全是审批、预算、运行边界：",[195,196,197,210],"table",{},[198,199,200],"thead",{},[201,202,203,207],"tr",{},[204,205,206],"th",{},"工具（版本）",[204,208,209],{},"装了什么\"刹车\"",[211,212,213,222,230,238],"tbody",{},[201,214,215,219],{},[216,217,218],"td",{},"Codex CLI v0.144.x（7-09）",[216,220,221],{},"writes 审批模式（只读免打扰、写才弹确认）+ 交互式 MCP 认证",[201,223,224,227],{},[216,225,226],{},"OpenHands v1.11.0（7-09）",[216,228,229],{},"Budgets 仪表盘 + Agent Profiles + BYO-key",[201,231,232,235],{},[216,233,234],{},"Claude Code v2.1.205–207",[216,236,237],{},"云上 auto mode 默认开，agent-teams 崩溃循环修复",[201,239,240,243],{},[216,241,242],{},"Zed v1.10.2（7-10）",[216,244,245],{},"llama.cpp 本地模型 provider，agent 完全跑在自己机器上",[62,247,248,249,252],{},"一句话总结这周：",[26,250,251],{},"前沿从\"它能不能写\"移到了\"你能不能放心让它无人值守地跑\"","。而 fire-and-forget 只有在刹车（审批范围、成本上限、BYO-key、本地选项）齐备时才成立。",[16,254,256],{"id":255},"五格局合并退场计费变天","五、格局：合并、退场、计费变天",[20,258,259,265,275,285],{},[23,260,261,264],{},[26,262,263],{},"Windsurf 被 Cognition 收购","，以 Devin Desktop 名义重启；",[23,266,267,270,271,274],{},[26,268,269],{},"Amazon Q Developer"," 进入向 ",[26,272,273],{},"Kiro"," 接班的下坡路；",[23,276,277,280,281,284],{},[26,278,279],{},"GitHub Copilot"," 2026-06-01 起从包月切到",[26,282,283],{},"按 AI Credits 用量计费","，平价时代结束；",[23,286,287,290],{},[26,288,289],{},"OpenAI 公开退出 SWE-bench Verified","——这个半数榜单还在引用的基准，公信力被自己人抽走。",[16,292,294],{"id":293},"aiho-观点","AIHO 观点",[62,296,297,298,301,302,306],{},"2026 年 7 月最值得记的，不是哪个模型又高了几个点，而是",[26,299,300],{},"整个赛道在用同一个动作表态","：把 agent 从\"演示玩具\"升级为\"可托付的生产系统\"。审批模式、成本仪表盘、本地选项——这些不性感，却是从\"能跑\"到\"敢让它跑\"的必经之路。这正呼应 ",[111,303,305],{"href":304},"\u002Fwiki\u002Fagentic-coding.html","Agentic Coding"," 在 2026 年从\"马力\"转向\"护栏\"的主线。",[16,308,309],{"id":309},"相关阅读",[20,311,312,322,340,345],{},[23,313,314,315,317,318],{},"概念：",[111,316,305],{"href":304}," \u002F ",[111,319,321],{"href":320},"\u002Fwiki\u002Fvibe-coding.html","Vibe Coding",[23,323,324,325,317,328,317,332,317,336],{},"工具：",[111,326,40],{"href":327},"\u002Fcoding\u002Fcli\u002Fclaude-code.html",[111,329,331],{"href":330},"\u002Fcoding\u002Fcli\u002Fcodex.html","Codex CLI",[111,333,335],{"href":334},"\u002Fcoding\u002Fide\u002Fcursor.html","Cursor",[111,337,339],{"href":338},"\u002Fcoding\u002Fagent\u002Fopenhands.html","OpenHands",[23,341,342,343],{},"模型：",[111,344,74],{"href":113},[23,346,347,348,352],{},"跨域呼应：",[111,349,351],{"href":350},"\u002Fwiki\u002Fembodied-ai.html","具身智能","（Agent 从数字走向物理）",{"title":354,"searchDepth":355,"depth":355,"links":356},"",3,[357,359,360,361,362,363,364,365],{"id":18,"depth":358,"text":18},2,{"id":67,"depth":358,"text":68},{"id":118,"depth":358,"text":119},{"id":152,"depth":358,"text":153},{"id":185,"depth":358,"text":186},{"id":255,"depth":358,"text":256},{"id":293,"depth":358,"text":294},{"id":309,"depth":358,"text":309},null,"2026 年 7 月 AI 编程圈主线不是'写得更好'，而是'管得更稳'：OpenAI 发布 GPT-5.6（Sol\u002FTerra\u002FLuna）并推 Ultra 多 agent 模式，Cursor 上 Grok 4.5 与 3.0 Agents Window，Claude Code 默认开启 auto mode，Codex\u002FOpenHands\u002FZed 集体加审批与成本护栏。","md","2026-07-23",{},true,"\u002Fnews\u002F2026\u002Fai-coding-july-2026",{"title":11,"description":367},"Dreaming.press \u002F Neodrop \u002F SDD 综合","https:\u002F\u002Fdreaming.press\u002Fposts\u002Fcoding-agents-shipped-guardrails-not-horsepower-july-2026.html","news\u002F2026\u002Fai-coding-july-2026","8VwayCP7TKwYBqA4pIUR4sZs-NSs37931du8ahAXoAA",1784900533855]