七月二十五日

Anthropic launched Claude Opus 5, available today on all paid plans and the Claude API at the same price as Opus 4.8. It's the default model on Claude Max, the strongest on Claude Pro, and also ships in Fast mode at around 2.5× the default speed.

Anthropic 发布了 Claude Opus 5，即日起在所有付费套餐和 Claude API 上线，价格与 Opus 4.8 持平。它是 Claude Max 的默认模型，也是 Claude Pro 上最强的模型，还提供约 2.5 倍默认速度的 Fast 模式。

According to Anthropic's automated behavioral audit, Opus 5 is its most aligned model to date, with the lowest rates of reckless or deceptive behavior and the strongest adherence to Claude's Constitution.

根据 Anthropic 的自动化行为审计，Opus 5 是其迄今对齐程度最高的模型，鲁莽或欺骗行为的发生率最低，对 Claude Constitution 的遵循度最强。

On cybersecurity, Opus 5 is stronger than Opus 4.8 but remains substantially behind Mythos 5 at developing exploits; its safeguards aim to let developers find and fix vulnerabilities while blocking high-risk uses.

在网络安全方面，Opus 5 强于 Opus 4.8，但在漏洞利用开发上仍明显落后于 Mythos 5；其安全防护的设计目标是让开发者能够发现并修复软件漏洞，同时拦截高风险用途。

Boris Cherny says the most exciting thing about Opus 5 isn't the eval scores: it's Anthropic's least prompt-injectable model yet. Buried in the system card, across prompt-injection evals and red teaming, Opus 5 is very hard to inject successfully, and when you layer defenses (model alignment, prompt injection probes, and Auto Mode in Claude Code) the attack success rate drops to roughly zero.

Boris Cherny 说 Opus 5 最令人兴奋的不是各项 eval 分数，而是它是 Anthropic 迄今最难被 prompt injection 攻击的模型。这一点藏在 system card 深处：在 prompt injection 评测和红队测试中，Opus 5 都极难被成功注入；而当多层防御叠加起来(模型对齐、prompt injection 探针、加上 Claude Code 的 Auto Mode)，攻击成功率会降到接近于零。

Thariq revealed that Anthropic removed roughly 80% of the Claude Code system prompt for its newest models, and shared what the team learned about writing system prompts, skills, and CLAUDE.md files for them.

Thariq 透露，Anthropic 为最新模型删掉了大约 80% 的 Claude Code system prompt，并分享了团队由此总结的为新模型编写 system prompt、skill 和 CLAUDE.md 的经验。

He calls Opus 5 an incredible daily driver that rounds out the Claude 5 family: pair it with Fable for planning, brainstorming, or fixing the hardest bugs.

他称 Opus 5 是极出色的日常主力模型，补全了 Claude 5 家族：日常用 Opus 5，规划、头脑风暴或修最难的 bug 时搭配 Fable。

Cat Wu highlights that Opus 5 is great at long-running autonomous work and invites people to try it out.

Cat Wu 强调 Opus 5 在长时间自主运行的任务上表现出色，欢迎大家上手体验。

Alex Albert says that just over six months on, Opus 5 now produces near-superhuman spreadsheets and slide decks that match what a consultant would make. He also shared his favorite launch graphs: the team put heavy work into making the model token-efficient across domains while still raising the intelligence bar, and he prefers it over Fable 5 for many coding tasks.

Alex Albert 说，仅仅过了六个多月，Opus 5 已能产出接近超人水平、堪比咨询顾问作品的电子表格和幻灯片。他还分享了自己最喜欢的几张发布图表：团队下了大功夫让模型在各领域都更省 token，同时还在拉高智能上限；在很多编程任务上，他甚至更偏爱 Opus 5 而不是 Fable 5。

The contrarian take of launch day: after a week of testing, Dan Shipper calls Opus 5 "a hard model to love." It argued with instructions, stopped before work was finished, and broke Every's existing skills and plugins, until the team deleted their elaborate workflows and started from scratch, at which point it got dramatically better and showed "flashes of brilliance." His tips: it breaks backward compatibility with existing skills, works better on medium or low thinking effort, and is "a poor man's Fable," with Fable's personality quirks but not its genius top end.

发布日最"唱反调"的观点来自 Dan Shipper：测了一周后，他称 Opus 5 是"一个很难爱上的模型"。它会跟指令抬杠、活儿没干完就停手，还搞坏了 Every 现有的 skill 和插件；直到团队删掉精心搭建的旧工作流、从零开始，它才有了脱胎换骨的表现，甚至展现出"天才的闪光"。他的建议：这个模型不向后兼容现有 skill；用中低思考强度效果反而更好；它是"丐版 Fable"，有 Fable 的性格怪癖，却没有 Fable 的天才上限。

Aaron Levie shared Box's test results running Opus 5 through its Complex Work Eval on real enterprise document work: due diligence +17%, life sciences target identification +30%, legal contract review +12%, technology +19%, and healthcare +13% over Opus 4.8, with the model staying thorough as checklists grow instead of catching the obvious items and stopping.

Aaron Levie 分享了 Box 用自家 Complex Work Eval 在真实企业文档任务上测试 Opus 5 的结果：相比 Opus 4.8，尽职调查 +17%，生命科学靶点识别 +30%，法律合同审查 +12%，科技行业 +19%，医疗健康 +13%；而且随着检查清单变长，模型依然保持彻底，而不是抓完显眼项就停手。

He also signed the open-weights letter on behalf of Box, arguing open vs. closed is not a zero-sum battle: open weights accelerate AI diffusion across the economy, enable dozens of vertical post-training attempts instead of waiting on a few players, and let high-end orchestration go to closed frontier models while workhorse tasks run cheaply.

他还代表 Box 签署了开放权重公开信，并主张开源与闭源并非零和博弈：开放权重能加速 AI 在整个经济中的扩散，让各垂直领域出现几十上百次 post-training 尝试，而不是苦等少数几家巨头；高端编排任务交给闭源前沿模型，跑量的任务则可以用更低成本完成。

Sam Altman backed the open-weights letter: "i want the US to win in AI both in open source and proprietary models, and i am glad to see this." He also asked users for feedback on "pro-ultra-superhard," a new top-effort mode.

Sam Altman 也为开放权重公开信站台："我希望美国在 AI 上既赢下开源也赢下闭源，很高兴看到这封信。"他还邀请用户反馈新的最高强度档位 "pro-ultra-superhard" 的使用体验。

Amjad Masad publicly pressed Anthropic on the open-weights letter: "Will Anthropic sign? If you work at Anthropic worth asking your leadership to sign or make their position clear: Are they for banning open weight models?"

Amjad Masad 公开向 Anthropic 施压，追问其对开放权重公开信的立场："Anthropic 会签吗？如果你在 Anthropic 工作，值得去问问你们的领导层：要么签字，要么把立场说清楚，他们是否支持封禁开放权重模型？"

He also noted the irony of VCs who passed on chip startup Etched in early rounds now waking up, and teased that Replit has changed a lot for anyone who hasn't used it recently.

他还感叹当年在早期轮次错过芯片创业公司 Etched 的 VC 们如今终于"醒了"，并预告许久没用 Replit 的人再打开会大吃一惊。

Thibault Sottiaux announced that ChatGPT Work is now available globally for all paid plans across mobile, web, and desktop: "Puts a jetpack on your ChatGPT."

Thibault Sottiaux 宣布 ChatGPT Work 已面向全球所有付费套餐上线，覆盖移动端、网页端和桌面端："给你的 ChatGPT 装上喷气背包。"

Josh Woodward announced Gemini Spark is live for all Google AI Pro subscribers in the US, expanding globally next. The pitch: "Gemini does more than talk. It gets things done." Example use case: drop in a school calendar PDF and tell Gemini to add every "No School" day to your Google Calendar.

Josh Woodward 宣布 Gemini Spark 已向美国所有 Google AI Pro 订阅用户上线，接下来会推向全球。他的卖点是："Gemini 不只会聊天，它能把事办成。"示例场景：扔进一份学校日历 PDF，让 Gemini 把每个"停课日"自动加进你的 Google Calendar。

Guillermo Rauch endorsed Figma2React with a three-word review: "It's good." He also jumped straight into stress-testing the newly released Opus 5.

Guillermo Rauch 用三个词点评了 Figma2React："It's good。"他还第一时间上手压测了刚发布的 Opus 5。

Garry Tan's macro take: to get economy-wide productivity gains from AI, managers and CEOs have to greenlight radically different staffing and workflow plans, and to date he doesn't think they've done it. "Be prepared for this to take 10 years not 2."

Garry Tan 的宏观判断：要让 AI 带来全经济范围的生产率提升，管理者和 CEO 们必须批准与现在截然不同的人员配置和工作流方案，而到目前为止他认为他们还没有这么做。"做好准备，这需要 10 年，而不是 2 年。"

He also amplified research finding that how fast a country adopted new technology over the last 200 years accounts for at least 25% of why some nations are rich and others aren't today.

他还转发了一项研究：过去 200 年里一个国家采纳新技术的速度，至少能解释当今各国贫富差距的 25%。

Matt Turck flagged a big week in model routing: Stripe is rumored to be acquiring OpenRouter for $10B, Cursor Router was announced Wednesday, Runway Router yesterday, plus routers in various forms at Databricks, Vercel, Cloudflare, Dataiku, AWS, and Google.

Matt Turck 指出这是模型路由领域的大周：传闻 Stripe 将以 100 亿美元收购 OpenRouter，Cursor Router 周三发布，Runway Router 昨天发布，此外 Databricks、Vercel、Cloudflare、Dataiku、AWS、Google 也都有各自形态的 router。

He also noted an under-discussed irony: the world's top AI researchers are researching their way out of their own jobs as they build recursive auto-research.

他还点出一个少有人讨论的讽刺：全球顶尖的 AI 研究员们正在通过构建递归式自动研究，把自己的工作研究没了。

Madhu Guru sees a massive opportunity over the next few years for people who can take messy real-world workflows and adapt foundation models to them: understanding how work actually gets done, designing evals, improving models through post-training, and building feedback loops. That's how a general-purpose model becomes exceptional for a specific domain, and today that skillset is still concentrated in a handful of labs.

Madhu Guru 认为，未来几年对能把杂乱的真实世界工作流与基础模型适配起来的人来说，存在巨大机会：理解工作到底是怎么完成的、设计 eval、通过 post-training 改进模型、搭建持续改进的反馈闭环。通用模型正是这样才能在特定领域变得出类拔萃，而这套技能眼下仍集中在少数几家实验室手里。

Zara Zhang says the #1 thing she wants from any model right now is speed: intelligence is already good enough, but waiting 1-5 minutes per task is the worst possible window, too short for deep work and too long to stare at the screen, so she ends up scrolling X. "Agents are making all of us more ADHD."

Zara Zhang 说她现在对任何模型的头号需求是速度：智能已经够用了，但每个任务等 1 到 5 分钟是最糟糕的时间窗口，做深度工作太短，干瞪着屏幕又太长，结果就是跑去刷 X。"Agent 正在让我们所有人都变得更 ADHD。"

She also observes that if you bring your agent into chat groups and meetings, the chat history and meeting transcripts become PRDs. Work culture used to favor written communicators; now verbal communicators have an equal chance because agents don't care.

她还观察到：如果把 agent 带进群聊和会议，聊天记录和会议转录稿就直接变成了 PRD。过去的职场文化偏爱擅长书面表达的人，现在口头表达者也有了同等机会，因为 agent 并不在乎形式。

Peter Yang's new favorite workflow: lying on the bed and talking to ChatGPT Voice to do work in Codex. His tip for doing it well: you have to remember the names of all your long-running threads.

Peter Yang 的新宠工作流：躺在床上跟 ChatGPT Voice 对话，驱动 Codex 干活。他的心得是：想用好这一套，你得记住所有长时间运行的线程的名字。

He also agrees that pure software is now really hard to monetize for indie developers; you need software plus something else, like services.

他还认同一个观点：如今独立开发者靠纯软件已经很难变现，你需要"软件 + 别的东西"，比如服务。

Peter Steinberger reports a new record for his autoreview skill: 66 rounds on a gnarly refactor.

Peter Steinberger 报告他的 autoreview skill 创下新纪录：在一次棘手的重构上连跑了 66 轮。

Swyx keeps shipping on SmolForge, adding customizable skins and spritesheet animations. He also says the reason he's building a new gsuite is frustration with "extremely stupid defaults" in the incumbent tools.

Swyx 的 SmolForge 持续迭代，新增了可自定义皮肤和 spritesheet 动画。他还说自己之所以在做一套新的 gsuite，就是被现有工具里"蠢到极点的默认设置"逼的。

Anthropic's security and product teams lay out how they cap the "blast radius" of increasingly autonomous agents across claude.ai, Claude Code, and Claude Cowork, and candidly walk through the failures. The core argument: design for containment at the environment layer first (sandboxes, VMs, egress controls), then steer behavior at the model layer, because "the deterministic boundary is what gets hit when everything probabilistic misses." Telemetry showed users approved roughly 93% of Claude Code permission prompts, so human-in-the-loop oversight decays with approval fatigue; the OS-level sandbox cut permission prompts by 84%. The confessions are the best part: an internal red-team phish got Claude Code to exfiltrate AWS credentials in 24 of 25 attempts because the malicious instructions arrived through the user; a third-party disclosure showed data exfiltrated through the api.anthropic.com allowlist itself, with an attacker-supplied API key uploading workspace files to the attacker's own Anthropic account. The hard-won lesson: an egress allowlist is not a destination filter but a capability grant. And a recurring theme: "the weakest layer is the one you built yourself." Battle-tested primitives like gVisor and hypervisors held; Anthropic's custom proxies were what broke.

Anthropic 的安全与产品团队详解了他们如何在 claude.ai、Claude Code 和 Claude Cowork 三条产品线上给日益自主的 agent 设定"爆炸半径"上限，并坦诚复盘了翻车案例。核心论点：优先在环境层做隔离(沙箱、虚拟机、出口流量控制)，再在模型层引导行为，因为"当所有概率性防御都失手时，扛住攻击的是确定性边界"。遥测数据显示用户对 Claude Code 权限弹窗的批准率高达约 93%，说明人工审批会因"审批疲劳"而失效；操作系统级沙箱则把权限弹窗减少了 84%。最精彩的是自曝部分：一次内部红队钓鱼演练中，Claude Code 在 25 次尝试里 24 次成功外泄了 AWS 凭证，因为恶意指令是经由用户之手输入的；另一起第三方披露显示，数据竟通过 api.anthropic.com 这个白名单域名本身外泄，攻击者植入自己的 API key，把工作区文件上传到了攻击者自己的 Anthropic 账户。血泪教训：出口白名单不是目的地过滤器，而是能力授予。还有一个反复出现的主题："最薄弱的一层永远是你自己造的那层。"gVisor、hypervisor 这些久经沙场的基础组件都稳如泰山，出问题的都是 Anthropic 自研的代理层。

The Takeaway: start from the use case and work backwards; DoorDash built its own delivery robot because everyone else built the technology first and went looking for a problem later.

DoorDash co-founders Andy Fang and Stanley Tang have quietly run a robotics program since 2018, starting as a skunkworks of "me and half an engineer." After years of partnering with everyone from sidewalk robots to robotaxis, they concluded neither fit: sidewalk bots at 2-3 mph can't cover the average 3-5 mile delivery, and a 4,000-pound robotaxi is overkill for a couple of burritos that can't walk the last block themselves. So they built Dot: a 300-pound, 20-mph, one-tenth-the-size-of-a-car L4 robot that rides both roads and bike lanes, delivering in Phoenix for two years.

Tang's recruiting pitch is blunt: "Do you wanna go work on prototypes and demos and be at a PhD lab, or do you wanna work on something where you can actually ship something in the real world?" The real world bites back in ways no demo predicts: leaves under only the right-side wheels changing the torque the controller must send, regen braking overpowering the battery in extreme stops, a hacked-together Jenkins boot script that crashed half the time once hundreds of robots needed booting every morning.

The moat is data nobody else has: 10 billion deliveries, including where human dashers actually drop food at apartment complexes, the "first and last 100 feet" that Google Maps doesn't know. On the AI side, Ask DoorDash is changing behavior: 50% of restaurant trajectories end in orders from never-tried restaurants, and grocery baskets run 40% larger. Fang notes there's now more agent traffic on the web than human traffic, and DoorDash's internal AI spend grew 20x from January to June before flattening as the DashBench benchmark forced ROI discipline. Tang's contrarian prediction: in ten years DoorDash will have more human dashers, not fewer; autonomy makes delivery cheaper, demand surges, and today's 9 million dashers can't 5-10x anyway.

要点：从用例出发倒推，而不是从技术出发。DoorDash 之所以自研配送机器人，正是因为其他所有人都是先造技术、再回头找问题。

DoorDash 联合创始人 Andy Fang 和 Stanley Tang 从 2018 年起就低调运作机器人项目，起点只是"我加半个工程师"的秘密小分队。在与从人行道机器人到 robotaxi 的各路玩家合作多年后，他们的结论是两头都不合身：时速 2 到 3 英里的人行道机器人跑不了平均 3 到 5 英里的配送，而 4000 磅重的 robotaxi 送两个墨西哥卷饼纯属杀鸡用牛刀，何况食物自己走不了最后一个街区。于是他们造出了 Dot：300 磅重、时速 20 英里、体积只有汽车十分之一的 L4 机器人，既能上马路也能走自行车道，已在 Phoenix 送货两年。

Tang 的招人话术很直接："你是想留在博士实验室做原型和 demo，还是想做能真正落地到现实世界的东西？"而现实世界的毒打是任何 demo 都预料不到的：只有右侧车轮压到落叶时控制器扭矩输出就得跟着变；极端刹车时动能回收系统会反过来冲垮电池；当年随手拼凑的 Jenkins 启动脚本，等到每天早上要启动几百台机器人时一半时间会崩掉。

护城河是别人没有的数据：100 亿次配送，包括人类骑手在公寓楼群里实际把餐放在哪的"最初和最后 100 英尺"数据，这些 Google Maps 里根本没有。AI 这边，Ask DoorDash 正在改变用户行为：餐厅场景下 50% 的交互轨迹最终下单了从没吃过的新餐厅，杂货购物篮平均大了 40%。Fang 提到如今网络上的 agent 流量已超过人类流量，而 DoorDash 内部 AI 开销从 1 月到 6 月涨了 20 倍，之后靠 DashBench 基准强制核算 ROI 才趋平。Tang 的反直觉预测：十年后 DoorDash 的人类骑手会更多而不是更少；自动驾驶让配送更便宜，需求随之暴涨，而如今 900 万骑手的规模本来也不可能再翻 5 到 10 倍。