June 10, 2026

Andrej Karpathy (deep learning researcher, ex-Tesla / OpenAI) calls Claude Fable 5 a genuine step-change on par with Claude 4.5 — especially for long problem-solving sessions on hard problems, where you can hand it far more ambitious tasks and it "just gets it" and runs. His bigger point: as working software comes "out on a tap," Jevon's paradox kicks in and his own demand for software is exploding — explainers, visualizers, dashboards, bespoke single-use apps, 10x'd test suites, auto-optimized code. He cautions the model still has quirks and the launch safeguards are tuned a little too trigger-happy.

深度学习研究者、前 Tesla / OpenAI 的 Andrej Karpathy 认为 Claude Fable 5 是一次真正意义上的跨越，分量不亚于 Claude 4.5，尤其是在面对难题的长时间求解场景里，你可以交给它野心大得多的任务，它能"领会"你的意图然后自己跑下去。他更深一层的观点是：当可用的软件像拧开水龙头一样源源不断时，Jevon 悖论开始发挥作用，他自己对软件的需求反而暴涨——讲解器、可视化、dashboard、为单一用途定制的小 app、把测试套件扩到 10 倍、自动优化代码。他也提醒模型仍有一些怪癖，且发布时的安全护栏调得有点过于敏感。

Boris Cherny (works on Claude Code at Anthropic) says Fable 5 is the biggest step up he has felt since Opus 4.5 — Claude has gone from a coding agent to a "thought and design partner" with judgment, taste, and "dimensionality" that previous models lacked. The detail he highlights: Fable debugs methodically on its own, taking measurements, adding logs, then verifying it truly fixed the issue before declaring victory — behavior nothing in Claude Code's prompting asked for, just "part of its personality" and a real "big model smell." He also amplifies why building self-verification loops lets a model run far longer and land closer to what you intended without you constantly checking in.

Anthropic 负责 Claude Code 的 Boris Cherny 说，Fable 5 是他自 Opus 4.5 以来感受到的最大跃升——Claude 从一个写代码的 agent 变成了"思考与设计的搭档"，拥有以往模型没有的判断力、品味和"维度感"。他特别提到的一个细节是：Fable debug 时会自己有条不紊地推进，先做测量、加日志，确认问题真的被修好了才宣布完成——而 Claude Code 的 prompt 里并没有要求它这么做，这纯粹是"它性格的一部分"，是一种真切的"大模型味道"。他还转发强调，为什么搭建自我验证（self-verification）的循环能让模型跑得更久、更贴近你的本意，而不需要你时时盯着。

Alex Albert (Research at Anthropic) ranks Fable 5 alongside Opus 3, Sonnet 3.5, and Opus 4.5 as one of the rare launches that change how people use models — it stopped feeling like a tool he directs and started feeling like something he collaborates with. With usage limits reset across products, he shares four tips: give it bigger, more ambitious tasks than before; default to xhigh/high effort (med for fast interactive sessions); rework skills and CLAUDE.md files, since instructions written for older models anchor Fable to stale patterns; and shift from giving tasks to giving objectives — describe what "done" looks like and how to verify it, then let Fable find the path (with /loop and /goal built for exactly this).

Anthropic 研究员 Alex Albert 把 Fable 5 与 Opus 3、Sonnet 3.5、Opus 4.5 并列，视为少数几次"改变人们使用模型方式"的发布——它不再像一个被他指挥的工具，而更像一个协作对象。借着各产品线用量额度重置的契机，他给出四条建议：交给它比以前更大、更有野心的任务；默认用 xhigh/high 的 effort（追求快速交互时再用 med）；重写你的 skills 和 CLAUDE.md，因为为旧模型写的指令会把 Fable 锚定在过时的套路上；以及从"派任务"转向"给目标"——描述清楚"完成"是什么样、如何验证，然后让 Fable 自己找路径（/loop 和 /goal 正是为此而生）。

Thariq (also works on Claude Code at Anthropic) frames Fable as a step-change with one message for users: it's time to be more ambitious. He's promising a series of posts on how the model has reshaped his team's work.

同样在 Anthropic 做 Claude Code 的 Thariq 把 Fable 定性为一次跨越，并向用户传递了一句话：是时候更有野心了。他预告会发一系列帖子，讲这个模型如何重塑了团队的工作方式。

Claude (official Anthropic account) announced Claude Fable 5 is available everywhere today. Alongside it, for a small group of cyber defenders and critical-infrastructure providers, Anthropic launched Claude Mythos 5 — the same underlying model as Fable 5 but with safeguards lifted in some areas. Mythos 5 is restricted to "Glasswing partners" for now, with stated plans to expand the trusted-access program toward defensive cybersecurity and biomedical research.

Anthropic 官方账号 Claude 宣布 Claude Fable 5 今日全面开放。与此同时，面向一小群网络安全防御者和关键基础设施提供方，Anthropic 推出了 Claude Mythos 5——它与 Fable 5 共用同一个底层模型，但在部分领域解除了安全护栏。Mythos 5 目前仅限 "Glasswing 合作伙伴" 使用，官方表示计划把这套可信访问计划逐步扩展到防御性网络安全和生物医学研究。

Swyx (DX / Cognition, AI Engineer) flags some "alpha": while Fable is temporarily not pay-per-use, run "review my code for issues" in Claude Code on it — and be ready to be horrified at what you shipped to prod without a "Fable Check" first. He also notes it was only 34 days from signing the Nvidia deal to shipping a Mythos-class model to GA, crediting the speed to building on the Nvidia stack.

DX / Cognition 的 AI Engineer Swyx 抛出一条"alpha"：趁 Fable 暂时还不是按量计费，在 Claude Code 里让它"review my code for issues"——然后准备好被吓到，因为你会发现自己在没做过"Fable Check"之前就把多少东西推上了生产。他还提到，从签下 Nvidia 这笔合作到把一个 Mythos 级别的模型推到正式可用（GA），只用了 34 天，并把这种速度归功于构建在 Nvidia 技术栈之上。

Guillermo Rauch (Vercel CEO) shipped a Vercel CLI update that lets you create AI Gateway API keys, cap their spend with a `--budget` flag, and set a `--refresh-period` for the quota — "virtual credit cards for AI tokens." He also shared a playful workflow where Opus wrote a VM and then Mythos verified it.

Vercel CEO Guillermo Rauch 发布了 Vercel CLI 的更新：你可以创建 AI Gateway 的 API key、用 `--budget` 给它们设定开销上限，并用 `--refresh-period` 设定额度的刷新周期——他形容这是"给 AI token 用的虚拟信用卡"。他还分享了一个有趣的工作流：先让 Opus 写一个 VM，再让 Mythos 去验证它。

Aaron Levie (Box CEO) argues the day's capability jump proves AI progress isn't slowing, and makes a bull case for applied-AI companies: there is still an enormous gulf between raw model capability and what it takes to apply models to specific corporate workflows. The durable moat, he says, is the "unglamorous work" — arranging a company's private reality so a model can act on it, handing it the tools to act, and doing the change-management and forward-deployed-engineering work that "never ends." Two things can be true at once: frontier labs keep growing massively, and a vast ecosystem of vertical applied-AI companies, new integrators, and infra providers emerges around them.

Box CEO Aaron Levie 认为，这一天的能力跃升正好反驳了"AI 进步在放缓"的说法，并为应用层 AI 公司给出一套看多的论证：模型的原始能力与把它真正落到具体企业工作流之间，仍隔着一道巨大的鸿沟。他说真正能形成护城河的，是那些"不光鲜的活儿"——把一家公司的私有现实组织好让模型能在其上行动、把工具交到它手里，以及那些"永远做不完"的变更管理和 FDE（forward-deployed engineering）工作。两件事可以同时为真：前沿大模型实验室会继续高速增长，而围绕它们也会涌现出庞大的垂直应用层 AI 公司、新型系统集成商和基础设施提供方的生态。

Zara Zhang (builder) makes a sharp adoption point: the barrier for non-technical people using coding agents was never the interface — chatting is the easiest UI ever invented — it's that they don't know what to ask for. A blank chat box assumes you already know what's possible, and most people don't. She praises Town's onboarding, where the agent proactively suggests workflows and things it can take off your plate rather than waiting for instructions.

Builder Zara Zhang 提出一个犀利的采用洞察：非技术人群用 coding agent 的障碍，从来都不是界面——聊天本身就是有史以来最简单的 UI——而是他们不知道该提什么需求。一个空白的对话框默认你已经知道有哪些可能性，而大多数人并不知道。她称赞了 Town 的 onboarding 体验：agent 会主动建议工作流、主动提出能替你分担的事情，而不是干等着你下指令。

Nikunj Kothari (FPV Ventures partner) shows what one-shotting looks like now: after listening to an Invest Like the Best episode on S-curves, he had Fable one-shot a website cataloging the major S-curves of the last 200 years, their inflection points, and commentary on whether each was a bubble — live at escurves.com. His recipe: dump the transcript into Claude's research mode, have it research all the S-curves and structure the sections, then generate a single Claude Code prompt to one-shot the build.

FPV Ventures 合伙人 Nikunj Kothari 展示了如今"一把过"（one-shot）能做到什么程度：他听完 Invest Like the Best 一期讲 S 曲线的节目后，让 Fable 一把生成了一个网站，梳理过去 200 年里所有重大的 S 曲线、它们的拐点，以及对每一个是否是泡沫的点评——网站已上线在 escurves.com。他的做法是：把节目文字稿丢进 Claude 的 research 模式，让它研究所有这些 S 曲线并把章节结构搭好，再生成一条 Claude Code 的 prompt 一把把网站建出来。

Thibault Sottiaux (works on Codex and ChatGPT at OpenAI) describes orchestrating Codex "like an orchestra — one /goal at a time," and is canvassing whether people use Codex's /goal occasionally or as their main way to get work done.

OpenAI 负责 Codex 和 ChatGPT 的 Thibault Sottiaux 把指挥 Codex 形容为"像指挥一支管弦乐队——一次一个 /goal"，并在征询大家：你们是偶尔用一下 Codex 的 /goal，还是把它当成完成工作的主要方式。

Josh Woodward (VP at Google Labs / Gemini), echoing the day's dominant theme, predicts the demand for software is going to be "off the charts."

Google Labs / Gemini 的副总裁 Josh Woodward 也呼应了这一天的主旋律，预测对软件的需求将"爆表"。

The Takeaway: As AI automates expert work, it doesn't erase jobs — it floods the zone with output that's "close but not quite right," which makes human experts more in demand than ever.

Dan Shipper, CEO of Every — a media-and-software company about as agent-native as it gets — unpacks his 8,000-word essay "After Automation." The paradox driving it: everyone at Every uses Claude Code and Codex daily, "swing a stick in our Slack and you're as likely to hit a human as an agent," and yet the company grew from 4 people to 30 and is still hiring. His explanation is that AI makes yesterday's expert competence cheap: anyone with a prompt can now produce code, essays, or designs. But because models are trained on past outputs, what they generate by default all looks similar and is close but not quite right — flooding organizations with abundant, devalued, almost-good work. That paradoxically raises demand for experts who can shepherd it across the finish line and build the systems (repo rules, review guidelines) that filter it.

His sharpest line on why the human gap persists: "The further away an agent gets from a human, the less valuable it is." Agents are getting good at autonomy (completing tasks) but lack agency — the self-motivated wants even a small child has. No matter how powerful it becomes, the model looks back at you to decide what matters, and what matters keeps changing precisely because AI keeps changing the world. On the wave of "we fired X thousand people" layoff posts, he's skeptical: companies that cut and blame AI are often just not doing well; AI reshapes workflows and org structure, which is messier than the clean "AI took the jobs" story.

His bottom line: "If you ride the models, you're going to be okay. You're going to have a job. You're going to do great work, and you don't have to worry." A bonus process nugget — he wrote the essay by monologuing into a doc each morning, asking Claude "what am I really trying to say?", and having Codex turn each draft into an audio reading he'd listen to on his commute.

核心要点： 当 AI 把专家级工作自动化时，它并不会消灭岗位——它会让大量"接近正确、但又不太对"的产出充斥进来，反而让人类专家比以往任何时候都更抢手。

Every 的 CEO Dan Shipper——这家媒体兼软件公司大概是 agent 原生程度最高的公司之一——拆解了他那篇 8000 字的长文《After Automation》。驱动这篇文章的悖论是：Every 的每个人每天都在用 Claude Code 和 Codex，"在我们的 Slack 里随手一捞，捞到的是人还是 agent 概率差不多"，然而公司却从 4 个人长到了 30 人，而且还在继续招人。他的解释是：AI 让昨天的专家能力变得廉价——任何人只要会写 prompt，现在都能产出代码、文章或设计。但由于模型是在过去的产出上训练的，它们默认生成的东西看上去都很像，且接近正确但又不太对——于是组织里涌进大量"过得去、却被贬值"的半成品。这反而抬高了对专家的需求：他们能把这些东西推过终点线，并搭建起过滤它们的系统（仓库规则、review 指南）。

关于为什么人与 agent 之间的鸿沟依然存在，他最锋利的一句话是："agent 离人越远，它就越没价值。" agent 在自主性（完成任务）上越来越强，但缺乏 agency——那种连小孩都有的、出于自我意愿想做点什么的冲动。无论它变得多强大，模型最终都会回过头来看着你，等你来决定什么才重要；而恰恰因为 AI 不断改变世界，"什么才重要"也在不断变化。对于那一波"我们裁掉了几千人"的帖子，他持怀疑态度：裁员然后甩锅给 AI 的公司，往往本身就经营不善；AI 重塑的是工作流和组织结构，这比"AI 抢走了工作"这套干净叙事要混乱得多。

他的结论是："只要你跟着模型走（ride the models），你就不会有事。你会有工作，会做出很棒的成果，不用担心。"额外的一个流程小技巧——他写这篇长文的方式，是每天早上对着文档自言自语口述一遍，问 Claude"我到底想说什么？"，再让 Codex 把每一版草稿转成一段音频朗读，供他通勤路上听。