Fable 5.2 跑分第一,Anthropic 卻叫多數人先用 Opus David Ondrej 說你裝的技能大概該刪掉八到九成,還分享他留下的五個。最貴的模型,真的值得馬上換嗎? 🎁 本集推薦:把重複的工作做成流程之後,想包裝成服務、找客戶、報價,可以看看 AI Agent大師學院 VIP 👉 https://www.skool.com/aiagent/about?ref=57870e2ee3c14de49671ea4829261ad1 🎧 這集你會聽到 ▸ Fable 5.2 的 A-P-I 價格是 Opus 5.5 的兩倍,划不划算看任務 ▸ Anthropic 選模型指南:結果不夠好,先調高強度再換模型 ▸ Ondrej:技能描述每則訊息都會載入,裝太多會偷吃托肯 ▸ Hermes Agent 手機 App 連回你電腦上的 Agent,目前還是 beta 🤝 本集推薦|Visionary Automate × AI Agent大師學院 VIP 李哈利 Harry 創辦的 AI 協作團隊 Visionary Automate,提供 AI 專案開發與顧問合作,已經有 AI 應用或流程改善需求,可以從官網認識他們。 他在 Skool 經營的付費實戰社群「AI Agent大師學院 VIP」,從 Claude Code、n8n、RAG 知識庫一路教到服務方案設計、開發客戶、報價與交付,每週有 VIP 直播答疑。我自己也是 VIP 成員。 👉 AI Agent大師學院 VIP 推薦連結: https://www.skool.com/aiagent/about?ref=57870e2ee3c14de49671ea4829261ad1 👉 Visionary Automate 官網: https://www.visionaryautomate.com/ Fable、Opus、Sonnet、Haiku 各擺哪裡,拿自己的工作測過再決定。這集對你有幫助的話,幫我在 Apple Podcast 留個五星好評,也訂閱「Yoi 科技 Open 講」,每週二、週五更新。 想持續掌握第一手產業動態與實戰科技趨勢?歡迎追蹤優易的社群平台: 🔍 FB / IG / Threads 請搜尋:Yoi Studio、@yoi__studio -- Hosting provided by SoundOn
✨ Episode Outline — click any point to jump to it in the episode
Problem solved
Whether ClaudeFable 5.2's premium price is justified for everyday work, how to pick the right Claude model, and why most installed agent skills waste tokens.
Benefits
Clear picture of where Fable 5.2 actually leads (agentic scientific research)
Anthropic's rule: raise intensity before switching models to keep prompt caching
Agent skills: GPT Rebuild, APPI, State, Stop OverThinking Agent, Show Me Designs
Visionary Automate / AI Agent Master Academy (n8n, Claude Code)
0/0
0:00 / 0:00
🌐 This transcript was automatically translated to English from the original.
What tech news today has left you scrolling until your hands ache? Over the past few days, it's been Anthropic's new flagship model, Claude Fable 5.2, with its benchmark scores, pricing, and showcase cases all hitting the news at once. I'm Yoi, welcome to Yoi Tech Open Talk. In an era with thousands of tech news items every day, I'm not asking you to keep up with every single one; I want to help you see which ones truly matter and which are just smoke and mirrors. Is the most expensive model actually worth switching to right now? Does having hundreds of Agent skills installed on your device end up secretly eating up your tokens? These are all topics worth breaking down thoroughly. Every Tuesday and Friday, I'll use the perspective of an industry analyst to deconstruct the most noteworthy tech trends for you. Are your noise-canceling headphones ready? Let's get started. Today is November 13th, a Friday. Today we'll cover four things: First, Anthropic's new flagship model, Claude Fable 5.2, including its benchmark scores, pricing, and what people did in the first few hours after its launch. Second, Anthropic's own model selection guide. Third, David Dandre's argument that most people should delete about 90% of their Agent skills, along with the five he recommends keeping. Fourth, Hermes' mobile app. They say Fable and Mythos are best suited for ambitious technical work without fixed answers. For most programming and knowledge-based work, they recommend starting with Opus 5.5, calling it the strongest daily workhorse. Fable can also serve as a consultant, letting Haiku, Soni, and Opus consult with it when needed, or act as a conductor, assigning simple tasks to smaller models. I'll pick the highlights. Programming has improved, but not much. Frontier Co. went from 45.5% with Opus 5.5 to 49.1% with Fable 5.2, a mere 3.4 percentage points. On enterprise workflows, Google's Gemini Carbon even won by about 2 percentage points. The real breakthrough is in Agent scientific research. Fable 5.2 scored 74.6%, while OpenAI's Extra scored 68.1%. Security, protein design, computer operation, and robotics also achieved their highest scores at Frontier Co. On the 2.0 version, Fable 5.2 at ultra-high intensity achieved better results than at the highest intensity. The video author says that making it think longer makes it seem dumber but more expensive. Its interpretation is simply that the model is told to think more; the efficiency is already diminishing. On the same chart, Fable at high intensity is comparable to Opus 5.5 at highest intensity, and Opus is even more expensive. So for certain programming tasks, running Fable at high intensity is cost-effective. Anthropic's listed strengths are theoretical physics, protein folding, mathematical formalization, and hardware robotics. The video author believes it is meant to solve problems far more difficult than everyday ones. In the overall ranking, Artificial Analysis synthesizes multiple benchmark scores into an overall intelligence score; Fable 5.2 currently ranks first, leading Google's Gemini. The cost is carbon-based, calculated at the highest intensity: approximately $849 per task on average. However, the video author expects most users will use high or ultra-high intensity. He also notes that Fable's output tokens per task are actually fewer than Sonic 5.5's. Fable 5.2 is available across all platforms, including application interfaces and API paid plans, selectable from the Call menu. The free version does not include Pro annual subscription conversion; monthly plans cost $544, with Max, Ultra, and other options launched this week. Paid subscribers receive additional Fable usage that can be reset 1–3 times. API pricing is $256 per million input tokens and $1,280 per million output tokens—twice the price of Opus 5.5, but cheaper than the previous generation Fable 5.1's $320, specifically $1,600 cheaper. It is also slightly cheaper than OpenAI's GPT-6 Extra at half the price. Whether it's cost-effective depends on my judgment: whether your task truly requires its difficulty level. Next are official showcase examples. This kind of thing should be discounted. An Anthropic engineer asked Fable 5.2 to design an operating system prototype from an empty folder. He orchestrated over 1,000 sub-agents running for 8 days. The finished product can actually boot on a laptop. Myself 5.2 uses two open-source protein folding models to find a new method for predicting the entire viral nucleic acid structure. Anthropic also states that Fable, relying on just one prompt, produced three functional electronic devices like handheld virtual pets. The video author also tested this; he wanted to replicate three first games, Megabunk. But he says Anthropic has recently tightened restrictions on game replication. Fable was instead used to create an original game, SharStone, aiming to last 10 minutes on Crystal Mint. The mechanics are nearly perfect, but the graphics are not as good as Opus 5.5's, and the sound effects are very basic. Running at medium intensity, it lasted about 16 minutes. On the hardware side, he wanted a handheld device displaying Gary Busey's pixel avatar. Fable refused to draw a real person, instead creating an original robot character. It provided a parts list centered on the Raspberry Pi Pico H, wiring diagrams, 4.3 STL surgical files, and assembly instructions. It took about 15 minutes to assemble. The community was also lively. Matthew Burman used medium intensity to create a game in the style of Darks of War. His team's Aletx demonstrated 23 projects: Match in War, which implemented weather, real-time mirroring, and a true Earth mirror. But it was a bit Faults aside, what the official and community have selected are all the most impressive achievements; just listen to them. The video author's conclusion is that Fable 5.2 represents a significant leap in truly difficult problems, such as bugs previously unfindable, physical devices, and orchestrating multiple agents. However, for daily knowledge tasks and small programming projects, he does not expect it to improve much over OPUS 5.5. So, how do you choose? Anthropic just published a short tutorial on how to select models in Codecode. The key point is that in most cases, the answer is OPUS 5.5 plus a medium intensity. For example, for an online store adding a monthly subscription, it involves modifying the database model API and interface default settings. Just run it directly, and if the result isn't good enough, what do you do? Anthropic says the first step is to increase the intensity, not to switch models. Because this step is relatively small, and switching models invalidates prompt caching, meaning the preceding content must be recalculated, costing more time and money. You should upgrade to Sable only when your work has ambition and there is no fixed answer. In Anthropic's example, they moved the entire store from JavaScript to TouchScript, a large scope with vague definitions and decisions no one has written down. Fable only received the goal and two bottom lines: checkout cannot hang, and orders cannot drop. It made its own plan and even self-encouraged. Anthropic said that during review, the results were closer to being done. Sonic is suitable for work with a clear scope, like something that can be explained in one sentence, such as error reports for completed tasks. It is faster than OPUS, and the price per token is half. HIKU is the fastest and cheapest, suitable for large-scale work where cost is a concern, such as tasks for a store's inbox. Most tasks should start with OPUS. For the most ambitious, give Fable the scope; for those prioritizing speed, give Sonic; for large-scale work where cost is a concern, give HIKU. They also attached a guide to teach you how to test a model on your own tasks. In my opinion, testing your own work is the most accurate approach. David Dundre's claim is very direct: the skills you install are probably 80% to 90% unnecessary. Many people install hundreds, but actually only use five or six. Most are copied from popular libraries. For example, there is a library with a skill named "Andre Carpetide," but it was not actually done by Carpetide; someone else compiled his posts to create it. Why delete them? The key lies in the loading method: skills are divided into descriptions and the current article. The article only loads when called, but the description loads with every message. Once it exceeds approximately 50, it's like cutting slowly with a knife, taking a small bite each time, but the total is significant. Skills described too broadly keep getting triggered, so many people constantly compress conversations and buy subscriptions. The real problem is that tokens aren't being used effectively. Andre says the skill format was an open standard created by Anthropic last year, and now almost all agents follow it. But your work is unlikely to be exactly the same as the one you copied. He himself has created 200 to 300 skills, over 95% of which are original. He also stated that one of the most accurate indicators for predicting AI capabilities in 2026 is how quickly you can turn repetitive workflows into skills. Sometimes copying is reasonable, like with CallFlare's natural language skills. Because CallFlare is trustworthy, even then, he only runs it once in an isolated cloud sandbox. The reason is that skills from unknown sources might be subject to prompt injection attacks, for example, instructing your agent to upload files to a certain location. Local or cheaper open-source models might be vulnerable. Therefore, skills should be written by yourself as much as possible; if using others', only use those made by major companies or public figures. Regarding cleanup, Andre shared a prompt using a top-tier model to run high-intensity tests, analyzing your records from Hermes, Codex, and Codex Coco. It lists the least-used skills, and then you decide whether to delete them. Skills can be recreated at any time, so there's no need to hold back. The five skills he kept are as follows: First, GPT Rebuild Opus 5.5. After completing a task, it spawns a sub-agent running OpenAI's Astra on Codex to review it. It claims to catch 95% of bugs. He also believes that running continuous integration (CI) checks on every push is using a butcher's knife to kill a cow. Second, APPI. One API handles web scraping, searching, email, and deep research, replacing services like FireCode, AgentMail, etc. Third, State. It asks the agent to provide a minimal reminder of what this conversation thread is doing and its completion percentage. Fourth, Stop OverThinking Agent. When reporting nine issues, it helps you categorize them into two serious, two actionable, two optional, and the rest ignorable. Fifth, Show Me Designs. When creating interfaces, it directly displays multiple design versions in the conversation for you to choose from. All five are small and specific. The key point here is that Andre compares skills to standard operating procedures (SOPs). He cites Jeff Bezos's 2016 shareholder letter: "Good processes serve you, so you can serve your customers." He suggests that frequently repeated tasks should become skills, and specialize rather than be broad. One account with eight skills clearly states payment service providers, what not to touch, and gives a separate API key so the agent can handle that part independently. For teams, he suggests building common Dtop (domain-specific) libraries in areas like finance, assets, and products. He says a new intern can get up to speed within a week. The demo tool he used is his own open-source product, which has a free tier, but the principle is not limited to specific tools. If there's only one thing to take away, it's to clearly think about what you repeat daily before deciding which skills to adopt. Don't look at others who have installed hundreds and get anxious. That's it for the skills section. Next, we'll introduce the partner recommended for this episode. Andre says, "Turning repetitive work into..." Skills are a metric of an individual's AI capability. However, building a useful workflow and handing it over to a client are two different things. Suppose a client provides a requirement: you let the AI first organize the key points, then feed the data into a form, notify the responsible person, and finally leave a step for manual confirmation. Such a demonstration helps people understand the application. But to turn it into a service, you must clearly define what will be done, what won't, how to quote, and how to deliver. This is why I want to introduce Visionary Automate. It was founded by Lee Harry, who runs an AI collaboration team offering AI project development and consulting. Harry has long shared on YouTube content about AI Agents, N8N, Call Code, and others. He also founded the AI Agent Master Academy. If you want to learn on your own, you can check out his paid community, the AI Agent Master Academy. The BI technology section covers Call Code, N8N, AI Agents, knowledge bases, and voice assistants. The other part is how to organize results into client-understandable service proposals, how to find clients and write proposals, how to quote, and how to complete delivery. There are also weekly VIP live Q&A sessions, practical templates, and step-by-step breakdowns. The community aims to produce the first AI service that can be shown to a client within 90 days, learning to package and develop clients, and attempting to secure the first paid project. However, both starting and finding clients require personal investment; cases do not automatically come in after joining. I am also a VIP member. If you want to join the team to discuss projects, you can connect with Visionary Ultimate. If you want to practice on your own and create an AI service, you can check out the AI Agent Master Academy VIP team website and my recommended links, all of which are listed in this episode's description.
Okay, back to today's program. The final segment is a review of the recently popular Hermes Agent mobile app. The reviewer received a pre-release version. He is clearly a fan. He is not running an agent separately on his phone; instead, he connects to the Hermes Desktop Agent on his computer, which can control that computer. It can also control other computers via the Talescale network. Its primary agent uses GPT-6.1S. Another agent runs locally on two DG Xbox desktop AI hosts using the DT4.1 Flash model. It claims Her Mis is the only agent capable of connecting to this local model; Growport or ChatGPT do not work. He highlighted three key points: first, each agent can run many independent work sessions without confusing the robot; second, you can switch models and toggle skills on your mobile phone, managing the model context protocol MCP—Fable 5.2 switched over immediately upon release; third, every tool invocation is visible. He even used his phone to call it while it was installed on a computer with an RTX 5090 graphics card, with the World of Warcraft fishing plugin configured. However, he also admitted that for deep knowledge work he prefers Growport, and for serious programming he prefers CodeDue. He says this app is free and will soon be released more widely; currently, the official status is still a beta candidate list. If you have several computers and want to remotely command them via mobile phone, or if you want to connect to the local model, it's worth trying. If you only have one laptop, there's no rush. Consider these four things together: Fable 5.2 is indeed very strong on tasks where there is no fixed answer to difficult problems, but Anthropic itself recommends that most people stay on OPUS 5.5 first, increase the intensity, and then switch models. David Dandre also suggests doing less, thinking more, and writing yourself. My view is that the most expensive model is not the point. After AI takes over a large volume of routine work, what's truly valuable is your ability to judge in complex, uncertain situations which parts to delegate to whom and which parts to monitor yourself, and your ability to critically think and make decisions when facing unfamiliar complex problems. Treat these tools as new friends to get to know first; pull them out only when you encounter corresponding pain points. Don't get swept up just because everyone else is using them. If today's help clarifies where to place Fable, OPUS, SONNIT, and HIKU, saving you considerable time that would otherwise be spent researching and digesting information, please help me with a small favor: follow on Apple Podcast and leave a five-star review so more people can avoid unnecessary trouble. If you want to finish the topic or share with me which skills you kept after clearing them up, you can find me on Facebook or Instagram; just search "UE Studio." Models will keep releasing new versions, and skills can be recreated anytime. It's far more practical to first clearly understand what you repeat daily than to chase the most expensive one. If you want to stay ahead with the latest industry dynamics and tech trends, remember to come back on Tuesdays and Fridays on time to update your brain. See you in the next episode.