X / Twitter
Deep learning researcher Andrej Karpathy
Karpathy gave unusually high praise to the release of Claude Fable 5. He noted that Fable 5 and Mythos share the same underlying model, with safety guardrails added, and that it is broadly SOTA across benchmarks by a clear margin. But he emphasized the more important qualitative experience: this is a leap worthy of a major version number, comparable in scale to Claude 4.5 last November, particularly in sustained work on extremely difficult problems. He said users can give it far more ambitious tasks than before; the model truly understands intent and advances independently. For the first time, he even felt tempted not to look at the code at all, though he warned against doing so in production. He also noted remaining quirks and overly sensitive safeguards at launch, hoping they would be tuned. He senses Jevons paradox in action: with software flowing like tap water, his own demand for software has grown dramatically. Explainers, visualization tools, dashboards, one-off custom apps such as an entire wandb tailored to a single project, test suites expanded 10-fold, automatic code optimization, and large research projects with custom HTML result pages—all are now things he can ask for.
https://x.com/karpathy/status/2064409694761054332
AI Engineer conference and Latent Space host Swyx
Swyx offered a practical tip: before Fable switches to usage-based billing, have Claude Code run “review my code for issues” on your codebase. He said the results would make you horrified that you had ever dared push anything to production without a Fable Check. He also laid out a timeline: only 34 days elapsed from signing the deal to making a Mythos-class model generally available worldwide. His assessment was that building on NVIDIA’s stack means “you can just get things done.” He also assembled comparison charts across subscription tiers, including what he called a vibe-shift chart; differences among tiers were less dramatic than at the flagship level. Overall, his focus is on reading the release in the context of industry cadence: model iteration speed, compute partnerships, and the pricing window.
https://x.com/swyx/status/2064492823781789969
https://x.com/swyx/status/2064421542503797186
https://x.com/swyx/status/2064415059003248694
Anthropic Claude Code team member Boris Cherny
Boris Cherny said Fable 5 was the biggest model leap he had felt since Opus 4.5 last November. He recalled realizing after Opus 4.5’s release that he had spent weeks coding 100% in the terminal, prompting him to uninstall his IDE. With Fable, he feels Claude has grown from a coding agent into a thinking and design partner in product building. Its judgment, taste, and multidimensional thinking make him comfortable assigning his most complex work to it. He described one moment behind that realization: asked to debug an issue, Fable was the first model he had used that was so methodical and precise, measuring first, adding logs, and verifying the problem was truly fixed before declaring completion. Nothing in Claude Code’s prompt instructed it to do this; it was simply the model’s character, which he called an unprecedented “big model smell.” He also emphasized self-verification loops. In an era of long-running models, self-verification is key to longer runs and results closer to user intent, removing the need to watch Claude constantly. He recommended delba_oliveira’s breakdown of the approach.
https://x.com/bcherny/status/2064469227499929651
https://x.com/bcherny/status/2064431111154053187
https://x.com/bcherny/status/2064426115255730578
OpenAI Codex and ChatGPT team member Thibault Sottiaux
Thibault Sottiaux’s theme over the past two days has been Codex’s /goal feature. He asked whether people use codex /goal occasionally or as their primary way to get work done. His own approach is to “play Codex like conducting a symphony,” running one /goal at a time and moving multiple goals forward in an orderly way. This interestingly echoes Anthropic’s promotion of /loop and /goal: both leading labs are moving the interaction paradigm from “give it a task” toward “give it a goal.”
https://x.com/thsottiaux/status/2064308436133716008
https://x.com/thsottiaux/status/2064307859903447396
AI tutorial and interview creator Peter Yang
Peter Yang shared the full prompt he used to test Fable: build an F-Zero-style antigravity racing game, with a pseudo-3D track using raycasting or mode-7-style scaling, 3 AI opponents, a boost meter that consumes health, a sense of speed at 400-800 km/h conveyed through track distortion and scrolling ground textures, arrow-key steering, shift to boost, and a 3-lap circuit with checkpoints. The requested style was neon cyberpunk: dark skies, glowing track edges, chromatic aberration while boosting, a HUD showing position, lap, speed, and health, and ships banking into turns. He stressed that “feel comes before realism; the sense of speed is the whole game.” The prompt itself is a useful template for game-focused vibe coding. He also offered negative feedback: browser calls made Fable noticeably slower for him. And he asked, “What does big model smell even mean?” poking fun at a phrase Anthropic’s team had been using frequently over the past two days.
https://x.com/petergyang/status/2064550073594446059
https://x.com/petergyang/status/2064577126385459265
https://x.com/petergyang/status/2064563041166090672
Anthropic Claude Code team member Thariq
Thariq called Fable a step change in models and said he hoped it would change how people collaborate with Claude. He previewed a series of articles on how Fable has reshaped his team’s work, with a one-line TLDR: it is time to be more ambitious. This aligns with Boris Cherny’s and Alex Albert’s statements: Anthropic’s internal framing of the release is clearly “change how you use it,” not just “improve performance.” He is currently at Code w/ Claude Tokyo and invited attendees to say hello.
https://x.com/trq212/status/2064437561930682672
https://x.com/trq212/status/2064521202622960058
Vercel CEO Guillermo Rauch
Guillermo Rauch announced a new Vercel CLI capability: directly create AI Gateway API keys, set each key’s spending cap with --budget, and set its allowance refresh interval with --refresh-period. He compared it to “virtual credit cards for AI tokens,” a practical governance tool for engineering organizations distributing model-call allowances to team members or subprojects. He also shared a cross-model collaboration example: having Opus write a VM, then Mythos verify it, demonstrating a workflow in which one model generates and another checks. Together, the updates show Vercel positioning itself at the billing and orchestration layers of AI infrastructure.
https://x.com/rauchg/status/2064551967461114111
https://x.com/rauchg/status/2064419055726215438
Anthropic researcher Alex Albert
Alex Albert announced that Anthropic had reset usage limits across its entire product line and offered four tips for users starting to test Fable. First, give it larger, more ambitious tasks than previous models could handle. Second, default to xhigh or high effort for best performance, using med when speed matters in interactive sessions. Third, rewrite your skills and CLAUDE.md: instructions written for older models anchor Fable to outdated patterns, so let it exercise its own judgment first. Fourth, move from assigning tasks to assigning goals: describe what completion looks like and how to verify it, then let Fable find the path. /loop and /goal are designed for this. He also placed the release in context as someone who had experienced every Anthropic model launch: the genuine step changes were Claude Opus 3, Sonnet 3.5, Opus 4.5, and now Fable 5. His central feeling is that the model has gone from “a tool I direct” to “something I collaborate with.”
https://x.com/alexalbert__/status/2064467657483829441
https://x.com/alexalbert__/status/2064394410004304003
Box CEO Aaron Levie
Aaron Levie strongly recommended an article about application-layer AI companies, calling it essential reading for applied AI founders. The core passage he quoted argued that applications earn their place in “corners models cannot train on” through unglamorous work: organizing enterprises’ private realities into forms models can act on, giving models tools for action, and helping customers change their organizational realities. This translation work is never finished, and companies doing it are difficult to replicate. His own assessment is that a huge gap remains between model capabilities and specific enterprise workflows. Some consists of technology yet to be built; a large portion involves obtaining and formatting the right data; and an even larger part is change management and implementation, such as the FDE model. He believes two things can be true simultaneously: frontier models and labs will continue growing rapidly, while an enormous software and services ecosystem emerges to bring their capabilities into real enterprises, leaving room for new infrastructure providers, applied AI companies in different verticals, and new kinds of systems integrators. On Fable 5, he was equally direct: if you thought AI progress was slowing, this is an immediate rebuttal. Broad capability gains will substantially improve agents in almost every category of knowledge work. He also discussed evaluation methodology: model performance is largely a function of inference-time compute, so compute-normalized benchmarks are the only reasonable direction. The difficulty is that choosing how much compute to use is subjective, model rankings can reverse at different thresholds, and there are nearly infinite ways to set those thresholds.
https://x.com/levie/status/2064569513023328268
https://x.com/levie/status/2064396746953023647
https://x.com/levie/status/2064379199629181139
Y Combinator CEO Garry Tan
Garry Tan’s assessment of Fable 5 was “the biggest model energy I have ever seen.” His practical experience, however, was not entirely smooth: he hit a wall while using Fable 5 to fix GStack issues and posted a long sigh. Excitement in the overall assessment alongside friction in real workflows makes these two posts together a more realistic portrait of users during this launch.
https://x.com/garrytan/status/2064573857911152710
https://x.com/garrytan/status/2064559225859416186
Independent developer Zara Zhang
Zara Zhang offered a counterintuitive view of coding-agent adoption: the bottleneck preventing nontechnical people from using coding agents has never been the interface, because chat is the simplest UI humans have invented. The real obstacle is that they do not know what to ask for. A blank chat box assumes users already know what is possible, and most do not. She cited Town’s onboarding: the agent proactively suggests workflows and things it can take off the user’s hands instead of waiting for instructions, which impressed her. She also previewed an online talk this Friday covering her full vibe-coding process. As a developer without a technical background who has earned 30k GitHub stars, she will discuss where product ideas come from, how she works with coding agents, how to design things that are not AI slop, and why she sees code as a storytelling medium. Registration with the code PREMIUMPASS is free.
https://x.com/zarazhangrui/status/2064587398529606082
https://x.com/zarazhangrui/status/2064486120386379950
FPV Ventures Partner Nikunj Kothari
Nikunj Kothari demonstrated an end-to-end workflow from podcast to finished website. After hearing extensive discussion of S-curves on the latest Invest Like the Best episode, he had Fable generate a website in one pass covering all S-curves over the past 200 years, their inflection points, and the controversies in which each was considered a bubble at the time. It is now live at escurves dot com. He shared the process: open research mode in the Claude app, feed in the entire podcast transcript, ask it to research historical S-curves and organize chapters, then produce a prompt that lets Claude Code complete the work in a single run. This chain—podcast transcript to research to one-shot prompt to live website—is a content production approach with a low barrier to replication for ordinary users.
https://x.com/nikunj/status/2064506504888373758
https://x.com/nikunj/status/2064508462034501997
Every CEO Dan Shipper
Dan Shipper announced Fable’s (Mythos’s) official release and revealed that the Every team had tested it for a week in advance. They immediately released their customary Vibe Check show, presenting the team’s overall assessment of the new model, with the video available on YouTube. Every has consistently been among the fastest media organizations to produce in-depth usage reports on model launch days, making this Vibe Check a priority for anyone seeking to understand Fable’s actual performance.
https://x.com/danshipper/status/2064395167658860859
https://x.com/danshipper/status/2064395458777108707
https://x.com/danshipper/status/2064398724495737180
Anthropic’s official Claude account
Anthropic announced the release’s two-model structure: Claude Fable 5 is available across all channels starting today, while Claude Mythos 5 is limited to Glasswing partners. Mythos 5 and Fable 5 share the same underlying model, with Mythos removing safeguards in certain areas. A small group of cyber defenders and critical infrastructure providers received the initial access to Mythos 5. Anthropic said it plans to broaden access through a wider trusted access program focused on defensive cybersecurity and biomedical research. This “same model, tiered safeguards, trusted access” release structure is a new template for how frontier labs balance capability and safety.
https://x.com/claudeai/status/2064394160522559632
https://x.com/claudeai/status/2064394158056386684
https://x.com/claudeai/status/2064394159318782217
Podcasts
AI & I by Every — We Automated Everything With AI and Tripled Our Headcount
Key takeaway: The more is automated, the more work people have. As long as you “ride the models” and keep up with new ones, you will not become unemployed.
This time, Every CEO Dan Shipper was interviewed by his own COO Brandon about his newly published 8000-character essay After Automation. Every is an ideal case for examining the question: the company is so AI-native that “if you wave a hand in Slack, you are about equally likely to hit a person or an agent.” Everyone uses Claude Code and Codex daily, yet the team has grown from 4 people in the GPT-3 era to 30 today and is still hiring. On one side, Dario says half of entry-level white-collar roles may disappear, and Citadel’s Ken Griffin is stunned by AI producing PhD-level financial analysis. On the other, the most automated companies keep expanding. This paradox deserves examination.
Dan’s central argument has three steps. First, AI makes “yesterday’s expertise” cheap: models trained on all human output let anyone write code, build apps, and create reports with a single prompt. Second, widespread adoption of cheap capabilities produces a flood of work that is “almost right but not quite.” Even operations staff at Every submit PRs, but senior engineer Willie often “thinks the idea is good, then rewrites the whole thing.” Third, this actually raises demand for experts: either to build systems—repository rules, review processes, editorial standards—that shepherd passable output into something truly useful, or to use the tools for previously impossible work, such as colleague Kieran independently building a complete inbox product in a month or two.
His most important line was: “The further an agent is from a human, the less valuable it is.” He thinks the word agent is itself misleading: its original meaning is acting on another’s behalf. These tools will become increasingly autonomous in executing tasks, but that is entirely different from the agency of “having things of their own they want to do.” The industry’s economic incentives push models toward greater obedience, not greater independent will. His definition of AGI is also pragmatic: an agent you never turn off because it is economically worthwhile to let it keep generating tokens and working. Even then, it will still ask, “What is important to do next?” And because AI is rapidly changing the world, deciding what matters requires even more continuous human judgment.
Dan was skeptical of the news that ClickUp’s CEO had prominently announced a one-fifth staff reduction: struggling companies already lay people off, and blaming AI is simply a more convenient narrative. He also observed that many companies that replaced customer service staff with AI were asking them to return two months later, because callers kept trying to establish “are you a robot?” and insisting on a human. His closing advice was clear: “If you ride the models, you’ll be fine. You’ll have a job, you’ll make great work, and you don’t need to worry.” He also shared his workflow for writing the essay: each morning he dictated the full argument into a proof document, had Claude help clarify “what am I really trying to say,” then had Codex turn the latest draft into a podcast so he could listen for problems during his commute.
https://www.youtube.com/playlist?list=PLuMcoKK9mKgHtW_o9h5sGO2vXrffKHwJL