The Every team, including Dan Shipper, Kieran Klaassen, Katie Parrott, and Mike Taylor, gives a hands-on "vibe check" of Claude's Fable 5.1 model, evaluating its capabilities and sharing a companion prompt library for working with it. # Channel Identity Observations: "FABLE 5.1 IS A BEAST" ## Observable Presentation Format - **Live-streamed panel review with named team members.** The video opens mid-stream with technical setup ("We are resolving some audio and video issues") and features multiple hosts: "With us is Kieran Klas. With us is Katie Parrott in her 'Every' cap" and a "Mike Taylor, our head of valuation." - **Day-of-release reaction format.** "Fable 5.1 has just been released. We've been testing it for probably about a week... We have your signature zero-day... verification, just like you've come to expect from us." - **Live demo-heavy structure.** Hosts share screens showing builds (a collaborative editor, dashboards, PowerPoint decks, 3D scenes, an "AI town"), with casual interruptions and banter ("Is everyone drinking enough water?"). - **A live audience is present and addressed.** "We have about 2,287 and the number is growing rapidly. So, if you're here, welcome," and viewer questions are read aloud ("B Mask Maker asks, 'Katie, do you still prefer Codeex...'"). - **A guest from the model vendor appears.** "We have a special guest today—Alex Albert from Anthropic... a researcher and product manager for this model." ## Explicit Self-Description Claims - **The channel is "Every" and positions itself as an essential AI subscription.** Repeated verbatim: "We—Every, we—are the only subscription you need to stay on the cutting edge of AI." - **Claims early access and day-one testing as a differentiator.** "When new models come out, we get them early and conduct a test on the day of their release." - **Claims a proprietary, multi-dimensional review methodology.** "We have our 'craving to use' test... do we gravitate towards it, and why?" plus sections on coding, writing, "intellectual work," and a written "vibe check" at "every.to/vibe-check/fable-5-1." - **Claims internal tooling and professional testing infrastructure.** "Mike created this to help us conduct an internal evaluation of this model... We have been doing this for a long time, professionally." - **Subscriber-only programming is offered.** "On Friday, we're doing live streams for subscribers only. We call them camps... You will get a chance to ask questions, meet the team." ## Intended Audience (as Addressed) - **AI practitioners and professionals making tool choices.** The hosts frame recommendations around switching decisions: "if you're using Opus 5, try Fable 5.1," and discuss cost-per-token and workflow integration. - **Non-expert "vibe coders" and knowledge workers are explicitly included in scope.** "This is the first model that I've used where we're really getting closer to truly delegating intellectual work or helping those who are doing 'vibe coding' but aren't professional engineers." - **Prospective subscribers are invited with a free entry point.** "It's free if you enter your email on every.to." ## Claimed Scope - **Model evaluation across work categories:** coding, writing, intellectual/knowledge work, design, and agentic use ("how it works inside the agent... Open Claw"). - **Comparative benchmarking against competing models:** repeated comparisons to "Opus 5," "GPT 5.6," and "Codeex/ChatBT" (transcript spellings), including claims like "about 50% more efficient per token than Opus 5." - **Forward-looking AI coverage:** "OpenAI Dev Day is in three weeks. So if you want to stay on the cutting edge of AI, you should subscribe to Every." ## Distinguishing Evidence Notes - The "only subscription you need" framing is a **creator marketing claim**, repeated at least three times—not an independently demonstrated fact. - The review verdicts ("Anthropic are back, baby"; "Fable 5.1 is kind of the Opus 5 killer") are the channel's evaluative opinions, demonstrated through their own tests and demos. - The channel demonstrates its own internal products in reviews (Proof, the "LFG" workflow, "Hands," the evaluation viewer), which shows it builds software with the models it reviews—but the transcript does not confirm which of these are public vs. internal ("we'll probably release it as experimental code someday, but for now it's just something we like to play around with internally"). # Delivery Observation — "FABLE 5.1 IS A BEAST" *Note: the transcript appears machine-generated/translated, so exact wordings ("Fable 5.1," "ChatBT/CHBT," "Claude/Clo") vary inconsistently. Quotations below reproduce the transcript as given. Names referenced (Dan, Kieran, Katie, Mike, Alex Albert) are treated as panelists/speakers as introduced.* --- ## Overall Tone - **Enthusiastic and informal, but structured.** The hosts open with production housekeeping ("We are resolving some audio and video issues. So give us a second.") before moving into a clearly segmented review format. - **Conversational asides punctuate technical review.** Mid-stream: *"Is everyone drinking enough water?"* and at the close: *"And remember, drink water, friends."* - **Excitement is expressed through exclamations rather than formal claims**: *"Anthropic are back, baby."* / *"it just struck me to the core."* --- ## Recurring Phrases and Vocabulary - **"Vibe" / "vibe check" / "vibe coding"** — the signature in-house framing, used repeatedly: > *"We have your signature zero-day 'weibo' verification, just like you've come to expect from us."* > *"helping those who are doing 'vibe coding' but aren't professional engineers"* - **"The only subscription you need"** — a promotional refrain repeated at least three times across the stream: > *"This is the only subscription you need to be at the forefront of AI."* - **Anthropomorphic pronouns for the model** — the model is consistently *"she/he"* and described in human terms ("personable," "she talks more like Richard Feynman"). - **Hiring/colleague metaphors** are a stable framing device: > *"We've all hired someone who's extremely smart but very difficult to communicate with, right?"* > *"I should have fired half of my company by now"* (cited as a good essay opener). - **Delight at unprompted model behavior** recurs nearly verbatim: > *"I probably wouldn't have guessed to ask you to do that, but I'm so glad you did."* > *"That's one of those details that I never asked to do, but it makes me happy."* - **Qualitative vocabulary** leans on feel-words: *"the 'sonnet feeling,'" "feels better in the hand," "hippie-dippy sentence structure," "AI flavor."* --- ## Teaching and Review Habits - **Named, repeatable tests.** The team relies on standing evaluation rituals: the *"writer's test,"* the *"craving to use" test,* the *"LFG engineering workflow,"* rewriting *"Proof from scratch."* One panelist explicitly frames method: *"well, let's just use me as a laboratory."* - **Show-then-interpret demos.** Each demo follows the pattern: run the artifact, then articulate what it proves. E.g., showing the Proof rebuild, then: *"there are many models now that can build something like this from start to finish. But one thing that Fable is still the best at for me is all these little details."* - **Explicit hedging and self-correction** is habitual, not occasional: > *"I am willing to take my words back if it turns out that there are some new pitfalls."* > *"I want to caution that this always happens with new models."* - **Concrete, quotable model output is read aloud** as evidence (e.g., *"raise the ceiling, not the floor," "I've never been so busy"*), followed by editorial judgment about why it works. - **Named comparisons** structure most claims: Fable 5 vs. 5.1, Opus 5, GPT 5.6 / "Soul," effort levels ("medium," "high," "extra high"). --- ## Communication Patterns - **Host handoffs with named roles**: *"Katie Parrott, share your thoughts on writing with us."* / *"We have Mike Taylor, our head of valuation."* - **Promotional pivots at regular intervals**, typically when viewer count or a section break is mentioned: *"if you're here, welcome. We—Every, we—are the only subscription you need..."* - **Parenthetical humor and pop-culture references** appear as connective tissue: > *"I felt like Gandalf: 'I don't remember this place.'"* > *"That's what AGI is, the problem with arrows, right? Forget about the pelican on a bicycle."* - **Direct address to the audience** recurs at transitions: *"Are there any questions about this, or shall we move on to the next one?"* --- ## Stable vs. Isolated - **Likely stable patterns** (repeated ≥3 times): "vibe" framing, "only subscription" promotion, anthropomorphic pronouns, unprompted-detail delight, hedge-then-praise structure, test-based methodology. - **Isolated moments** (single occurrence; weaker evidence of habit): the Richard Feynman analogy, the Gandalf quote, the *"pelican on a bicycle"* joke, the *"hippie-dippy"* coinage. These are characteristic color but cannot be confirmed as recurring habits from one stream. # Fable 5.1 Livestream — Recorded Positions & Reasoning **Source:** Every.to livestream on the release of "Fable 5.1" (transcript appears machine-translated; model names like "Fable," "Opus," "ChatBT/Chbt," "Soul/GBT 5.6," "CodeX," "Claude" are used inconsistently and recorded as expressed). Speakers identified from context: **Dan** (host, name inferred from being addressed), **Kieran Klas**, **Katie Parrott**, **Mike Taylor**, and guest **Alex Albert** (Anthropic). Some speaker attributions rest on contextual cues (e.g., "Dan said…" references) rather than explicit labeling. --- ## Overall Verdicts ### Dan (host) - Fable 5.1 is a "faster and more friendly Fable" — in coding "a little bit better than its predecessor," and "now fast and friendly, like she speaks English." - Claims it is "**about 50% more efficient per token than Opus 5**" and faster, so Opus 5 users should switch. - Framing: "Anthropic are back, baby." - On his own usage: he still spends most of his day in ChatBT, but uses significantly more tokens in Fable 5.1, delegating large parallel tasks. ### Kieran Klas - Fable 5 was great for "fire and forget" work but "hit a ceiling in depth, especially in coding." - Fable 5.1 is also good for **collaboration** ("the Sonnet feeling" — pleasant to work with in loop mode), so it handles both agent-style iteration and autonomous long runs: "**This is Fable for everyone**," not an elitist model. - Compared to Opus 5: "you just feel this dumbness coming on" when switching back; Fable inspires trust in large multi-threaded projects. - Recommendation: effort level "**extra high**" gives the best long-task results, and notably "knows when it's better to do nothing" — good judgment about what not to add, unlike Opus 5's extra tokens. ### Katie Parrott - She found Fable 5 "too difficult"; Fable 5.1 "really can do it all" — worked in her iterative writing process and asked insightful interview questions. - **Explicit qualification:** "this always happens with new models, and over time we discover their flaws, so I am willing to take my words back." - Notes residual "AI flavor": still a bit literary with "hippie-dippy" sentence structure, and "x, not y" antithesis patterns remain, though reduced. - Effort levels: on "X High" the model "gets in its own way" (long padded intros); **medium** gives a speed boost and she likes those outputs. ### Mike Taylor - "Fable 5.1 is kind of the **Opus 5 killer** — better quality, faster, about the same price." - **Main criticism:** default design quality is "boring" when given no design guidance (dashboards, a Typeform competitor); better with explicit design direction. --- ## Recurring Arguments & Themes ### Communication/intelligence theme (Dan, Kieran, Alex) - Dan: friends rejected Fable 5 because it was hard to work with; likens 5.1 to **Richard Feynman** — smart *and* communicative, vs. a "genius savant." - Kieran cites a philosopher's idea (source unspecified): thinking at too high abstraction makes one impossible to understand. - Dan (paraphrasing Kieran's point): the model "gets to the heart of the issue" faster; example — opening an article with "AI responds to 95% of my emails. I've never been so busy" instead of "telling, not showing." ### Discernment (Dan) - Claims this is the **first model** whose meeting-summarization "feed" surfaced genuinely relevant items (e.g., 50/50 decisions needing his input); previous models' picks were "ridiculous." Sees this enabling personal natural-language feeds without millions of data points. ### Demos & evidence presented - **Kieran:** rebuilt "Proof" (collaborative Markdown editor) from scratch in ~6 hours via "LFG" workflow, with built-in AI integration; notes subtle animations, sidebar toggles, and detail handling improved over Fable 5. Context: Dan's original Proof crashed after launch and couldn't be fixed because it was vibe-coded with weaker models. - **Katie:** writer's test — essay intros, paragraph filling, content repurposing; found Fable 5.1 made "very human choices." - **Mike:** internal 12-task evaluation harness; writing task showed clean, publishable structure vs. Opus 5's "crazy headline" and hell metaphors; PowerPoint task nailed brand style and arrows (which GBT 5.6 failed) — "the problem with arrows" as AGI joke. NPS dashboard caveat: chose not to lead with a story ("strong foundation, clear promise") unlike GPT 5.6, though it surfaced a genuine insight ("interesting" predicts passives, not promoters). - **Dan:** "Hands" — a Mac computer-control agent operated via Slack; first-attempt build succeeded where Fable 5 and 5.6 were "unstable." - **Kieran:** 3D/spatial demos — sun reflections, birds that stop flapping, butterfly pathfinding in a breathing-exercise garden: details "I never asked to do, but it makes me happy." - **Alex Albert:** Blender headless rendering demo (architectural render with cinematic tour). - **Mike:** generative-agents AI town (Valentine's-party message-spread simulation), invention-idea loop, and a Slack-simulating "god mode" (built in ~3 hours; he notes accuracy still needs work). --- ## Alex Albert (Anthropic) — Stated Positions - Describes it as a "**friendly Fable**," smoothing "small bumps" from Fable 5; goal of wider adoption and out-of-the-box usefulness. - **Qualification:** "training models is not a perfect science… whatever you do, it always entails some compromise." Defends Opus 5/Sonnet 5 as great, productive models despite flaws; Fable 5.1 "combines the best of Fable and the best of Opus." - On Kieran's complaint about **manual effort-level selection**: acknowledges the friction; says medium/high defaults suit 90%+ of tasks; notes customers migrating from Opus 5 to Fable 5.1 Medium; says "in the future we will explore how to make this even easier" — wants fewer settings himself. - On breaking the model: each generation has an adaptation period; cites Opus and 3D games on three.js as past viral use-cases; currently seeing 3D rendering/spatial thinking trends internally. --- ## Unresolved Tensions & Contradictions (preserved) - **Tool loyalty:** Dan, Katie, and Kieran each state they still spend most time in ChatBT/Codeex/Soul respectively, despite praising Fable 5.1 — "regaining trust" framed as a process. - **Katie's Hands claim:** she called Hands in Slack but had never downloaded it; the agent may have "lied" about using it. - **Default-design disagreement:** Kieran praised design sense; Mike called default output "boring" vs. GPT 5.6's more narrative choices. - **Katie's effort-level contradiction:** X High improves coding (Kieran) but "gets in its own way" for writing (Katie). - Dan himself notes the tension between praising unprompted detail and the NPS case where the model **didn't** make the editorially interesting choice by default. - Katie's live CMS demo remained "in the planning stages" — never completed on stream. # Speaker & Provenance Record — "FABLE 5.1 IS A BEAST" ## Video Overview - **Format:** Live stream (livestreamed model review, ~2,287 live viewers mentioned) - **Publisher/Host organization:** **Every** (every.to / every.co) - **Subject:** Release-day "zero-day" review of the fictional "Fable 5.1" AI model (referred to as from **Anthropic**) ## Identified Speakers & Roles | Speaker | Role | Notes | |---|---|---| | **Dan** (implied host, name used by others: "Dan," "Daniel") | Creator / Host | Leads the stream, initiates the vibe check discussion | | **Kieran Klas** ("Kieran") | Guest / Team member (Every) | Shows coding demos; built "Proof" collaborative editor; created "compound engineering" repo and "LFG" engineering workflow | | **Katie Parrott** | Guest / Team member (Every) | Wears an "Every" cap; presents writing tests and writer's test results | | **Mike Taylor** | Guest / Team member (Every); identified as **"our head of valuation" (evaluation)** | Shows internal model evaluation tool, dashboards, PowerPoint demos, AI town simulation | | **Alex Albert** | **Special guest — Anthropic researcher and product manager for the model** | Introduced late in the stream; discusses model development, effort levels, and use cases | ## Provenance Facts ### Referenced Sources & Third-Party Claims - **Full written review** at `every.to/vibe-check/fable-5-1` — includes "craving to use" test (model received "two golds"), coding, writing, intellectual work, and agent-behavior sections; claims to be free with email signup. - **Anthropic's official statements** are said to be included in the Every review ("we'll tell you what Anthropic says"). - **Alex Albert posted on Twitter/X ~20 minutes before the segment** about rendering / 3D / spatial thinking demos (architectural plot render + cinematic tour). - **"Generative agents" article** (referenced by Mike for the AI town demo); aesthetics credited to **ConcernedApe** (Stardew Valley-style). - **Kieran's own Twitter/X post** with Hadestown set-design renders; Mike mentions possibly sharing invention-loop outputs on Twitter. - **NPS survey data** from Every's own past surveys used as test material. - Comparison models referenced: **Fable 5, Opus 5, Opus 4.8, Sonnet 5, GPT 5.6 ("Soul"), ChatGPT/CHBT, Codeex, Claude Code, Clo/Claude Design, Gemini** (API key integration in Proof demo), Blender, Three.js. - A **"philosopher" quoted from an unnamed source** about high-abstraction thinking — speaker explicitly states they don't remember who said it. - **Richard Feynman** referenced as an analogy (fifth-grader explanation). ### Internal / Uncertain Attribution - **"Hands"** — internal Mac computer-control agent app built by the Every team; described as not yet released ("we'll probably release it as experimental code someday"). - **Proof** — Kieran's internal/spin-off collaborative editor project. - Katie notes **uncertainty about whether the "Hands" agent actually used the tool or "lied"** about it — explicitly flagged as unverifiable in the stream. ### Promotional Segments (flagged) - **Repeated promotion of Every subscription** throughout: "the only subscription you need to stay at the forefront of AI," free review with email signup, prompt library, "vibe check" graphic. - **Upcoming subscriber-only "camp" livestream on Friday** (promoted at close). - **OpenAI Dev Day** mentioned as upcoming in three weeks (forward-looking promotion). - Viewer shout-out/thanks to **"Bassmaker/B Mask Maker"** for a submitted question. ### Comment / Interaction Availability - Live viewer question read on-stream (from "B Mask Maker," asking Katie about Codeex vs. Claude). - Live audience present (~2,287 viewers); viewers instructed to mute microphones; no external comment section discussed. ## Availability Limits - No paywall explicitly mentioned for the review (free with email), but stream recordings/replay availability is not addressed in the transcript. OK. And I also need access to the camera. So it's on my computer, or you can join. It does n't matter which is more convenient. But yes. Hmm, I'd say it's probably better on my computer . Yes. Uh, just, can you connect the camera? Yes, just connect the camera to my computer and a microphone. Hello everyone . We are here. We are resolving some audio and video issues. So give us a second. With us is Kieran Klas. With us is Katie Parrott in her "Every" cap. We love these “Every” caps. Um, if you're here, you should know: Fable 5.1 has just been released. We've been testing it for probably about a week , and it's just a cool model. We have your signature zero-day "weibo" verification, just like you've come to expect from us. We will tell you everything we did with him. We tested it thoroughly. They checked it in writing. Checked on programming. They tested it on intellectual work. Um, a full review of the vibe will be on the every.to website. You should watch this. every.TO. This is the only subscription you need to be at the forefront of AI. Um, I want to start. I want to go through our vibe check. Um, I need two seconds to get ready and show you something. So please wait . While I'm at it, Kiran, would you like to give an overall rating of this model? Yes. Hmm, that's a great model. And for whom? Because Fable 5 was a great model too. Um, Fable 5 was great for people who like to give a bunch of work to just get it done . It's like he's hit a ceiling in terms of depth of work, especially in coding. This model is also very good for collaboration. So she has a feeling that Katie also called "the sonnet feeling." It's just a pleasure to work with her. It's, um, nice to work with in a "loop" mode, which is great because it means I can use Fable for any job since I do both. That is, I either communicate with the agent, iterating, or let it work without stopping. And he does both tasks brilliantly. We will delve into the details. But really, my conclusion is this: this is Fable for everyone. This is not some elitist model that only people from narrow circles will understand. This is actually something everyone should try and use. So if you were a skeptic or didn't like Fable 5, please give this version a try. I agree. I think there's something interesting here. When we say "Fable for everyone," this manifests itself in several ways. The first is, as Kiran said, it's more convenient, faster, and more efficient with tokens. The previous Fable model was powerful but too expensive. She was so wordy or communicated in such technical nonsense that you couldn't get the full benefits of Fable unless you were very tech- savvy for large projects. And this is the first model that I've used where we're really getting closer to truly delegating intellectual work or helping those who are doing " vibe coding" but aren't professional engineers, and it's very, very exciting. For me, there is a question about Anthropic's strategy: we create a genius in the data center and then use it to build all our other models. I felt that one of the drawbacks of being a data center genius is that they are hard to communicate with, and they seem to have lost some of their " cloud" identity, like the 3.5 or 4.5 versions that we loved so much. And it seems to me that this is the case when they returned. They figured out how to combine an incredible coder working all day long while still remaining very personable. Katie , Mike, do you want to add anything? I just want to chime in: as someone who found Fable 5 too difficult and wasn't 100% confident in Fable, I agree with Kieran—this model really can do it all. I tested her in my iterative writing process , constantly exchanging ideas with her , and I understood everything she replied. She asked insightful questions during my interview workflow. And it just seems like a much more affordable model that is capable of incredible things. I'm just checking to see if she can do the whole project for me in one go while we're on this call. So, I'm looking for a CMS to manage my Bible study. In short, I need an interface that will allow me to manage something like a Bible study hub. Let's see. We'll see what Fable 5.1 can do for me while we're on this call. Cool, Mike. I need to run. I will be back soon. I need to change my browser. Mike, keep going. Yes. So, one of the things that all my friends were saying to me when I was like, "Oh, you have to try Fable 5. It's the best model." They replied: "No, it's not the best model , because I don't get good results from it, it's very difficult to work with ." And we've all hired someone who's extremely smart but very difficult to communicate with, right? And it felt a little bit like that. And I remember talking to one of my friends, and he said that the true sign of intelligence is when you're smart and yet you find a way to communicate with people, like Richard Feynman's approach, where you can explain everything to a fifth-grader. And I feel like that's exactly what happened here with Fable 5.1. She talks more like Richard Feynman than like some genius savant. Daniel, I love this. I saw a thought somewhere, very apt, about Claude models, where some philosopher—I don't remember who—was talking about models of thought and how when you think at a high enough level of abstraction, you become impossible to understand because you're simply no longer connected to the real world. And Fable 5 and Fable 5.1 seemed to have fixed that. That definitely sounds right, and that's probably why I like to read extremely esoteric philosophy when I'm stressed, because it disconnects me from the real world and helps me calm down. Ahem. So, okay, I want to provide...we've been testing this for a while now. I want to show people... we'll go into specific tests that we've run, show what we've built and stuff . But the first thing I want to start with is: how does this change my work? And we actually have a very interesting approach, we use... well, let's just use me as a laboratory. Can you see what happened to the way I use AI when I got access to Fable 5.1? And what we found, or what I found, is , um, I used ...and there, there's another site that I need to find. But I used ChatBT much more often even after I got access to Fable 5.1. However, here you see that I use significantly more tokens in Fable 5.1. So I still spend most of my day in ChatBT. So if you're a fan of chat, I think it's likely to continue to be so . But if you're a fan of CHBT, I think you'll find that you can delegate large tasks to it and successfully run them in parallel while working at Chbt. And then , because of how we... because of some of the tests that we did, it's actually more tokens. It's somewhere... even if the price is better, even if it's higher than Fable's, the overall cost of use is about the same. So I think this is one of those models that is worth upgrading from Opus 5 to Fable 5.1. It will be about the same price, but much better quality, and he is really quick and easy to communicate with. Kiran. Yes. Like, yes, the price is the same, but it works much better. I was switching between Opus 5 a bit, especially when we had early access, and then we lost it for a while , and I went back to Opus—and you just feel this " dumbness" coming on. And this is especially noticeable in large projects where many things are happening at the same time, it's just that the trust you can have in someone who really understands everything is there in Fable, but it's not there in Opus 5. And Opus 5 does a great job with design and other work, but why, why not choose a model that gives more for almost the same price? It simply uses tokens more efficiently. So yes, if you're using Opus 5, try Fable 5.1, because it's very similar to work with , only it's just bigger and better. And yes, I feel like we have n't been in the saddle here in a while, guys. This is fun. It's nice to be back. This is fun. Is everyone drinking enough water? I have water here. I also have I have coffee. Sorry . You have tea. Mike and I are on a diet. Well, this chilled " cigarette" drink, I think, is appropriate for Fable 5.1. Who has something interesting they want to show? Kieran, I know you created something there . Maybe you could show it. Mike, I'd be happy to have Kieran start cooking this. I would like to go through some cases. For example, if you showed us the slides or that blog post that he wrote—it was really cool, especially for SOS 5. Yeah. I'm just desperately trying to hide it . Yes. Secret model names. In short, if we can do it, it would be great. I can show some materials while they both get ready, because no one listened when I asked to check everything in advance. Well, yes. Do you need time, Kieran? I have them here. I can make one. They are already ready. Do one, and then we'll hand it over to Katie while Mike tries furiously to hide the names of his code models. And then we'll get in touch. Here, come on. Class. Yes, I shared the screen. If you can show it. So, my favorite benchmark for Fable is to rewrite Proof from scratch. Proof is an amazing collaborative editor we built; It's like a Markdown editor where you can see the origin of the data, it's agent-oriented , has a bunch of features, and it's very complex. And I just said, "Hey, can you take a look at this idea?" I have a description of the problem and I simply pass it into the LFG engineering workflow, which brings the idea to life from start to finish. And that's what he did. And that's very cool. It took me about six hours to work on it, and it's essentially an entire editor with "accept/ reject" functions. There 's even a built-in AI that found the Google API key and integrated Gemini to do the work for me. So you can ask AI to do things, like make it shorter. Let's see what he does here. Functions. Well , I want to accept it. Boom. So , it looks very polished, and you might ask, what's new in 5.1. My opinion on Fable 5, which was the best before, and Fable 5.1—what I noticed is that the new version handles the details a little better. Well, details like the ability to focus on text, for example, there 's a toggle here that just hides that. The sidebar toggle was not there in the other version and the activity did not scroll with me. I also noticed that the design has gotten a little better, they have improved their design sense. They also adjusted the animations. I think the animations used to be a little too flashy, with the appearances and all that, but here they're very subtle. So they have their purpose. Here, for example, it seems to push things closer to how they were. And I like this clarity. Also, if I refresh the page, there seem to be very subtle animations. So they set that up too. Of course, I also looked at the code, it looked really good, but Fable 5 was incredible too. This is the "extra high" level. I think if you run these long tasks on "extra high", you get just great results. "High" is very good, "medium" is also good, but if you want the best, choose " extra high", it goes the furthest. And the best thing is that he also knows when it's better to do nothing . So he won't add anything extra . He will just make it better. So, "extra high" means better results. This doesn't mean, like in Opus 5, a bunch of extra tokens and things I don't need. The ability to judge is very well tuned: whether something is worth adding or not. This is really very interesting to see. Are there any questions about this, or shall we move on to the next one? I think let's move on to the next one. I think it's ...well, actually, before that...I think it's important to emphasize how hard it is to build an app like this, it really is... Can you explain what kind of challenges you faced with your version and how you, well... It's a long story at this point , but I built the first version of Proof, I think, about six months ago, we launched it, and it crashed on the first day. I tried to fix it, but it was completely impossible because I completely "vibe-coded " it, and the models at the time were not powerful enough to: A) " vibe-coded" it from scratch so that it worked, and B) fix it if it broke. Now they , now they definitely can. So, comments work, change tracking, real-time text updates—it's kind of crazy. Actually, Kieran, there are a lot of things here that we should just add to Proof. I know. I know. That's the trick. I have my own spin-off project, Proof, which was Fable 5, I made it in one go, put it in the open access, and now I think: this is better than that version. So, yeah, there 's some really cool stuff here, like in another intro I just do this, bang— and it works. There's so much that looks incredibly smooth, and it's very difficult to implement in software—to make it look smooth and actually work, there's CDT, multi-user mode, you can see who's where, and you can click " view." So the basic idea is that you need to review it, mark it as verified, and you can even confirm it. You see levels of how much you support the text and you can click here. So you go through what you have to do. You can see how much was written by artificial intelligence and how much by humans. You can see what kind of AI it is, whether it's built-in AI. You can share this via the CLI link and invite your own AI to collaborate. There is no web MCP here, but I have to add this to the font. But it's just crazy what can be built in one go. And that's a really, really strong move for Fable—to do it tastefully and aptly. And this is what I want to emphasize: there are many models now that can build something like this from start to finish. But one thing that Fable is still the best at for me is all these little details that it just does, and you look at it and think, " I probably wouldn't have guessed to ask you to do that, but I'm so glad you did." This is exactly what I wanted. And that is very difficult to do. Yes, and it costs a lot of time . If you do this with other models like the Opus 5, you'll get very far, but it will take days of iteration to get it to that level. And the fact that you don't have to be in the development cycle as a person to evaluate and guide is very valuable. So I think that's the real power of this demo. Yes. I can do other demos, or maybe someone else will do one now, and then we'll come back to something. So, I want to take this moment and just say to everyone: if you're here, welcome. We—Every, we—are the only subscription you need to stay on the cutting edge of AI . When new models come out, we get them early and conduct a test on the day of their release. We just published our review after a week of testing the Claude 3.5 Sonnet. Anthropic are back, baby. They drove. If you go to every.to/vibe-check/fable-5-1. Or just go to every.co and click on the link. We have a detailed review with all our practice tests. You can read it with ChatGPT, you can read it with Claude. He is quite wonderful. We have our “craving to use” test. This, you know, is for those who evaluate the model: do we gravitate towards it, and why? This model received two "golds ", which is very rare. We talk about programming. We've already talked about evidence here. We are going through the writing of texts. We are going through intellectual work. If you're interested in this model, if it's important to you to learn what it's good for, what it's not good for , and where it fits into your workflow, you need to go to Every, subscribe, and read this. It's free if you enter your email on every.to. And welcome to our broadcast. We will analyze this in detail . Katie Parrott, share your thoughts on writing with us. Yes, we asked Claude to take our writer's test, which, let me remind you, is a set of tasks that reflect what we might ask the model to do. This includes writing an essay introduction from scratch, filling in a missing paragraph when your brain “ shuts down” and you don’t know what to say but want the model to try, and adapting content from an essay for a LinkedIn post, email, or X-post. So I can show you my screen, and demonstrate to you. Ahem. Cut, cut, cut. I need to set everything up. OK. I can't see anything. Did you click " share"? Um, share. Window. Let me...Oh, here it is. So this is a viewer that I created for myself to be able to evaluate all the model output. Well, just to make them more understandable for myself. So here you see on the left the introduction to the actual article that we publish. It comes after the mega post about automation, which you should definitely find if you have n't read it yet. And here the model X is Fable 51. Hide, hide the names of your models, friends. And we, and then Fable 5 here on the right. And what I really noticed about Fable 51 is that it gets to the heart of the issue we're trying to address much more quickly. So you see in the published version that there is a paradox at the heart of AI. You know, we immediately create tension there and this notion that we've automated everything we can, but there's more work than ever. If you look at Fable 5, it deviates a little bit from the theme of how crazy it is that Dan can do his mail. It just takes too long to get to this part about it creating more work. While the Fable 51 model immediately gives what, frankly , I find to be accurate: the AI responds to about 95% of my emails. I've never been so busy. So, I found that this model makes really good, very human choices, and you may disagree with them , but they don't seem synthetic. Now I want to caution that this always happens with new models, and over time we discover their flaws, so I am willing to take my words back if it turns out that there are some new pitfalls that we should be aware of. But so far I find it quite humanistic. Still a bit literary, as we said. You know, she likes this more, let's say, " hippie-dippy" sentence structure. I don't know if that makes sense to me. I don't know if that makes sense to me. Yes, completely. This is a scientific term. Yes. Hippie-dippy structure, right? What are the vibes like there? And one more thing I'll say quickly, so this was a comparison of Fable 5 and Fable 5.1. We can also consider a model based on comparing levels of effort. And you know, Kieran mentioned that when you switch to "X High" she really improves in programming and then she can kind of run long distances. Actually, I noticed with this model and the past ones, like ... Oh, I need to remove this example. Well, whatever, I noticed that on "X High" she was getting in the way of herself a bit . Here she has a very long introduction about everything we do. It can be manual, and it's not as fast and interesting as what we got here with our hit, like "I've never been so busy." So I like that middle line too. By all reasonable standards, I should have fired half of my company by now. I think this is another very good way to start the article. So the "Medium" level for Fable is worth checking out, I would definitely say, because you get a speed boost. I noticed that it takes on a bit more of an "AI flavor" with changing effort levels. So, this is a report on the state of writing. I like it. I just want to summarize this a little bit. There is something interesting here. It's really interesting to watch him rework my text and think, "Oh, do I like this or not?" So, I definitely... The introduction that I made was, "There's a paradox at the heart of AI ." What's interesting is that in this introduction I now see something I didn't see before: I'm telling, not showing. And what I like about this version of Fable 5.1: "AI replies to 95% of my emails." "I've never been so busy" is just a show. I think it would be a little stronger if I added "and yet I've never been so busy" to emphasize the contrast , because I may not immediately catch that connection in my head, and I want to connect one sentence to another. But I really like it. And I really think this demonstrates a step that AI has historically had a hard time taking: understanding what's interesting and putting that first. And the fact that this model is stronger in this, uh, what I would categorize as what I call a distinction , uh, a distinction of whether it's appropriate for the situation, or my interests, or something like that . Um, the fact that she's better at it is good for writing, and it's good for a lot of other things too. For example, one of the things I did with this model that worked for the first time, I'll try to show it in a demonstration. But, um, but what she did, what this model did — she created a feed of all the meetings that had taken place and created short summaries, like little tweets, that she would pull out for me. Okay, here's a meeting where there was a 50/50 decision , and it would be helpful if you could give your opinion. Usually the models found all these solutions, and I thought they were ridiculous. Like: I understand why you think it's appropriate, but it's actually not appropriate at all. Your, your , your discernment is not good enough. Um , and this is the first model that pulls real stuff out of meetings, and then I think, "Oh my god, I have a new perspective on my company." And getting to the point where the models have good enough resolution to do that, I think, opens up a bunch of different possibilities. It's like writing, but it's also about how you run your company? If you can, if you can, um, isolate useful information for yourself that is relevant to you. You know, feeds used to be the domain of social media, and you had to have millions and millions and millions of data points to create feeds. Now you can create your own natural language feed, and it's getting pretty good, and I think in a year it's going to be a lot better, and I think Fable 51 is the first model where I think, "Oh yeah, that's something I would check out." Um, that's why I want to continue, I want the show to go on. We have Mike Taylor, our head of valuation. Mike, do you want to show us something? Is there anything you want to show us? Yes. Yes, I'll show you something. Uh, let me share the screen. Eh, that's all . OK. I hope you see that. Er, one second. Please wait . Yes, we see. Okay. Before you start, I should say that Mike created this to help us conduct an internal evaluation of this model. It allows you to see each model and each result of its work. So it's very easy to compare . We have been doing this for a long time, professionally. Um, and using it is a pleasure, it's really cool. I'm so glad Mike is showing this. Yes. Yes. No, no, it's good that we put it aside and released it into the world. But yes, these are a number of tasks that were very important to me at the beginning of the year when I was testing AI. And, in fact, I can run all 12 tasks that interest me, get input and output data, and each one is executed separately. So this is essentially a fair test. For example, one of the things I often do is create dashboards. And you can see that this is the panel that he created. I wanted to share this first because Fable 5.1 didn't feel perfect for me , as the default design is n't very good. Kiran, I know you liked it from a design perspective. But for me, it's a pretty boring panel. I also have another task: I asked to create a Typeform competitor where you can speak instead of type, and it produces the same standard "vibe" code. The caveat is that this is the default design . I didn't give him any specific design guidelines. I didn't say it had to follow a certain design. So I think it would be much better if you gave him more design guidance. But that was the only downside I could find in this model. Mm. Mike, can you open the writing assignment for a second? Yes. This is one of my favorite quests for Fable. What's interesting about this result? By the way, would you like to describe the task itself first? Yes. So, this task is what I actually did in real life. I asked AI to write this post. Originally it was Opus 4.8. Um, and I'm like: I created a cool tweet and thought it should be a separate post. So, I gave him a huge transcription of my thoughts, a kind of monologue about everything I was thinking when I wrote that tweet. And then I gave him a sample of my text and said: this is how I write. And then the test was, really, how well he did at writing it for me. Yes. This is a completely sentimental post, and I just want to read part of it. So raise the ceiling, not the floor. The best thing about it is that it's quite interesting, he highlighted the right interesting point. The problem is that it's X, not Y, and it's not specific enough for a post, but it's still not bad. Seven things I learned about getting teams to actually use AI in my first few months of running a tech consulting business—that’s a perfect, straightforward subheading, exactly what we would publish right away. I think he did a good job with the introductory paragraph. Sampar asked me a question in X, and so on , now that we've defined what this article is about, positioned it correctly , and then, you know, every section here, if we continue, when a company decides to implement AI, the first instinct is to check the tools. Someone puts together a spreadsheet with vendors: cloud, or Codeex, or Gemini under the hood, and the team spends six weeks. This is exactly what I would say, if I saw it, "Yeah, that's a perfectly reasonable way to start a paragraph." I want to show you for comparison, look at Cla...Yes. Ahem. OK. Nobody wants to use your AI tools. It's the same task. That's some crazy headline. It's very similar. This is very similar to what the Opus 5 did in our tests. Like, he said, "we didn't fire anyone." Why are you starting with this? That's right. What are you even doing, dude? And so I think this is a very good example, because this is exactly what people didn't like about Opus 5, and which is clearly fixed in Fable 5.1. And, um, you can see it in other parts if you scroll down a little bit on Yeah, on this one. No, go back to Opus 5. I want to keep There's some really crazy stuff here. Um, wait, wait, wait. Well, okay. go to go to number two. Hmm, I don't know. There's , there's, there's one of these numbers where it just kind of just goes , it just completely went off the rails, talking about hell and there's a special place in hell reserved for something like XYZ, and it was just, you know, so Opus 5, as you know, as everyone knows, we gave Antonify pretty bad reviews, and it's really nice to see that they figured out how to package it, and so I think Fate 5.1 is kind of the Opus 5 killer , it's just you get better quality, it's faster and for about the same price. Yes, and where he didn't get 100%, it was somehow insignificant. Like, he was a little too verbose, um, but like most models, they fail at that, um. He had, uh, a couple of AI traits, um. So, um, you know, we didn't automate the workflow; we used new opportunities to make the work an order of magnitude better, right? Like this antithesis of "x, not y," so there's still a bit of that in it, but it feels much more human. It doesn't seem like an obvious Claude, you know ? Do you have, um, do you have PowerPoint? I want to show the PowerPoint because there are a few things in it. Yes, PowerPoint. That's what the public likes, right? Well, yes . So, um, prepare the groundwork for us . Yes. So, so this assignment is basically, here's a link to the repository for compound engineering, um, that Kieran created. Uh, no, sorry, but, uh, create a PowerPoint to teach this, you know, in a short training session. And that's what I really had to do at a certain point. And, uh, I used compound engineering a lot, but I probably didn't know all the intricacies of the repository at the time. So this was a task that was particularly close to my heart, uh, because I was kind of teaching myself at the same time. Um, but, but one of the things I really like about this is that, um, we have a pretty specific brand style and it finally really lives up to it. Well, so it adds all these extra, like, uh, swirls, and background gradients, and all that stuff that previous models couldn't handle. Uh, and he doesn't try to cram too many different ideas onto one slide. Uh, so, you know, you can see that here, well, just nice spacing, you know, it's not difficult to read at all. Uh, so I would say it's very similar to something I could teach myself. And another thing I just wanted to point out is that this slide immediately made me think, "Oh, we need to borrow that." Because I have n't shown it this way before. I was just showing it in dots, and I thought, " Oh, this visualizes much better, with and without the stacking effect." Uh, so I even had a few moments where I was like, "Oh, okay. Yeah, this is better than what I had ." There are also a few little things here that I find very interesting. You were showing...you know about that slide with the arrows pointing in... uh, in a circle? I think it's somewhere closer to the beginning. Uh...trying to find him. I'm flipping through. Almost, almost there. Here. Here he is. So, okay. Force the model to automatically create a layout. Okay. We have a "plan," "work ," "review," "accumulation "—and these arrows point in the right direction . If we go to the seventh slide in GBT 5.6. Uh, yeah, I want to show you what it looks like. Or maybe it's not the seventh slide, because it's non-deterministic, but one of the slides is just like that. Uh, yeah, this one. Oh, yes. Uh , yes. Well, look at this. You know, she's trying to do the same thing, but she just can't quite manage it. It turns out that placing the arrows correctly is very, very difficult. Yes. That's what AGI is, the problem with arrows, right? Forget about the pelican on a bicycle. Okay. Creating arrows in PowerPoint is very, very difficult. Uh, and this is one of those things: you look at it, it's a small detail, but if you really want the model to do a presentation for you in one go, those details have to be right. Otherwise you just think, "I'll have to do it myself." And Fable 5.1 is exactly what we were talking about: it's the first time you can truly delegate intellectual work. She's really, well, she's, she's, she's really much better in many of these ways. Yes, definitely. Yes. Yes. As for the design, Clo Design, the tool itself is great because it pushes you a little further. So when I say he's good at design, I mean he's good at executing design tasks, following instructions, and working in a way that doesn't make mistakes. So in that respect it is very useful. And I've been using Clo Design a lot with version 5.1, and it's really good. Now you can synchronize data with Clo Code very easily. But of course, is this the best- looking design? Not always if you don't have the right skills or clear direction, but at least if you say, for example, sync Figma with your site or Clo Design with the site, it actually works well, which is great. What I mean by default design is this , for example, the Fable 5.1 control panel. It's, you know, it's not bad, and it has a lot of useful things in it. But it's a little less engaging and less narrative-oriented than what GPT 5.6 Soul did. So he came up with a better headline and, you know, chose more interesting things. So I would say there's still a little bit of ...you know... that's something we should talk about. Yes, let's talk about it. Yes. So, look at this. This is GPT 5.6. It's one thing, I think Fable 5.1 can do it, but it's a challenge where I think, "Oh, it's interesting that he didn't choose that for this case." 5.6 had a strong foundation and clear promise, right? For context : we took our NPS survey from a few years ago . We loaded it into the model, and it's a test we run all the time. And GPT 5.6 decided to tell a story like: strong foundation, clear promise. Customers see a rare combination of thoughtful AI and media in everything. Thank you. and useful products. The level of support increases dramatically when the combination turns into a simple result, practical help without unnecessary noise. This is actually a very useful summary that is not about numbers. And if we go back to 5.1 um, he decided not to do it. He just decided to immediately post every survey question from the NPS dashboard and so on . Would Fable, if you asked it to complete a task like GP 5.6, would it be able to handle it? Apparently. But it's always interesting to see what a model decides to do on its own, without being asked, with any request. And what's also interesting about this is that we praised her willingness or ability to correctly capture all the little things without you asking. And this is the case when for some reason it doesn't work so well. Yes. And I have a strong belief that if I had directed her a little, she would have done well, maybe even better. But I just like to see what the default settings are , you know? Yes. And if you scroll down, here were some things that she highlighted. This again brings us back to the question: what is her level of insight ? Can she highlight anything interesting? Um, and here are examples of people who can't describe why they don't recommend it. This is an interesting point. Or Oh, that's a good example. Um, scroll up a little. "Interesting" is a word of caution . "Useful", " great" and "smart" predict promoters. So people who just say, "Oh, yeah, that's interesting," are unlikely to actually become our real promoters. And the fact that she just highlighted this, without any prompting from us, is a really interesting insight that could change the way we build businesses. Yes, 100%. That was really helpful, and, you know, the other thing I would like to point out and that I liked is the adjective scale, which is where I think she gets that observation from. I had n't seen this design for an NPS survey before, and I thought it was pretty cool. It was like she just did some keyword analysis and then presented it in a visual way that made sense to me . Yes, that's really cool. Um, okay. So, uh, what I want to do now, we have a lot of people gathered here. We have about 2,287 and the number is growing rapidly. So, if you're here, welcome. We—Every, we—are the only subscription you need to stay on the cutting edge of AI . See, it's written right below. Uh, not in line, right below. Uh, every.to. What, uh , what I want you to know is, if you want to know the verdict on this point, you should watch our stream right now. But you can also go to the Every website, which is right here, every. Uh, and we have a vibe check, a written vibe check for Fable 5.1. Verdict: Anthropic are back again. Uh, but overall Fable, uh, what's really cool about this Fable 5.1 version is that in terms of coding, I would say it's a little bit better than its predecessor Fable , which was by far the best coding model in the world . Uh, but she's also now fast and friendly , like she speaks English, and that's a big, big, big achievement. So we are fans of this model. Uh, she works, she's been working for days on end. Oh, that's something I can demonstrate to you in a second. She works for days on end, completing tasks. Uh -uh, and she, uh, and it so happened that someone was typing. So just turn off your microphone if you're typing. It turns out to be about 50% more efficient per token than Opus 5. It's also faster than Opus 5. So if you're using Opus 5, this should be your new go-to tool. You should read this "vibe" check. We'll break this down further on Every, but you should read this. It's completely free. Free. You just need to enter your email. I have everything from what Anthropic says to our usability test, where we talk about, you know, I think the best way to test if a model is good is if we've switched to it and are we using it every day? If so, why, and who are we? This is our test of the model's demand. We have coding that we, that we parse. We have writing services. What else do we have? We have intellectual work that we have just examined in detail. But it really is a faster and more friendly Fable. And I think that's a great achievement. Uh, what I want to talk about now in terms of coding, I actually asked her to create this thing for me called Hands. And Hands is a pretty cool thing. It seems Katie uses Hands too, which is pretty cool. And Katie, if you have anything to say about this , let me know. So, Hands is an agent for controlling a computer. It's a Mac app that runs on my computer and I can send tasks to it. So, uh, yesterday I asked him to create and share a history podcast about the history of everything. I did this through Slack, and the message from Slack just goes from my computer to Hands . He uses my computer. He simply enters the text into the codices. It's a little tricky to demonstrate live, but it injects data into CodeX, ChatGPT, Claude, or my browser. It gets the result and then sends it back to Slack. So I can actually use my computer fully through Slack. I think computer control is one of those things where I think, "Oh, this is really, I think it's improving quickly, and I want to learn how to do it well." This is interesting to me, and it's nice to be able to control my computer from Slack. And this was a first- attempt build from 5.1, and regular Fable and 5.6 both failed to do it. They were close to success, but the result was very unstable. And this is the first option that actually seems to work, which is quite, which is pretty cool. And I use it all the time. Katie, it seemed like you were using Hands. Is that so? Well, I called Hands in our Slack, but maybe I didn't have it set up and the agent could have just said they were using it and actually lied. Oh, really? Yes. Did n't you download it? I didn't know it needed to be downloaded. I thought it was a skill that could just be acquired. Sorry. Batman. OK. Well, in any case, you should try Hands. I think that's pretty cool. This is a Mac application. We'll probably release it as experimental code someday, but for now it's just something we like to play around with internally. For those who have joined, Kiran, is there anything new you want to talk about or show? Uh, I do, I can share a few short things. Uh-huh. So, with LFG bench, which is a complex engineering skill that does everything on a single request. I have some results here that I can look at and maybe say what's cool about it. So it's mostly for long tasks, but in terms of, just oh, wait V-code, yeah, on the VI- coding side. Yes. Yes. So I really like this model for VI coding as well. But the examples here are, um, results. This is the design , what I'm testing here is whether it can just do the design, and it's a little bit oversaturated, but I'm looking here to see if I can find anything that gives away AI, or a new feature of AI, or something annoying, or something that I wouldn't publish. I think we're at a point where he can easily create complete landing pages. Um, the other thing is 3D, uh, spatial. It's a cozy island and you can see it looks pretty good. The other models look good, but the new addition here is the display. So, you can see a reflection of the sun that I've never seen before, which is really cool. You can see birds that, uh, they sometimes stop flying. So they don't flap their wings all the time, which is also new, something I've never seen before. Um, you see it's very detailed. I think there are no mistakes here. Um, so yeah, I think this one is pretty... yeah, there's always something to add, which is pretty cool. And here you can see butterflies flying. Um, more fireflies and fireflies. Yes. I've seen this before, but all these details are just cool. This, this is a garden for breathing exercises. So it's basically isometric. Um, and yes, you can just go into the desert, do a breathing exercise session, and if you finish, it will turn green and a flower will grow . And if you do it longer, you can get more flowers. Things are cool here. Like this butterfly here. I never said that. He flies from flower to flower. So, he flies from this flower to that one, and flies around. That is, there is a whole mechanism for finding a way for a butterfly flying to flowers, which is something new. And this is one of those details that I never asked to do, but it makes me happy. Like, oh, that's smart. I should have done it. And Dan said that with proofreading and writing, it's just fun to have a model where you see the details and think, "Oh, this is it." And this is very unique. We usually get annoyed by all the mistakes she makes. uh, it doesn't work, and now we're putting in the effort because we're excited about the model's results, and that's something special, and I think that's what makes working on these things very interesting. Uh, I guess the more effort, the more pleasure, so the " extra-high" level brings much more joy than "high " or "medium", although "medium" is very suitable for coding "by vibe" or just for cyclical tasks. Uh, well, yeah, those are probably the main things from the LFG launch. Surprisingly. I really like the way the breathing exercise garden looks. Oh, and by the way, on the subject of little things and their importance, we have a special guest today—Alex Albert from Anthropic. Greetings. Greetings. Alex is a researcher and product manager for this model. So, you helped make this model what it is. Thank you for all your work. Tell us about this model. Tell us about your thoughts on it and the process of its creation. Yes. Uh, well, thanks for inviting me, Dan, Kieran, Mike, Katie. Nice to see you again . Uh, this is an interesting topic for me. I was very involved in this release and it was a long process. Of course, it was a great team that introduced this model to the world. Uh, and we're all really excited about this . I think it's best described as a " friendly Fable," as you, Dan, called it—it's like the icing on the cake to what Fable 5 was in the middle of the summer. Well, I think there were a lot of small bumps that we smoothed out in this release, which makes it, firstly, very easy to interact with the model and use as an everyday tool, and secondly, it contributes to the wider adoption of this model. We focused on a bunch of things to make sure it would be really well received by anyone who wanted to implement it in their company, in an organization, or just use it on their own. Uh, so there's a lot to love here. I would be happy to just talk about all of this. This is great. This is amazing. Uh, well, tell us a little bit about, you know, I think there was this arc of Fable's development: it came out, and we were like, "Damn it, this model is crazy ." “And from our perspective, the Sonnet 5 and Opus 5 launches seemed to take on a bit of the Fable character, which is hard to communicate with, and, you know, they just lost that Claude ‘zest’ . And now it seems she's back. So can you tell us a little more about that journey, and if there's anything you can share about what was going on behind the scenes and what you were trying to achieve, I'd love to hear it. Yes, I think that training models is not a perfect science , right? So, whatever you do, it always entails some compromise, and we are constantly balancing between one and the other. This doesn't mean that functionality is necessarily sacrificed entirely for the sake of the feel of the model or behavioral aspects, but there is no formula where you can simply plug in the right data and get a perfect solution for everyone. Honestly, I think the Opus 5 and Sonnet 5, despite all their reported flaws, are still great models that millions of people use every day. In addition, they are very productive. For example, Opus 5, if you just run it in the background, it will do a great job and be very efficient. I think we learned a lot from those models, especially through feedback from talking to many users, customers, and people like you. It really helped shape Fable 5.1 into something great, combining the best of Fable and the best of Opus, bringing it all together . Which is very valuable to me. And I think this is actually the beginning of a long growth trajectory where we've already warmed up, and I think it's going to be very interesting with this model and future models. Surprisingly. Kiran, do you have any questions? Yes. I like everything about this model except one thing, and I'm curious what you think about it: I still have to choose the effort level, which is a bit annoying because it's so good at long and short tasks that I don't have to think about choosing for a short or long task, but I still have to think about the effort level. I know you used to have this dynamic routing system, and I'm curious because that's probably the only thing that still annoys me about the model: I have to decide if I want extra- high or medium, and even though they're well-tuned, I still have to choose. Yes. Oh, thanks for acknowledging that they are well-adjusted. We definitely spent a lot of time choosing the right default settings . So for most users, the high level, which is standard in cloud code, and the medium level, which is on our other platforms or services, is what you should stick to for 90+ percent of your tasks. It just works. This is the best balance of intelligence, efficiency, and speed. And this is what I myself constantly use by default. I think medium, extra-high, and even going down to low gives you a range of choices, which, you know, Kieran, since you're an advanced user, you want to customize the model for a specific task. And I think for you, it's a little bit like, " Okay, should I switch down or up?" If you switch to medium-low, it's truly a cost-saving option while maintaining intelligence and speed. You'll get better performance than the Fable 5, for a lower price and faster. We see customers even migrating from Opus 5 to Fable Medium, Fable 5.1 medium, due to the high token efficiency and low costs. And choosing an extra- high level means you will use the model more autonomously. Maybe you want it to carefully work out all the details and produce the most polished result. I think in the future we will explore how to make this even easier. From a user perspective, when I think about my own use of models, I want fewer settings and toggles to think about every time I make a request— I just want it to work. Um, so we're very aligned on this approach, and that's something we'll continue to work on. Perfectly. Class. I wanted to, um, I wanted to ask you something, if you don't mind. Yes. So, Alex, one of the things I realized when I was running benchmarks is that my tasks aren't ambitious enough to really show how good this model is. And I had to spend most of my time rethinking: what can I do? Where can I try it out? And I do n't even really feel like I managed to " break" it anywhere. How do you guys think about this within the team? Do you see people on the team who may have had it longer, or are they starting to take on more ambitious tasks? Oh yes. Of course. I think it's a funny thing—this " possibilities pool" that exists in every new generation of models. It usually takes a few weeks, you know, even a month for a model to fully unlock its potential in what's possible. And for me, the clearest example of that was Opus four or five when it came out last November . Then it really became clear only after the winter holidays, closer to January, when people started saying: " Oh, wow." Okay, this is a really special model for standalone background tasks. Everything is finally falling into place for a lot of the things I wanted to do. I think every model has this adaptation period. So over the next week or two, we'll start to see examples emerge. Internally, our people have already started testing this model on everything. On everything, without exception. And I think that's the best way to understand boundaries—is to just give free rein to your ambitions. Ask yourself: why can't she complete this task? Have I tried this before? Maybe I'm just artificially limiting her. This is the first thing. And secondly, I think every model has a moment where some new type of challenge opens up, which usually goes viral on Twitter or is very common, and a lot of people start trying it. And as with Opus 5, one case I remember is 3D games; a lot of people suddenly started making games on 3JS and, well, really testing the model in that sense. Um, one thing that we're noticing internally , and I just posted about this about 20 minutes ago, is rendering, 3D models, and spatial thinking. Um, in this example, I gave the model an image of a plot of land and said, here's the plot. Can you design a complete architectural design for her house, render it, and then create something like a cinematic tour, you know, like a realtor would add to their listing? Do you want to show us? Um, yes. I, uh, I don't know how I can open my file here. Maybe someone can just open my Twitter, it's there. Yes, I can. Um, my new favorite benchmark is to have each model try to build a set for the musical " Hadestown." And I found that it demonstrates the capabilities of the model very well. And I got a pretty good result from Fable 5.1. Yes, a cool test. We love him. Uh, uh, wait, wait, wait. We are going to share the screen. Share your screen. That's it . Do you see my...Oh , don't you see. Not yet. But you'll see. Okay, here he is. I think that's what you're talking about, right? Yes, that's exactly what I'm talking about. So, I, I mean , it just struck me to the core. I didn't specify almost anything in the request, and she just figured out: oh, there's a Blender library that can be used without an interface for launching and rendering. And then she started taking screenshots to check everything as she worked. I eventually came back to it a few hours later and saw the video it had processed and rendered on my CPU quite efficiently. Um, and I think that was something I didn't even know was possible to do, I just had never personally used Blender before. I did n't know that was even possible. Just by providing such a description. Blender is having a moment of glory right now for some reason. Blender. Yes. and Linux. Linux and Blender. I just downloaded it today. I'm like that , I guess I should switch to Blender. So, um , yeah, I mean it was just a library that the model found and used. Uh, but I think even without getting tied to it specifically, it's just the type of task where you think, wait, what if I wanted to redo my backyard? Can I just ask the model to do it? It's like I think Mike has a few demos of this that I want you to see, because I think you'll be excited about what this model that you've been working on is capable of. Mike, do you want to show us something? Yes, of course. Let me, uh, I also got to the point where I thought: you know what I could do that seems too ambitious? Uh, so I really liked the article about generative agents a few years ago, I've been thinking about it a lot , and I like the idea of creating a little AI town where they just run around and talk to each other. So I said, here's the article, can you do this, and I came back an hour later, and it was already working. Thanks to ConcernedApe for the aesthetics. Yes. Yes. And this is specifically from the article. I think it's a Pokemon-style aesthetic, which I also really like. Eh, but that's pretty cool. Uh, and then I thought that this could be used to understand how messages spread through the population. I thought, let's make it even crazier. So, uh, you know, it was kind of throwing in new things. This is also from the article, but the point was that if we tell someone there will be a Valentine's Day party, how will it spread through the community? And I just think that this is, of course, a toy version of what can be done, but for economists or politicians, it would be something incredible. You know, they could actually do A/B testing of different messages and see what works. Yes, that's impressive. It's like a combination of this analysis with visual generation and all these different skills mixed into one. Uh, I ran it in a loop to generate ideas for inventions, things that don't exist but that should exist for me . And he started releasing some crazy stuff, like a walking stick that lifts itself up or a whiteboard that erases itself. I may share some of them later on Twitter, but this is just incredible. These were full-fledged engineering diagrams with parts lists and more, as well as animations of how it would actually work. Well, just crazy stuff, I don't know much about that kind of engineering. I wouldn't have been able to guide the model in such matters, but she knew how to turn this vague technical specification into a working product. Yes, that's brilliant. There was one more thing, because I thought I did it without really thinking about it. So can we be even more ambitious? So I made him mimic our Slack. He just pulled up everyone from our Slack, their message history, and now you can compose a message, paste it in here, and get a kind of " god mode" for that message: who will respond, what will people think Dan will say? Of course, there's still room for improvement in accuracy here, and I think there are Simile AIs that do this as a product, but I was impressed that it only took three hours, and it really taught us how information spreads internally: if you post to a channel for everyone, it spreads much faster than it does within a team of consultants. So I feel like we're not even taking a lot of things into account here. Yes. Yes. I have one question, because it's great, but I think the strongest point of this model is that it actually does a good job with smaller tasks. Have you seen people within the company switching from Opus to Fable for smaller, operational tasks like this, because Fable 5 could already do crazy things, and now it's even better, but in everyday life I won't be using it all the time. this is really cool. And have you noticed a shift in how and when people within the company approach Fable? Yes. Yes. I think this is, of course, the other side of the spectrum. You have this incredibly smart and capable model that can do some very crazy things. But at the same time, many people have work to do— very practical work. And we are focused on making this model just work “out of the box ” for users. So how does it actually integrate into our products like Cloud Code, into our services, into collaboration, or anything else? Within the company, this is the model that everyone turns to for these tasks. I find her ability to be token-efficient and very focused in code edits or when writing blog posts and editing text to be extremely useful. I use it all the time for such tasks. And I feel like, as you mentioned earlier on the stream, it has a better—it's hard to define, but it just " feels better in the hand ." And that's something that can't really be captured in a benchmark or an assessment. This only really shows up after a few days of use. So I'm really excited that people will start to notice this when they try this model. Surprisingly. Alex, I know you have a busy day. I appreciate you spending this time with us, telling us about the model, and thank you for your work on it. Of course. Thank you for using it. I appreciate it. Of course. Of course. Goodbye. Thank you. Have a nice day. Well , you heard it from the man himself. I think it's a really interesting combination of programming skills, creative thinking, taste—I don't know what to call it—all in one package. This is a really good model. Who has a demo? We just got a bunch from Mike. Katie, is there anything else you want to share? Is there anything we haven't talked about on the writing front? Not on the writing front, but my attempt at live production is still in the planning stages. So, okay. Are you, uh, good to know that she's so thorough. Can you show us the Hades Town render she made? Yes, of course. Um, I'll open it now. While you're at it, I'll say that this is a real trend right now. Everyone is creating 3D worlds. Kieran seems to have been the first to start doing this, back when technology couldn't really create 3D worlds. And now everyone is making 3D worlds in one go. Um, and I had to think about what kind of 3D world I wanted to see. So I had the models do 3D reconstructions of the Battle of Waterloo, which I would love to show, but my computer would just die if I did that. I urgently need a new computer, something like this. I'm on video calls all the time , and people ask, "Oh, you have bad internet?" And I say, "No, it's just some agent trying to create something and overloading my processor." "Um, so you can try Linux next time. Maybe I'll try Linux. You're really convincing me. Um, I think to be honest , I'm just waiting for the RAM shortage to end so I can get a better MacBook, but we'll see. Um, oh, we have a good question. Um, so B Mask Maker asks, "Katie, do you still prefer Codeex for writing or intellectual work?" ""And I have a good answer for that too. I'm curious what you think, Katie. Hmm, I'm currently debating whether to return to Claude or stay with Codeex. Codeex seems more familiar to me, because it's been a few weeks since I've been using Claude consistently, because , you know, I didn't find much use in the previous models. Um, but I joked in our chat that when I went back to the Claude desktop app , I felt like Gandalf: "I don't remember this place." Hmm, but I'm gradually getting used to it, and I'm having very good results for my writing style and voice with Claude, with Fable version 5.1; but, um, Soul is still very strong, so I'm a little hesitant, I'm in that situation again where I don't want to choose one thing. Um, because if I do n't like the result in one, I'll try another, and sometimes, well, sometimes I get a surprise that way. If we looked at how you use them today, which app do you spend the most time on? Today I am still more into Codeex. Hmm, as I mentioned in my test, I'm starting to use this again. I've been let down by a few models, but they're slowly regaining my trust. This is a process. I like it. Yes. Me, me too, I'm similar in that. I'm still at ChadBT most of the time . Um, but I've been such a devoted CHAGT supporter for the last six months that I think it's amazing that I'm now starting to think, "Oh yeah, I'll go back to Claude from time to time for bigger assignments." And yes, I agree. What I like about the 5.6 models and CHBT in general is that, in my opinion, their text is simply cleaner. He is so clear, direct. He's just, like, doing his job. Um, and I think Claude, and we've been talking about this for a while, has a little more variety , a little more creativity, a little more literary. And I think that this heterogeneity of it means that sometimes I think, "The Claude model would be a good fit for this." I really like the Claude model because it helps me understand my own thoughts. Um, but for direct sentence writing, I think Chad just does a better job with the clean style. And for my writing style, that's the most important thing. Yes. Inside Compound Writing, I found that the model captured my style better than Fable 5.1. So I think he's also a more flexible writer. Although I personally like Claude for my writing style. Hmm, if I'm writing a section for the context window, where it should usually be more descriptive, I find myself still reaching for Soul. Steeply. So, we come to the end of our time together on this live broadcast. However, if you are interested in Fable 5.1, if you want to learn more about it, if you want to hear everything we discussed and more, you should visit every.to. We have a full, in-depth written review on Every. We also have a library of prompts for Fable 5.1 so you can get the most out of Fable. And we also have this really cool graphic. Look at this. This is so cool. Um , we've been testing this model for the last week or so . You can read this with your agent. We have all these different sections. We'll tell you what Anthropic says. We will tell you what we are trying to achieve in the tests. Um, who we are, what we use it for, what we don't use it for , and how it's changed our approach. We look at everything from the coding, all the tests we ran, to how it changed our programming usage and everything we built. We went through the writing that we've already discussed on this stream. Um, but where does it fit in and how does it perform compared to Opus 5 and 5.6. Um, we're looking at intellectual work: from creating NPS dashboards to developing presentations . And we also figured out how it works inside the agent. So if you use this in Open Claw or something like that, you'll be able to learn how it works in Fable 5.1. And we also have our verdict. You should check this out if you are interested in AI. There 's a ton more out there, this is the only subscription you need to stay at the forefront of AI. Look how beautiful it is. What a nice home page. Ahem. Um, and there's a lot more exciting stuff to come, so stay tuned. We have a lot planned for the next few weeks. OpenAI Dev Day is in three weeks. So if you want to stay on the cutting edge of AI, you should subscribe to Every, and by the way, thanks Bassmaker. Um, we have camp on Friday . Um, on Friday, this Friday, we're doing live streams for subscribers only. We call them camps, and we'll discuss our use of Fable 5.1 in more detail. You will get a chance to ask questions, meet the team, and chat. So, if this is your first time here, welcome. We are Every, and we are glad to see you. If you come back to us, we love you too. We are thrilled to be able to share these moments with you . And we hope you have a great Tuesday exploring Fable 5.1. And remember, drink water, friends. Cheers. OK.