18:08AI Edge · Tutorial
How I Turned Claude Into My Personal Assistant
An 18-minute build diary of a custom personal OS: Telegram voice notes, live net worth, habit tracker, CRM, and the Supabase memory layer that ties it all together.
May 14thThe argument in one line.
Claude Fable is the most capable public coding model ever released, but its 2x cost and slow output speed mean the optimal strategy is surgically deploying it at the planning and QA endpoints of a build while running a cheaper model for the execution middle.
Read if. Skip if.
READ IF YOU ARE…
- You are actively building with Claude Code or vibe-coding products and want a realistic take on whether Fable is worth switching to immediately.
- You are a business owner trying to decide whether to increase your AI API budget now that the frontier model costs 2x.
- You want a concrete model-selection framework that maps specific task types to specific models.
- You are trying to understand the Mythos/Fable naming confusion and what the public release actually restricts.
SKIP IF…
- You want independent, reproducible benchmark analysis — this is a first-hours impression using Anthropic's own charts.
- You are on a fixed AI budget with no flexibility; the cost section surfaces pain without a cheap path forward.
The full version, fast.
Claude Fable 5 (the public version of the Mythos-class model) posts benchmark gains of roughly 2x on agentic coding and 6x over GPT 5.5 on frontier code, but ships with two catches: sensitive queries silently reroute to Opus 4.8, and after June 22 the model is API-only at 2x the token cost of its predecessor. The reviewer's answer is the 10/80/10 approach — use Fable for the first 10% (architecture/planning) and last 10% (bug checks/QA), and let a cheaper model handle the execution middle where most tokens actually burn.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:46
01 · Intro
Announces Fable 5 public launch, frames the video as an honest first-look review
00:47 – 01:15
02 · What Is Fable?
Anthropic describes it as a Mythos-class model with safeguards for general use; cybersecurity capabilities restricted
01:16 – 01:54
03 · Benchmarks
2x agentic coding improvement over Opus 4.8; 3x spatial reasoning; strongest gains on coding
01:55 – 02:36
04 · What's Restricted
Sensitive queries silently redirect to Opus 4.8 rather than being refused; cybersecurity and exploit topics flagged
02:37 – 04:02
05 · Blog Post Review
Screen-share of Anthropic blog showing frontier code accuracy vs cost chart; Fable dominates Opus 4.8 and GPT 5.5
04:03 – 05:31
06 · Agentic Coding Benchmarks
~6x over GPT 5.5 on frontier code; demos include Pokemon FireRed completion and solar system eclipse prediction
05:32 – 07:33
07 · Societal Implications
Drug design, space exploration, labor market discussion; human + AI competitive framing
07:34 – 10:30
08 · Early Feedback
Cursor, GitHub, Replit, Lovable endorsements; JPMorgan early access; misalignment score lower than Opus 4.8
10:31 – 14:27
09 · Third-Party Reviews
Matt Berman notes: verbose, slow, wildly autonomous, asks clarifying questions, effort-level trick for speed
14:28 – 16:35
10 · Is It Worth It?
Free on Pro/Max/Team until June 22; then API credits only at 2x Opus 4.8 price; $100 build cost estimates
16:36 – 18:51
11 · How I'm Using It
Fable = premium model for high-stakes tasks; Opus 4.8 = daily driver workhorse; task-type breakdown
18:52 – 20:34
12 · 10/80/10 Approach
First 10% Fable for planning; middle 80% Opus 4.8/Codex 5.5 for execution; last 10% Fable for bug-checks and QA
20:35 – 21:09
13 · Outro
Subscribe and newsletter CTA; promises ongoing Fable deep-dive content
Lines worth screenshotting.
- Fable 5 is roughly 6x more capable than GPT 5.5 on frontier code benchmarks, which Anthropic defines as the hardest real-world coding tasks.
- Sensitive queries on cybersecurity and exploit topics silently reroute to Opus 4.8 rather than being refused — the model never acknowledges the redirect.
- Fable is free inside Claude subscriptions only until June 22; after that it requires API credits at 2x the price of Opus 4.8.
- A build that costs $40-50 in Opus 4.8 API spend will cost roughly $100 in Fable, making full-project use prohibitive for most solo builders.
- Fable is wildly autonomous — significantly more willing than any prior model to work for hours unsupervised without human checkpoints.
- The model asks clarifying questions before acting: a single prompt triggers a Q&A loop, confirmation, then a spec, which slows output but improves result quality.
- Fable is reportedly slower to write and slower to code than Opus 4.8, which compounds the cost problem for iteration-heavy workflows.
- Opus 4.8 remains the better model for creative writing and casual language tasks — Fable over-explains and makes some readers feel genuinely dumb.
- The 10/80/10 framework: use the frontier model for the first and last 10% of a build; run a cheaper model for the middle 80% where tokens actually accumulate.
- Anthropic says 90% of its own code is written by AI — the implication is that AI acceleration of AI development is already a closed loop.
- The public Fable is a capability-restricted version of the full Mythos model that enterprise partners have been using for months.
- Competitors like Cursor, GitHub, Replit, and JPMorgan had access to the Mythos model before the public launch and have already shipped products built on it.
Deploy frontier models surgically, not universally.
WHAT TO LEARN
When a new frontier model costs twice as much and runs twice as slow, the smartest move is not to switch everything over — it is to identify the two decision points in every build where its reasoning advantage actually changes the outcome.
- The most expensive part of any AI-assisted build is not the planning session — it is the iterative middle, where you go back and forth fixing mistakes and generating assets. A cheaper, faster model belongs there.
- A model's benchmark advantage is most valuable at the two leverage points of a project: the architecture decision at the start, and the vulnerability and QA scan at the end. That is where the premium model earns its cost.
- When a model asks clarifying questions before acting, that is not friction — it is the model recognizing that a poorly scoped prompt produces a worse outcome, which saves rework time downstream.
- Slow output speed is not just an inconvenience; it compounds into a real time-cost at scale. If a build takes five hours instead of 45 minutes, the relevant question is whether the quality difference justifies the wait.
- A model that outperforms on coding may still underperform on creative writing — different training objectives produce different output textures, and using the frontier model for everything often means using the wrong tool for half your tasks.
Terms worth knowing.
- Fable 5
- The public-release name for the Mythos-class model from Anthropic. Mythos is the capability tier; Fable is the specific model within that tier made available for general use.
- Mythos class
- Anthropic's internal designation for its highest-capability model tier, previously held back from public release due to cybersecurity risk concerns.
- Frontier code
- Anthropic's benchmark category for the hardest real-world coding tasks — complex, multi-file, long-horizon problems that simpler benchmarks do not capture.
- Agentic coding
- AI-driven coding where the model works autonomously across multiple steps — planning, writing, debugging, and iterating — without a human approving each action.
- Vibe coding
- A colloquial term for building software primarily by prompting an AI model in natural language rather than writing code directly.
- Harness
- A wrapper or orchestration tool that manages an AI model's access to tools, context, and memory during long autonomous runs.
- API credits
- Pre-purchased token budget on the Anthropic API, required to use Fable after the subscription free-access window closes on June 22.
Things they pointed at.
18:50toolHermes Agent
18:55toolOpenClaw
15:50productCodex 5.5
Lines you could clip.
06:25
“Fable five absolutely wipes the floor with Opus 4.8.”
09:10
“It's not just human versus human — it's human and AI versus another human and AI.”
09:11
“In six months time, this is gonna be one of the crap models.”
14:16
“It felt like it wanted my hardest tasks and anything less wasn't good enough.”
15:32
“I was already racking up $40-50 bills a day at some point with OpenClaw. This will now cost you $100.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphoranalogy
00:00It is here. This is one of the biggest days in AI history. Certainly, the biggest day for AI of the entire year because Claude Fable is out, previously known as Claude Mythos.
00:11The model so good that it scared governments. The retail version is now officially live in Claude. You can now go into Claude and use it.
00:19I've been using it for the past few hours and, boy, am I impressed. There are some caveats as well. So today, I wanna give you my full in-depth honest review so far and first look at Claude Fable five.
00:32I'll run you through everything that you need to know. I'll run you through how I am personally switching my AI usage now that this new model is live. So by the end of the video, you'll know exactly how you can use it and get the most out of your Claude subscription.
00:46So let's begin by diving into what Claude Fable five is. Anthropic describes it as a mythos class model that they've made safe for general use. We know there was all of the conjecture around the last mythos model not being safe for the public because it could literally go through code bases and potentially exploit code bases.
01:05So they've basically muted this model. They've implemented some safeguards to mean that it can't be exploited. Its capabilities exceed those of any model they've made available before.
01:16The benchmarks are, of course, very, very impressive. It beats Opus 4.8 by a mile, especially when it comes to coding. That is really where this model is going to shine.
01:25For Agentic coding, it's almost a two x improvement. And even for things like spatial reasoning, it's almost a three x improvement. Knowledge work has an improvement.
01:34So pretty much across the board, this is a better model, but it's very clear that the biggest leap in performance comes specifically on the coding side. If you're using Claude code, if you're building applications, if you even build a personal operating system or you consume some of my content, obviously, the builds will now be a lot more effective and accurate if you do choose to use Claude Fable five.
01:54Claude says releasing a model this capable comes with risks. Without safeguards, Fable five's capabilities in areas like cybersecurity could be misused to cause serious damage.
02:03So queries on a narrow range of topics will instead receive a response from our next most capable model, Opus 4.8. This is absolutely fascinating. So instead of training the model to answer things differently, they're just basically going to scrap usage altogether for the public on Fable five and instead redirect you to Opus 4.8 if your question involves certain topics.
02:26And these topics are probably going to include things like cybersecurity, exploiting applications, looking for attack vectors, even potentially reviewing some code.
02:36You may run into some roadblocks. This is something that I'm really gonna be testing over the next few days as I get my hands dirty. So make sure you subscribe because there'll be tons of Claude Mythos content coming your way.
02:46I'm gonna show you my use cases. I'm gonna show you the changes that I'm making to my AI workflows and some of the cool things that you can do with this brand new model that'll be coming over the next few days. So make sure you do subscribe and turn on notifications.
03:00Because when a new model like this releases, when new technology comes out, there is such an advantage being early. You wanna be ahead of everyone else. You wanna be at the cutting edge, especially when it comes to AI.
03:10You wanna continue to use it to level up your life and level up your business. So I think just as a general practice in life, it's great to be early. It's great to be on the cutting edge.
03:18That's why we called the channel AI Edge, and I'll be here to help you absolutely maximize the usage out of your cloud subscription with Fable five. So let's get off x and actually look at their blog post because there's much more details in here, and it really exhibits how smart it actually is.
03:32Like, if you look at the frontier code, accuracy versus cost, you can see Fable five . You can see that every single accuracy level comes in higher, low, medium, high, extra high, and max than the previous model by quite a mile.
03:49And they've also plotted out GPT five point five's performance, of course, according according to their proprietary scoring mechanisms, which I'm sure would be different from OpenAI. But nonetheless, it's very clear that it is a better model than GPT 5.5 at coding. And in terms of agentic coding, the biggest jump is actually in frontier code.
04:07So on the benchmark test, it is the best model. It's better than 4.8, and it's, of course, better than GPT. But on Frontier code, it comes in over a five x, almost a six x more powerful than GPT 5.5.
04:20Like, this is a massive leap in improvement. You can't really underestimate just how big this is in terms of a model release. If you're a coder, you're gonna have a field day with this.
04:29And not just coders because, you know, you can use Claude code for building workflows. You can use Claude code for running agentic for running agentic flows across your business using Fable for these use cases is going to probably get you much more accurate results. And, obviously, this is something over the next few days I'm really keen to experiment with.
04:47I've already done a little bit of experimentation, but because of some of the limits, which we'll discuss, I haven't been able to fully test it to the max. They're still rolling it out in Claude code and rolling it out across the API system.
04:58Let's look at some examples now. So firstly, for knowledge work, Fable five shows a strong performance on complex analytical tasks. So, you know, if you are analyzing a spreadsheet, if you are doing really important, you know, business strategy or personal strategy, Fable five is gonna be really good at this kind of work.
05:14It has really smart advanced reasoning. So it's not just for coding. It's gonna be the smartest model in a variety of areas.
05:19You know, if you're coding a game or coding graphics, you can see that Fable is really good. It actually beats Pokemon's fire red. So it actually goes through and completes the entire game on its own.
05:29This is a a demo that's completely Claude controlled. So it's smart that it can even complete a video game for you.
05:35I mean, in the future, these models are gonna be so smart. It could probably complete an even more complex game like, you know, GTA for you. Of course, you would probably never use that in in your normal life because that defeats the purpose of a video game, but that just shows you that it has advanced reasoning.
05:50That just shows you that it can do something that, you know, some humans used to struggle with. Like, I used to play Pokemon as a kid. Used to get stuck with this kind of stuff.
05:56It can now, you know, complete it. So we're moving towards a world where clearly the artificial intelligence, the AI models are smarter than us, and that is just the new reality we have to accept and instead learn how to leverage these models in our daily lives to become more productive. I really think we're entering a world where it's not just human versus human, but it's human and AI versus another human and AI.
06:19If you are better at AI, if you are better at running agents, if you understand AI better, you're gonna have a massive advantage across all your competition if it's in business or just in the workplace or, you know, in your specific field just in general. You can also see it's really good at coding.
06:34It simulated the entire solar system and actually predicted a solar eclipse. I mean, some of the visuals look fantastic. I've been experimenting a little bit with what I've been able to.
06:43And, you know, in terms of coding visuals, coding UIs, I think it's gonna be amazing, especially paired with Claude design, which is all also industry leading. You know, if you create websites, if you do graphics, I still think Claude, you know, it's always been the best model on the creative front, and this is only going to cement its place on the design front.
07:00You could see it's even being used for drug design. Models like this, you know, I I do think people get caught up in, you know, the narrative of the labor market and people losing jobs, and I totally understand that. But you have to understand, as these models get smarter, and this is only the beginning.
07:13AI has only been around, you know, for a few years. But as they get smarter, there's gonna be genuine drug discoveries. It's gonna be able to solve some of the most complex issues in society.
07:20Like, these are tools which can really revolutionize the human race. So although there are definitely risks, although there are definitely issues being created, not denying that side of things, it is also going to open up a new world of possibilities in medical, in science, in space travel, you know, rocket science, pretty much any advanced industry.
07:40And I think this is a really exciting thing about these models and the the thing that I'm personally really excited for, you know, in terms of humanity as a whole. You know, clearly, it's been able to solve already some complex problems some companies have have been struggling with. In fact, you go down and look at some of the early feedback, Cursor said that Claude Fable is a state of the art model.
07:59It's opened up a class of long horizon problems that were out of reach from earlier models. GitHub said it's a massive step forward. These are all companies that got access to the Mythos model before anyone else.
08:08A lot of big companies, Replit, Lovable. And we also know that, like, JPMorgan and some of the investment banks have already been using the the Mythos model, and now it's available for the first time to the public. It's really interesting because the biggest risk that people flagged about Mythos was the fact that it could, you know, potentially be used in the wrong hands for, you know, exploits, hacks, you know, all this bad stuff.
08:30But interestingly, it actually has the lowest misaligned behavior at least that they're reporting of any of these models actually lower than OPUS 4.8. Clearly, some of the safeguards that they've also implemented, which we discussed, will help it, you know, be even more safe for consumer use.
08:44The crazy thing is that this is technically a dumbed down model though compared to, you know, the full Claude mythos that other companies have had access to. So the fact that Claude has something even better is kind of crazy to me. And the other crazy thing that I think about is, you know, at a certain point, Opus 4.6 was like the best model that we were all using, and it was like, you know, absolutely mind blowing at the time.
09:04You know, then it was 4.8, then, know, at a certain point, was also GPT 5.5. And now we have this. In six months time, this is gonna be one of the crap models.
09:12Like, you know, you already look back at four point four point six and or 4.5, which was groundbreaking at the time, and you're kinda like, you know, it's meh now. I wanna use the the Frontier model. In six months, this is going to be a meh model, and we're gonna have models that are just so so so much better.
09:27So the rate of improvement here is, I think, the really crazy thing. The fact that we already have massive jumps, if you wanna go back and look at the benchmarks, like such a big jump on frontier coding. You know, what happens in six months, twelve months, eighteen months?
09:38This is just accelerating. AI just helps AI to code faster. I think Anthropic said 90% of their code is actually written by AI.
09:45This is just . So, I mean, if you aren't dedicating a significant amount of time to learning how to use this stuff, I don't know what to tell you. But, obviously, if you're watching this channel, you're probably trying to figure that out.
09:58So, you know, I'll be coming with a lot of content to help you guys do this. Another thing that I recommend is that you actually sign up for my newsletter. I'll leave a free link in the description below because we give you full builds like we did, for example, a tutorial in the newsletter of how to build your first plugin.
10:12So the newsletter is for just practical guides. We spit game once a week. I truly believe it's one of the best newsletters in the AI space.
10:19I know there are a lot of kinda crappy AI written newsletters out there, but we genuinely take our time with this. We curate it as a team, and I think you'll get a lot of value from it, especially if you like the YouTube content. So I'll leave a link in the description below if you would like to check it out.
10:31So now let's look at some opinions from some other people I respect in the space that have had even longer than me to use it. And then I would like to talk about whether it's worth it because I have done some thinking about how I'm actually going to use Fable and where I'm still gonna use some other models.
10:47So I do wanna talk about that with you. But first, Matt Berman, who's been testing this for the past week, he got early access to it. He says workflow mode is the standout.
10:55He taught it to do a full code review and want to spin up hundreds of agents in parallel. This is gonna be really, really cool, actually, you know, building these multistep workflows with AI agents. It can basically assign an individual agent to every file in an application to find bugs, edge cases, missing documentation.
11:11I mean, VibeCoders are gonna be able to build apps even quicker now. You basically be able to build an app in an hour. One comment that he mentioned that I've already found in my test is that it's wildly autonomous, so it's way more willing than any previous model to go off and work for hours at a time on its own.
11:26So once you actually strap in a harness like Hermes agent, which I'm gonna do a video about soon, or Open Claw, the possibilities to have it just self improving over time and actually working on its own code while you sleep, I mean, it's just gonna be insane. I I just can't wait to test it out even more.
11:42The problem is gonna be cost. We're gonna speak about that soon because it does come with a big caveat. I'm gonna be honest, and it's gonna make it challenging for some people.
11:50I'm gonna discuss exactly what the pricing structure looks like soon. You you will have access though for a couple of weeks to this model before Claude makes a big change, which we'll discuss.
12:00Matt says it's incredibly verbose. Explanations get deep in the weeds fast. This is something that I noticed.
12:06I actually kind of prefer the language and the creative writing still of Opus 4.8. So I'll show you where I'm still using 4.8, and one of the areas is definitely in writing. I do think sometimes that, you know, over words things.
12:21I think for some creative writing, it could be really, really good, but that is something that is probably gonna be tweaked over time. Matt actually said the way it explained things made me feel genuinely dumb. Yeah.
12:32That that is definitely a a drawback of the model, but I think that only just speaks to how smart the model actually is. He mentioned Fable loves clarifying questions. A single prompt turns into questions, a summary of answers, confirmation on the summary, then a spec.
12:45This is how I love to use AI. Like, I always like to get AI to ask me questions, answer those questions, get AI to ask me questions about my questions because you really need to give the AI context. So to see it asking clarifying questions to do a better job, as Matt pointed out, it just shows that the AI now has more willingness to actually help you, which I do think is an improvement because some of the models, you know, you give it a prompt and then they just go away and and they and they just do stuff, but you'll actually get a worse product effect.
13:10Especially for people that aren't as AI native, maybe we'll get more hand holding for some certain tasks. I think that'll be really beneficial. A comment he makes, which is something I a 100% have experienced already is that it does feel slow.
13:22It is damn slow. It's slow to write. It's slow to code.
13:25It's expensive. We're gonna get into that. And this is one of the reasons why I'm not switching all of my work or all of my prompting over to Fable because it just is really, really slow.
13:34And I've actually come up with a new approach to use this model, which I think is way better, which I'm gonna discuss with you guys today as well because I think it's really gonna, you know, help you out if you do choose to code on this. Matt's conclusion was that he turned down the effort level even more than you think you need to because it helps a lot with the speed.
13:51All of these quirks are fixable. Model optimizations and increase compute capacity to increase speed, more fine tuning, which, of course, we'll get over the next weeks. I'm sure we get 5.1, 5.2, etcetera, and it will help with verbosity and being overly cautious to make decisions.
14:05His ultimate verdict is that it's extremely capable. He's still figuring out how to get the most out of it. That's something that I'm gonna go real deep into over the next couple of weeks.
14:12It felt like it wanted my hardest tasks and anything less wasn't good enough. This is the first look at a brand new training run and it's already the most capable model I have ever used. That's the part I can't stop thinking about.
14:24It is absolutely crazy. It's it's just so so smart. But now let's talk about is it worth it, and then let's talk about how I'm actually gonna use it.
14:32Because these are things that obviously really, really matter perhaps more than anything else I've spoken about in today's video because we do wanna get practical here. So here's the kicker. Anthropic said it will make Fable five available to all users on its pro max team and seat based enterprise plans at no additional cost.
14:49Here's the annoying thing. Until June 22, so you really only have a couple of weeks to use this for no additional cost within your existing Claude subscriptions.
14:59After that, users will only be able to access the model by buying and utilizing extra computing credits, which basically means you'll need to use the API, which is doable. You you can just purchase, you know, credits through that.
15:12If you're using a harness, you can go onto the Claude site and actually add a budget. But because this is twice as expensive as Opus 4.8, the credits are gonna be much more expensive than the previous models.
15:25So it's basically two times more expensive. So if you're doing coding, you know, I was already racking up $40.50 dollar bills a day at some point with OpenClaw. This will now cost you a $100.
15:33So, obviously, it depends on how much you're doing. But if you were coding an application, like, let's say you built my personal OS dashboard, which I did, you know, a video about from start to finish just with Fable with no other model contributing, I see that potentially as a 100 to $150 task, which is obviously exorbitant and is gonna price a lot of people out of the market.
15:54Now Anthropic's aware of this. They said that they aim to incorporate Fable into as many subscription plans as quickly as they can. But, of course, they don't have unlimited compute.
16:02Even with Elon helping them out with his compute, which is probably one of the reasons this model was able to be released, they still didn't have enough compute. And this is probably going to be, you know, a a one, two, three month thing before everyone has access. So I'm just gonna be frank with you.
16:16If you do have capital, you do have a big advantage. So if you're a business owner or someone that is willing to spend maybe a few thousand bucks to really use this, you're gonna probably get a massive leg up. And if you are a business owner, I probably suggest, not financial advice, that I probably suggest you you consider upping your AI budget.
16:31If you aren't, you're just gonna have to accept that you'll have to use this in a limited capacity. And personally, in terms of the capacity I'm using it in, this is what I'm going to be doing.
16:41So I'm basically treating Fable as my premium model, and I'm treating Opus 4.8 as the workhorse. As I said earlier, it's kinda crazy that we're talking about, like, 4.8 now as a workhorse model, and Opus 4.6 is like a cheap model, but that's just the nature of the game. And probably GPT is gonna come with a response to this.
16:58You know, they, Mythos, now called Fable. Mythos is the class. Fable is the specific model, was coming for a long time.
17:04They they've probably got an ace up their sleeve as well. So, you know, this constant competition keeps, you know, kicking these guys to one up each other.
17:12But Fables are spent essentially like an expensive high stakes model. You only wanna use it on the highest stakes tasks. And then OPUS 4.8, you wanna use it as your daily driver.
17:21For example, for business strategy, when I need the correct response on a decision, you know, potentially something legal, potentially something strategic that that could have, you know, massive financial implications. I want and I said this in my video, like, I was losing trust in Claude.
17:37I want the smartest model in the world to give me business strategy. I'm not gonna trust any other model. I always wanna trust the best model.
17:43So that is a case where Fable is probably worth it. Anything mission critical, so any build that I I need done correctly. If I'm working on vibe coding, my team, you know, is helping us build products for AI Edge.
17:55That is obviously going to use Fable at at least following the strategy I'm gonna talk about in a minute because I think it's very important that the best model does the most important stuff. For in-depth coding and refining coding, I'll use Fable. And for planning and final bug checks, I'm gonna run a strategy which I'll speak to you about in a second, which basically is going to so I make sure Fable touches the most important parts of each specific build.
18:22Now for Opus 4.8, it still, I think, speaks in a more understandable and more human manner, which is better for creative work. So a lot of writing, you know, copy on pitch decks, etcetera.
18:32I'm gonna keep using 4.8 for quick prompts, maybe light strategy. I'm still gonna use 4.8 for, you know, brain dumping voice notes and condensing information and quickly reviewing information and working on spreadsheets, I'm still gonna use 4.8 unless the math is really, really important. And for the majority of my builds, I'm still gonna use 4.8 or Codex 5.5, and I'm essentially gonna follow the following approach.
18:56It's called the ten eighty ten approach where the first 10% is gonna be used by Fable. So this is you know, let's say I'm doing an OS build. Right?
19:04An operating system build. Fable will do the first 10%. It's gonna make sure the planning is really strategically thought out because it's a very smart model.
19:11Then I'm gonna use the execution. So the grunt work, the building, the back and forth on Opus 4.8 or Codex 5.5 are cheaper, but still very good coding model.
19:20Then once it's done the majority of the work, the last 10%, I'm gonna switch back to Fable and I'm gonna use it to do bug fixing. It's gonna find potential vulnerabilities. It's going to suggest tweaks.
19:32And this way, I have Fable touching the strategy. I have it touching the final execution, but I'm not absolutely , which to be honest is where the majority of the tokens are burned.
19:45Going back and forth, editing mistakes, you know, creating graphics, all of the stuff especially with operating systems and agentic workflows for business, that is the stuff that really kills you. So the ten eighty ten approach is an affordable way to use Fable. And not only affordable because sometimes I have higher budgets for certain projects, but just a faster way to get things done.
20:05If you use Fable for the entire process, it's going to take a long time. And some things, you don't wanna take a long time. You wanna do an OS build in forty five minutes.
20:12You don't wanna do it in five hours. You wanna do I I put a lot of trading bots as well to automate some of my trading execution. I want I wanna spin something up in thirty minutes, not an hour, hour and a half.
20:21So this is where this approach can come in. You maximize the best strengths of Fable whilst, you know, getting cost down and and getting speed down by using Opus 4.8. But this is how I'm using it at the moment.
20:32It's likely to change over the next few weeks, so I will keep you up to date. Make sure you do subscribe to the channel so you get all the latest content from me. I'm gonna be absolutely smashing the content on Fable as I continue to test it out.
20:43Also, if you haven't already, you can check out the newsletter down below where I'm actually gonna share my builds with you as we've been doing over the past few weeks. I truly believe we've built one of the best, if not the best newsletter in AI. And if it's not the best, I'm gonna make it the best because I I want everything I do to be the best for you guys.
20:59And, yeah, hopefully, you enjoyed today's video. I'll see you in the next one. Good luck, and enjoy using this crazy, crazy model that came out today.
21:07See you later. Peace out.
The Hook
The bait, then the rug-pull.
On launch day for Claude Fable 5 — the public version of the model Anthropic had been quietly deploying to enterprise partners for months — AI Edge creator Miles Deutscher sat down with a few hours of hands-on time and a clear question: is this actually worth it, and if so, how do you use it without burning your budget?
Named ideas worth stealing.
18:52model
The 10/80/10 Approach
- First 10%: Fable for planning and architecture
- Middle 80%: Opus 4.8 or Codex 5.5 for execution
- Last 10%: Fable for bug-fixing and QA
A token-budget framework that deploys the frontier model only at the high-leverage planning and QA endpoints of a build, running a cheaper model for the execution middle where most tokens burn.
Steal forAny agentic coding workflow where cost and speed matter
16:36list
Fable vs Opus 4.8 Task Split
- Fable: business strategy, mission-critical builds, legal/financial decisions, in-depth refactoring, final bug checks
- Opus 4.8: creative writing, quick prompts, voice note condensing, spreadsheet work, daily driving
A use-case matrix for deciding which model to invoke based on stakes and task type rather than defaulting to the frontier model for everything.
Steal forModel-selection policy for teams or solo builders with multiple models available
How they asked for the click.
VERBAL ASK
20:35newsletter
“Make sure you do subscribe and check out the newsletter down below”
Double CTA — YouTube subscribe and paid newsletter (aiedgehq.co). Mentioned multiple times throughout but hard push in final 90 seconds. Somewhat buried behind the framework reveal.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
