13:23Peter Yang · Tutorial
How I Turned Codex Into My AI Life Coach in 13 Minutes
A 13-minute walkthrough of the four plain-text files that give an AI enough context to provide genuinely useful life and career advice.
June 17thThe argument in one line.
GPT-5.6 and Claude Fable 5 are close enough in raw capability across six real build tasks that the deciding factor for a daily driver ends up being cost and rate limits, not intelligence.
Read if. Skip if.
READ IF YOU ARE…
- You already pay for both a ChatGPT plan and a Claude plan and are trying to decide which one earns the daily-driver seat.
- You build with AI coding agents (Codex, Claude Code) and want a recent, hands-on read on browser/computer-use reliability, not just a benchmark chart.
- You can only afford one AI subscription and need a practical tie-breaker beyond raw model quality.
SKIP IF…
- You want a rigorous, controlled benchmark methodology -- this is one builder's subjective build-and-compare, not a formal eval.
- You don't use AI for coding, app-building, or browser automation -- most of the comparison is production/dev-workflow specific.
The full version, fast.
GPT-5.6 is priced about 50% cheaper than Claude Fable 5 on input tokens and rarely hits rate limits on a Max plan, so Peter Yang opens by framing this as an economics question, not just a capability one. Across six hands-on builds -- an interactive travel site, a one-shot 3D Star Fox-style game, browser/computer-use video publishing, a mobile app feature, life/business advice, and an audit of his own AI-skills repo -- the two models trade wins roughly evenly, with GPT-5.6 clearly ahead only on browser-use reliability and Fable only clearly ahead on remembering an unrequested game mechanic and surfacing a sharper business insight. The conclusion: GPT-5.6 becomes the daily-driver engineer for its price and rate limits, with Fable reserved as the planner/architect/designer for anyone who can run both subscriptions.
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 · Cold open
Peter Yang frames the video's central question -- GPT-5.6 or Claude Fable 5 -- and recaps GPT-5.6's three-tier lineup (Sol, Terra, Luna) plus its narrow coding-benchmark edge over Fable at ultra mode.
00:46 – 01:45
02 · Model tiers, price, and coding performance
GPT-5.6 Sol is shown priced about 50% cheaper on input and 40% cheaper on output than Claude Fable 5, and Peter argues OpenAI's real edge is efficiency and generous rate limits, not raw model quality.
01:45 – 05:44
03 · Use case 1: interactive travel website
Both models build an interactive itinerary site from a Google Doc. GPT-5.5's earlier attempt looked functional but generic; GPT-5.6 adds a 3D torii-gate hero and closes most of the design gap with Fable, landing the round as close to a tie.
05:44 – 09:39
04 · Use case 2: 3D Star Fox-style game
Each model one-shots a playable browser game in roughly 10-15 minutes. Fable gets a slight edge for remembering to implement the barrel roll, an unrequested but genre-defining move GPT-5.6 left out.
09:39 – 15:10
05 · Use case 3: browser/computer-use video publishing
GPT-5.6/Codex clips a podcast, adds captions, and autonomously navigates YouTube Studio, TikTok, and Instagram to upload and schedule posts across 20+ steps without errors; Claude Fable via the Chrome extension finishes the same task but more slowly, relying on screenshots to decide where to click.
15:10 – 19:34
06 · Use case 4: mobile app feature build
Using a shared design.md, component library, and a 'compound engineering' planning skill, both models produce nearly identical nutrition-tracking tabs for a fitness app -- evidence that a written design system narrows model differences more than model choice does.
19:34 – 22:17
07 · Use case 5: life and business advice
Both models read Peter's personal strategy document and give candid advice. GPT-5.6's answer is better organized; Fable surfaces a sharper, non-obvious insight about a broken YouTube-to-Substack funnel.
22:17 – 24:37
08 · Use case 6: personal AI OS repo review
Both models audit Peter's messy personal-automation repo and independently land on the same recommendation -- consolidating four separate sponsor-related skills into one -- with neither model clearly ahead.
24:37 – 26:36
09 · Strengths and weaknesses recap
A crowd-sourced pros/cons summary: GPT-5.6 is relentless and reliable but slightly weaker on intent and front-end polish; Fable is wiser and better at design but expensive and rate-limited.
26:36 – 28:29
10 · Daily-driver verdict
Peter's final call: GPT-5.6 is his daily driver for its price and rate limits, with Claude Fable 5 kept as planner, architect, and designer for anyone who can run both subscriptions.
Lines worth screenshotting.
- GPT-5.6 Sol is priced about 50% cheaper on input tokens and 40% cheaper on output tokens than Claude Fable 5 ($5/$30 per million vs $10/$50 per million).
- On a ChatGPT Max subscription, rate limits stop being a daily concern, while heavy agentic browser/computer-use work makes Claude Fable 5's usage limits and API cost bite harder.
- GPT-5.6 closed most of the front-end design gap that made GPT-5.5 look far worse than Claude models -- a travel-site build that was a near-tie between GPT-5.6 and Fable would have been a blowout for Fable under 5.5.
- Explicitly asking a model to add a 3D/WebGL hero element is a repeatable trick for making an AI-generated web page look more animated and expensive.
- Claude Fable 5's Star Fox-style game remembered to implement a barrel roll without being asked, a genre-signature move GPT-5.6's build omitted despite matching it on shooting, power-ups, and a boss fight.
- GPT-5.6/Codex completed a 20+ step browser-use flow (uploading and scheduling one clip across YouTube Shorts, TikTok, and Instagram) with zero mistakes across weeks of use.
- Claude Fable 5 completed the same multi-platform publishing task through the Chrome extension but ran slower, because it relies on taking screenshots of the page to decide where to click next.
- The reviewer flags that the browser-use gap he observed may belong to ChatGPT/Codex vs Claude Code as apps, not to GPT-5.6 vs Fable as models -- an attribution he explicitly can't resolve.
- Given the same shared design.md and component library, GPT-5.6 and Fable produced nearly identical mobile app UI -- a written design system narrows the gap between models more than model choice does.
- For a strategic advice task, GPT-5.6's answer was better organized while Fable surfaced a sharper, non-obvious insight: a broken YouTube-to-Substack funnel the creator hadn't noticed.
- Both models independently recommended the same fix when auditing a messy personal-automation repo -- consolidating four separate sponsor-related skills into one.
- The reviewer believes model writing quality has quietly regressed as both OpenAI and Anthropic optimize further for coding, rating Claude Opus 4.6 as a better writer/editor than Opus 4.8.
- Default guidance for both models: run at medium or high reasoning effort, not ultra/very-high, unless a task is genuinely stuck -- the top settings burn tokens without proportional quality gains.
Price and reliability beat marginal intelligence for daily AI work.
WHAT TO LEARN
Across six real build-and-compare tasks, GPT-5.6 and Claude Fable 5 traded wins roughly evenly, so the deciding factor for a daily driver ended up being cost and rate limits, not raw capability.
01Cold open
- A benchmark chart or price sheet only tells half of a model's story -- real adoption is decided by hands-on use across the tasks you actually do.
- When two frontier models keep releasing counter-updates within weeks of each other, treat any 'best model' verdict as a snapshot, not a permanent ranking.
02Model tiers, price, and coding performance
- Price and rate limits are legitimate product features: a model that's 50% cheaper and rarely throttles you can beat a marginally smarter model for daily, high-volume use.
- Before picking a default model for agentic or high-token workflows, check per-million-token input/output pricing side by side, not just headline benchmark scores.
03Use case 1: interactive travel website
- Explicitly asking a model to add a 3D/WebGL hero element is a repeatable trick for making an AI-generated web page look more expensive and animated.
- The design gap between model generations can close fast -- don't assume a model that lost a design comparison months ago still loses today.
04Use case 2: 3D Star Fox-style game
- A single well-specified prompt (mechanics, obstacles, power-ups, boss fight) can produce a playable 3D browser game in 10-15 minutes with either frontier model.
- Small, unrequested details -- like remembering a genre-signature move -- are often what separates a merely functional AI build from one that feels considered.
05Use case 3: browser/computer-use video publishing
- Browser-use and computer-use agents are worth the setup cost for any workflow that touches a poorly-APIed platform, like social upload flows or compliance forms.
- Reliability, not just capability, matters for agentic multi-step tasks: an agent that completes 20+ steps without a single mistake is more valuable than one that's occasionally smarter but needs supervision.
- When comparing two products' agent behavior, separate the model from the app wrapped around it -- a slowdown might be the browser extension's approach, not the underlying model.
06Use case 4: mobile app feature build
- A shared design-system document plus a component library makes outputs from different AI models converge -- invest in the design doc, not just the prompt, if you want consistent UI across tools.
- When you only need to validate a UX direction, tell the model to use local or mock data instead of building the full backend -- faster to review and cheaper to iterate.
07Use case 5: life and business advice
- Give an AI advisor your real goals, financials, and past conversations, and ask it to be candid rather than encouraging -- the useful output is the blind spots, not the reassurance.
- Different models can surface genuinely different insights from the same input data -- worth running important strategic questions past more than one model.
08Use case 6: personal AI OS repo review
- Periodically asking an AI to audit your own tool or skill collection for redundancy is a low-cost way to catch scope creep before it compounds.
- When two independent models converge on the same recommendation, treat that agreement as a stronger signal than either verdict alone.
09Strengths and weaknesses recap
- When you can only afford one subscription, the deciding factor for daily-driver work is often the boring stuff -- rate limits, cost, and reliability -- not which model wins more comparisons.
10Daily-driver verdict
- It's reasonable to split tools by role: use the more expensive, design-forward model for planning and architecture, and the cheaper, more reliable model for daily execution.
- Model quality on any single axis (writing, design, reasoning) isn't static -- labs trade off strengths as they optimize for whatever axis is most competitive at the moment.
Terms worth knowing.
- Browser use / computer use
- AI agent features that let a model directly control a web browser or a computer's local files and apps instead of only calling APIs -- used here to upload and schedule videos on platforms with no good API.
- GPT-5.6 Sol / Terra / Luna
- The three-tier model lineup within OpenAI's GPT-5.6 release, with Sol positioned as the tier competing most directly on coding performance.
- Compound engineering
- A free, shared library of AI planning prompts and skills, available on GitHub, used to generate more thorough and consistent build plans before an agent starts coding.
Things they pointed at.
16:33toolCompound engineering skills repo (GitHub)
Lines you could clip.
01:36
“GPT-5.6 is about 50% cheaper than Claude Fable 5 based on API pricing.”
24:46
“It's got a dog in it -- it will grab your problem by the throat and it will not let go.”
26:36
“GPT-5.6 is my daily driver... I would think of Fable as my planner, my architect, and my designer, and GPT as my engineer.”
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.
metaphor
00:00Hey, everyone. OpenAI just made GPT 5.6 available to everyone, so I wanted to answer your most burning question. Is GPT 5.6 or CloudFable five the best AI model in the world?
00:13And the answer is it depends. Let me show you what I mean by testing both models head to head across six different use cases. Building an interactive travel website, creating a three d retro space shooter like Star Fox, editing and publishing short videos programmatically, adding a new feature to your mobile app, getting life and business advice, and finally upgrading my personal AI OS.
00:39At the end of this tutorial, I'm gonna give you a clear recommendation of what model I'm gonna use as my daily driver moving forward. So just a quick recap, GPT 5.6 comes with three models, Soul, Terra, and Luna.
00:55And based on this terminal bench chart, GPT 5.6 Soul is slightly better than Fable on coding, especially when it's set to ultra mode. But OpenAI's biggest competitive advantage is actually not model quality.
01:11Instead, I think it is how efficient OpenAI's models are. G p 5.6 o is about 50% cheaper than CloudFable five based on API pricing. And when I use my ChatGPT Max subscription, I don't even think about my rate limits at all because I almost never hit them.
01:30And I think this is a really big deal because increasingly, we're gonna see both consumers and enterprises look for not only the best, but also the most cost effective model.
01:42But these benchmarks only tell half the story, so let's run both models through our six real use cases. Let's start by comparing GPT five point six's front end capabilities to Fable.
01:54Now I'm planning a trip to Japan in December, and here's my itinerary. Looking at all these bullet points is boring, so let's ask GPT 5.6 to do this.
02:04Build a beautiful interactive travel website from the most recent Japan itinerary that I have in my Google Doc, organize it by day, for each day include a date, area, activities, and dinner, plus images for major attractions, make it feel like a real travel catalog that I'll share with my family.
02:25And here is a useful line. So one trick to generate beautiful web pages is to ask AI to add a three d and WebGL element to the hero so that it looks more immersive and animated. Now let me show you what GPT 5.5 generated the previous version of the model, then we'll see 5.6, and then finally, we'll take a look at Fable.
02:47Alright. So here is the page that 5.5 generated and it is perfectly functional.
02:53You scroll down here, you can see that it has the travel schedule, itinerary, arriving in Tokyo, it has going to animal cafes, seeing Tokyo lights, and so on, and it goes down here.
03:06Right? So this is perfectly functional, but I think some of the images that I picked seem pretty out of place and boring like this one.
03:13And also on the hero image, the three d route that I made seems kinda random. Right?
03:19It doesn't really make any sense. So that's GPT 5.5. Again, that's the previous version of the GPT model.
03:26Now let's look at 5.6. As you can see, we now have a beautiful three d tori gate that's somewhat interactive, and then there is a moon or a sun in the background.
03:37And let's scroll down, and the rest of the cars look pretty great too. Right?
03:41The journey begins. Hello, Shinjuku. It found more relevant images like Meiji Shrine to go to and so on and so forth.
03:50And in some cases, it even included multiple images, although I can see here that the image got cut off. And let's keep going down and, yeah, it looks great. This is something that's perfect that I can show my wife and the rest of my family to get them excited for their trip.
04:05And I also like some of the coffee here. Steam in the cold air, no hurry, nowhere else to be. It kind of creates a nice mood for me to go visit some onsens in December.
04:16Now let's take a look at the website that Fable generated. So here we are, Japan in winter lights. So it has a beautiful image in the background, and it looks like it used WebGL to generate these snowflakes that look awesome.
04:30Let's scroll down, and it's got all the cities that I wanna visit here. It has the daily itinerary.
04:38The images are kinda small, but I like how they're kind of on point to the places that I actually wanna go to. And, yeah, it looks pretty good too. I love how I included, like, multiple images for each place that I wanna go to.
04:50And let's just scroll all the way down here and all the way down, and you can see that it's included some Japanese that I don't understand and some summaries and so on.
05:02So overall, I think GPT 5.6, the travel catalog that it made is definitely much better than GPT 5.5, And I think it's pretty comparable to Fable's travel catalog too.
05:16Fable, you can see here some of the text got cut off. Right? And 5.5 is very clear.
05:21I I feel like this is almost a tie down, which is actually very impressive because 5.5 was a lot worse than any kind of cloud model when it comes to design and 5.6 has really closed that gap.
05:34So now I think GPT 5.6 is a totally viable option for design and creating beautiful websites. Okay. Now let me show you something super fun and impressive.
05:45I don't know about you, but I grew up with Nintendo and Super Nintendo games, and one of my favorite games is Star Fox, which is a three d space shooter. So I asked both GPT 5.6 and Fable to build a single level for a Star Fox style three d retro game that I can play in the browser.
06:03The player should be able to fly forward through a planet surface with computer controlled windband. I should be able to dodge obstacles, shoot enemies, collect power ups, and even fight a boss at the end and make it look and feel amazing.
06:16Right? And 5.6 basically worked through it. And, yeah, I I think overall, it probably took maybe a little bit over ten minutes to build this game.
06:26So let's go ahead and actually play this game now and see what it looks like. Okay. So this is the game that GPD five point six made.
06:32Let's launch it. And here we have our plane with our wingman, and it's flying.
06:39As you can see here, there's some controls. I can fire. I can boost to move faster.
06:43I can collect power ups, and let's just go ahead and play all way to the end. I made it a pretty short game.
06:49So there's enemy ships coming as you can see, and we keep flying. There's different types of ships, and let's see what the boss looks like, shall we?
06:58Alright. Here we go. The dreadnought, and it looks pretty strong.
07:02I have a bomb button. I can hit x to hit the bomb. Okay.
07:07There we go. It's shooting at me, and let's just kill the boss and see what happens.
07:12There's different phases. Alright. Here we go.
07:19Boom. We hit another bomb, and the boss blew up. And there you go.
07:23The sky is yours, and we finish the game. Right? So, again, a game that you can build in just ten minutes that looks like this is pretty impressive.
07:31It could not have been done with the previous version of the models. Now let's take a look at the game that Fable generated. It's called Starwing.
07:40Press enter to launch. Okay. Here we go.
07:44So it looks quite different from the GPT game. We can also go faster. One thing that I noticed right off the bat is it's given the women some more personality and it's also added rings for me to fly through, which is pretty cool.
07:57Yep. There's a lot more chatter as you can see. There are buildings that look kinda weird.
08:02Hopefully, I won't die here. And let's get the power ups. Alright.
08:07And just buildings coming to view. Some of the buildings look weird, they're floating in the sky, but I like how I can power up my laser to be much stronger.
08:16And as you can see, it even supports barrels, which is a key feature of Star Fox. Alright.
08:23Let's go to the end here and let's see what the boss looks like. Alright. Here we go.
08:28Here's the boss. Alright. The boss even taunts me, and here it is.
08:34And let's keep doing barrel rolls, and let's see what happens. Okay.
08:37So for the boss, you have to beat up the cannons first so it has different hit points and hotspots. And then let's keep shooting at a boss, And, hopefully, we can kill it pretty fast.
08:48Oh, it even has a laser that we have to dodge. I don't know I don't actually know how dodge it. Alright.
08:53There you go. The boss blew up, and that's about it.
08:57So there you go. And and Fable, think, took a little bit longer, maybe a little over ten or fifteen minutes to build a game, but it also one shot at the game here.
09:06Right? So overall, these two games are both pretty impressive, but I'm gonna have to give a slight edge to Fable on this one because it actually remembered to build support for a barrel roll, which is, again, a key feature of Star Fox.
09:22And I didn't specify this in the requirements, but Fable remembered it and GPT forgot. Right?
09:27And I think overall, the game is a little bit more exciting, although that's totally up to taste. But the point is you can now build totally awesome games in three d or two d with these models in just a little over ten minutes with a single prompt. Okay.
09:42So now let me show you something that GPT 5.6 is definitely best in class at, which is browser and computer use to edit and publish short videos.
09:52So browser use basically means letting the AI navigate your web browser and computer use means letting it work through your local files. Now this may sound scary, but these features are literally magic once you get used to it. In fact, browser use is the primary reason why I switched away from ClawCode to using ChatGPTN Codecs as my daily driver.
10:15The reality is that most websites and apps do not have great APIs or MCPs, so good browser use is so important to save time on common workflows.
10:26Let me show you exactly what I mean. So I have this video post skill that I built that makes clips from my long form podcast videos and then post those clips to several different social platforms at once.
10:40And here's how it works. So first, I type slash video post and give it a link to one of my YouTube videos. Then it's gonna suggest five parts of the transcript that it thinks is interesting to clip.
10:52So you can see here, there's five different snippets, right, from the transcript. This interview that I had with Kun, who is an AI engineer. Now what I need to do is tell it which transcript I like, so clip four.
11:06And then it's gonna go off and use FFmpeg, which is open source library to basically make the actual video.
11:14So you can see here, if you look at the landscape view, and let's open this, it's basically made the video here.
11:23It's even added captions. Right? Alright?
11:26And you can see here that it can also make the vertical version as well at the same time. Now the vertical version looks a little bit off because the screen is cut off, but I can give you further instructions to make that better.
11:39As you can imagine, this is already magic and saves me a lot of time because otherwise I have to go into manual video editing tools and clip stuff myself. But now let's take a step further. Let's ask g p d five point six to schedule this clip on Instagram, TikTok, and YouTube shorts for 5PM on Thursday.
11:59Okay? So what it's gonna do now is it's gonna use browser use because Instagram, TikTok, and YouTube shorts don't have the best APIs to customize the same the way I want it to.
12:12And while we're waiting for this, let's ask Claude Fable five to do the same thing. So I'm gonna go to my video post here, and it's also made some video clips for me and I'm gonna say schedule this. Let's schedule it on Friday instead.
12:25Alright. So now let's take a look at how GPT and CloudFable works.
12:30So let's go to GPT first. Okay. So you see here that it navigated to YouTube Studio because it's gonna do the shorts first.
12:38It's clicking a bunch of buttons, and let's just go over here to Chrome and see what it's doing. So you can see here it's navigating to the clip that we want and it's gonna upload the clip.
12:51And eventually, it's also going to add a title and description that we want. Let's take a look at it right now.
12:59Alright. So it pasted in the title and the description here. Alright.
13:03Now let's take a look at what Fable is doing. So you see here that Fable has also read through a bunch of documents and is using the Cloud extension in Chrome. So the Cloud extension, it just works a little bit slower than GPT in Codex.
13:18You can see here that it has to take screenshots of everything to figure out where to click. Let's check back in on GPT. It looks like it's uploaded everything, and let's take a look at Chrome.
13:29It is clicking through the ad suitability stuff. None of the above submit rating. Oh, by the way, I didn't give it any instructions for this stuff.
13:36It just kind of figured it out on the right things to do. And at some point, it's gonna finish uploading the file.
13:43Okay. So I'm gonna stop it now on both GPT and Fable because it's gonna take a while. And it's gonna basically go through over 20 steps to do this to not only upload and schedule the stuff on YouTube shorts, but also TikTok and Instagram, which are even more difficult to deal with.
14:00For Instagram, for example, we have to use some sort of Facebook tool to make this work. And as complicated as this is, I've been using GPT 5.6 for this flow ever since it launched, and it hasn't made a single mistake yet. And and Claude, it can also do this, but the problem is Claude Fable is expensive to use.
14:21There isn't that much limits to use. And doing all this browser stuff just takes a lot of tokens. Okay?
14:29So I'm not sure if this is actually a gap between GPT 5.6 and Clock Fable or more of a gap between ChatGPT and Codex app versus ClockCode. But from my testing browser and computer use is definitely much better in ChatGPT and Codex than ClockCode.
14:47So I cannot emphasize enough how important this is because there are so many different ways that this stuff can save you time. Just think about all the terrible forms and PDFs that you have to fill out online or the hours of compliance training that you have to take.
15:02Browser and computer use can help with all of this. So this is a clear area where OpenAI and GPT 5.6 is ahead.
15:10Now let's move on to our next use case, which is adding a new feature to a mobile app. Many of you have asked me to make more videos about how to build mobile apps, and that is coming soon. But for this video, let's just focus on adding a new feature to an existing app.
15:27So I have this mobile fitness app that I've vibe coded to help me track my workouts. And, basically, I submit the same prompt to both GPT 5.6 and Fable, which is I would like to add a tab to track daily nutrition and meals in my fitness app.
15:43Please build a detailed plan in interactive HTML first with requirements, designs, tech stack, and more. Take a look at MyFitness app repo if you need context. Right?
15:54So a tip here when planning is to always ask the model to create HTML plans instead of the standard markdown files just because HTML is way more interesting and engaging to read. Another tip is to use compound engineering, which is this free skills repo that you can find on GitHub that I'll link in the description below to build more robust plans.
16:15Okay. So now let's take a look at GPT five point six's plan. This is the plan that I made and let's see what sections it has.
16:23So it has objective, which is to ship a force tab for nutrition. It has the problem framing, who the users are, what the requirements are, what the acceptance criteria is.
16:36And this is really important section, especially if you start using goal or long running agents because the agents have to verify what they're trying to do. And it has a bunch of stuff about success criteria. Some of this stuff is a little bit too much in my opinion, but I love how I included the design direction so I can actually see what the nutrition tab will look like.
16:55Right? And it looks pretty standard. And scrolling down, there's key technical decisions, there's high level tech design, data model, tech stack, and more.
17:04Okay? And let's take a look at Fable's plan. So Fable's plan looks very similar because I'm using the compound engineering scale for both.
17:13And just so you guys know what I'm talking about, just search for compound engineering and you'll see it here.
17:19And all you gotta do is just, like, point your AI at this GitHub link to download the scale. So there's a compound engineering planning scale that will help you make these robust plans.
17:30Anyway, let's keep scrolling down on the Fable thing, and you can see things are very similar. I want you to look at the design. And, yep, the design is also very similar to GPT.
17:39And I think this is because of the component engineering plan, but more importantly, I've given this project a design. Md and a set of components that you should always use in its designs to have this kind of consistent red theming. So after asking it to come with a plan, I then basically asked it to just go build a prototype.
17:58And what I asked for it to do is to just use local data because I don't want it to actually build the whole back end and the database for nutrition stuff. I just wanted to use local data so I can see what the new tab looks like. And this is what GBT came up with.
18:13So this is the new nutrition tab. So I can track my nutrition every day.
18:17I can add foods. I can add, let's say, chicken and rice, and it'll update everything here.
18:23And, yeah, I can add to dinner as well. I can add eggs and toast for dinner. And there you go.
18:28So we have our tracker here, and here are some other tabs in my app for workouts and more. Now let's take a look at Fable's output.
18:36Honestly, it looks very, very similar to GPT's output. I mean, let's go back and forth. I think these are slightly different colors, and GPT actually has some sort of words of encouragement, but Fable's output looks very similar.
18:48It has different side of things to add. Alright. Add to breakfast.
18:54And So both apps, look very similar because I have a shared design. Md and components library that I'm using for this app. So if I ask GPT 5.5 to generate this, it would look a lot worse because I've tried this before.
19:09Right? So again, as you saw from the travel example and this example, g p d 5.6 is much more on par with Fable and Anthropix models when it comes to front end design and experience.
19:20I still think it's maybe slightly worse, but I think it's a lot closer now. So from my point of view, mobile app planning and building, it's a tie between these two models.
19:30Now let's go back and quickly cover our two remaining use cases. So getting life and business advice and upgrading my personal OS. So as I shared in my previous Fable video, I have a plan document where I keep my long term goals, business strategy and what gives me and drains my energy and I have this advisor skill that I built that I use to basically talk to my plan all the time and get advice.
19:55So let's go back to GPT and let's open our plan review. And basically, what I asked GPT five point six to do is to read my plan doc, give me direct advice on what I should focus on, be candid but encouraging like a trusted friend, and try to identify insights that I may be not aware of based on our past conversations.
20:12So here's the advice that it gave me. It says that revenue is doing well. Here's what I should focus on.
20:18I'm working on a Codex course because I love Codex so much, and I'm working on some other stuff here. And I love how included this blind spots I think you may have missed section. So for example, the first blind spot is your ideal customer profile is still too broad.
20:34Right? It's saying that I'm building for PMs, engineers, creators, and operators, and they have very different reasons for buying.
20:41So try to narrow down the profile a little bit more, which I think is totally true, something I didn't really think about. And I can see that this output here is a lot more personal and engaging than previous versions of GPT models. But let's take a look at Fable now.
20:56So, again, I asked Fable the exact same question, and this is what Fable came out with. Its output is not really as organized as GPT 5.6 output, but if you actually carefully read through some of this, it does have some pretty unique insights.
21:11So for example, insight number three, you're close to a 100,000 YouTube subscribers, but YouTube only drives x percent of your sub Slack traffic.
21:19I assume that means it's a very small percent. Right? There's something really broken about this pipeline here.
21:25So maybe there's a better way to actually bring awareness to your newsletter and your substack on the YouTube. So this is advice that I don't think GPT 5.6 recognized.
21:36I think overall here, Fable has a slight edge on just like understanding my intent, but I think GPT's responses are actually more clear and better organized than Fable.
21:48And just one thing I found is that as both OpenAI and Anthropic have optimized for coding, the writing of these models has actually kind of regressed in my opinion.
22:00So for example, I think Opus 4.6 was a better writer and editor than Opus 4.8. But these results are pretty encouraging.
22:09I think GPT 5.6 is now a much better writer, has more personality and Fable as well.
22:16Okay. Last but not least, let's take a look at my personal OS project. So this is where I keep all my skills and workflows like the video post skill that I shared before.
22:26So over time, this repo can easily get messy because I keep adding new skills and automations to it. So periodically, I like to ask something like this.
22:35Take a detailed look at my repo based on your understanding of my workflows and our conversations. How can we improve it? What skills can we build or consolidate?
22:44Please list your suggestions in a number list. And here's what GPT five point six came up with. It's suggesting that I built an ops review skill.
22:53That's like a Uber skill across all my other skills to keep me on track. It's saying that I have four different sponsor skills. I should consolidate all of them, and it has a bunch of other suggestions here.
23:04Now I like to do this because then I can actually read through each of these suggestions and say, hey. Go do one, two, and three, but skip four. So I can actually give some feedback before it just goes off and does all of this.
23:14But my one piece of advice is if you have a very active report like like this, schedule a recurring job so that every week or so, it it will run this scale and automation to clean things up for you and get your approval. Right?
23:28But now let's take a look at what Fable came up with. So it's saying that there's a bunch of document drift across different skills and I should clean stuff up there.
23:39It's also suggesting that I should consolidate a bunch of thinking skills and merging a bunch of other skills around research and so on and so forth.
23:48And it has the same suggestion to try to consolidate my sponsor skills into one. So it's encouraging that both g p t five point six and Fable are suggesting the same thing. And maybe there's actually a way to get 5.6 and Fable to basically bounce off each other.
24:02Pretty sure in Codex and ChatGPT, you can actually call the Cloud CI to trigger Fable. Like, just go back and forth a little bit, figure things out before making the suggestion list to me.
24:14So, again, if if you have an active project repo, I highly recommend that you also do this on a regular basis and maybe even set up an automation for it so that you can keep all your skills and all your files clean. So in this case, I can't really tell which model is better.
24:29I I think, ideally, I would have both models and have them, like, make suggestions off each other. Alright. Now before we wrap up, I wanna share some perspectives from other builders and creators online about GPT 5.6 and Cloud Fables pros and cons.
24:44So GPT 5.6, basically, its biggest advantage is . I'm not sure if you heard of this phrase.
24:52It has a dog in it means that it will grab your problem by the throat and it will not let go. Right? It will not give up.
24:58It will keep working until it's done. Okay? And that is a very great quality.
25:01It's extremely hard working. I think that's what it means. Number two is is very reliable especially for browser and computer use.
25:08As I demonstrated with my video post scale, it would just like check all the boxes, go through 21 steps to make things happen for you manually using a browser. So I can basically trust it to get the work done. And finally, it's also fast and token efficient, and the negative is it's slightly worse on understanding your intent and also on front end, although I think it's closed the gap on the latter a lot versus the previous version of the GPT model.
25:34Alright. So now, CloudFable is magical model. It's incredibly intelligent and wise.
25:39It's like a , someone said. It's very thoughtful and very well spoken. It is still really good at front end.
25:45It will craft a better UI from scratch. And I think it also pushes back more, although I think GPT does this too. But it has it's more principles.
25:53Like, something that I really hate is when models just really agree with me or change their mind all the time. And I think with both Fable and GPT, I've seen that it sticks to its principles more.
26:04And the biggest con with Fable, of course, is that it's expensive and limited. It's about twice as expensive as GPT 5.6 on API pricing.
26:13In contrast with GPT 5.6, I can just keep using it without thinking too much. By the way, when you use both these models, I recommend that you use them mostly on medium or maybe high.
26:24I recommend that you do not go to very high or ultra core or what whatever unless it's really stuck on a problem. Default to medium or high to save on tokens. Alright.
26:34So let's recap all six use cases. Building an interactive travel website, I think it's pretty much tied between GPT 5.6 and Fable.
26:42Creating a three d retro space shooter, I give the slight edge to Fable. Editing and publishing short videos or any kind of browser use, GPT definitely wins. Adding a new feature to a mobile app, getting life and business advice, and upgrading my personal AI OS, I can't really tell a clear winner across these three bottom use cases.
27:00So I think for now, they're pretty much all tied. So which is my daily driver? GPT 5.6 is my daily driver.
27:10And that's because GPT 5.6, as I mentioned earlier, just has a much more generous rate limits than Fable.
27:18Basically, it's as good as Fable on everything except for front end design and maybe long term planning, but it's just much cheaper and faster to use. So if you can only get one subscription, I will get ChatGPT Pro or Max, not only for GPT 5.6, but also for the new ChatGPT app, which has best in class browser use and basically the codecs backend supporting ChatGPT.
27:40I'll make another video on ChatGPT work and my favorite automations to save time at a later date. But if you can afford both subscription plans, I would think of Fable as my planner, my architect and my designer and GPT as my engineer and my daily driver.
27:59Alright. So that's my recap of GPT 5.6 versus Cloud Fable. We're super lucky to have these two next generation models exist and competing at the same time.
28:09And I'm really excited to go off and build more things and save more time with both of these models. Now if you enjoyed this video, please like and subscribe for more practical AI tutorials like this. And also check out behindthecraft.com to get my best AI skills and courses.
28:27I'll see you soon.
The Hook
The bait, then the rug-pull.
Two frontier models, six real build tasks, and one blunt question: if you can only afford one AI subscription, which one actually earns the daily-driver seat? Peter Yang runs GPT-5.6 and Claude Fable 5 through a travel site, a 3D game, a social-video pipeline, a mobile feature, a life-advice session, and an audit of his own AI tooling to find out.
Named ideas worth stealing.
16:33concept
Compound Engineering planning skill
A free GitHub skills repo that generates more robust HTML planning documents (requirements, tech design, acceptance criteria) before an agent starts building, used identically with both GPT-5.6 and Fable in this video.
Steal forany AI-assisted feature-planning workflow
How they asked for the click.
VERBAL ASK
28:26subscribe
“please like and subscribe for more practical AI tutorials like this. And also check out behindthecraft.com to get my best AI skills and courses.”
Direct, low-friction close-out CTA at the very end after the verdict, paired with a plug for his own paid AI-skills site.
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.
