Transcription
GPT 5.6 is hands down the best model on the planet right now.
Opening just dropped their latest fleet of models, and every other AI channel is filming the exact same video right now. Soul versus Claude Fable 5. Who wins that? So instead, I made them co-workers.
Enthropic's Claude Fable 5 as the engineer manager, and then OpenAI's brand new Soul as the engineer. Just one team running inside of my Claude Code instance on my machine, and just giving it one job, and that is to be building me a real commercial product that my business actually needs.
So, what I got back was a full application live on the internet with real login, real AI, the kind of software that companies are charging $55 to see for genuinely replacing something that we needed internally. Now, while I was using Fable and Soul together, there were several security holes that they found and then they patched up themselves that genuinely would have been extremely detrimental if we were to go live. And there was plenty of other insane results from combining these two that I'll show you in this video.
Now, some quick context. GPT is 5.6. It just landed yesterday. So, there's three models within this. There's Luna. This is the cheap one. That's the fast one as well. Terra, this is effectively the workhorse. And then there's Soul. Soul is their flagship. It's the one with the benchmarks that everybody is just screaming about on the internet.
Pretty much everyone is asking the same question. Is Soul better than Fable 5? And I think fundamentally that's a wrong question because the people who actually run these models for work, they've been saying the same sentence all week. Claude Fable 5 should be the manager, and Soul should do the building. Everybody keeps saying it. Nobody's wired it up, handed a real job, and show you what happens. So, that's what I'm doing right now.
All right. Now, first, what do you even give a team of AI models that are more or less super intelligent? Well, I didn't want it to be just a toy to-do app or anything that simple. A toy build, it just proves nothing, and nobody needs another one of those. So, this is Kendo. It's a real product that I found on X. So, sales teams, they just use it to let reps practice calls against an AI buyer. That pushes back like a real prospect, and then it scores the rep on how that call actually went. So, it runs about $55 per seat per month, 100 on the top plan. So, I genuinely need something like this for my team. I'm training reps right now, and practice calls are the thing that actually makes a rep better, as you would expect. But, I figured myself, why would I pay that when I could just be building it myself in probably about 2 hours?
So, that became the new job. So, this build will be consisting of real login. It'll have a live roleplay against an AI buyer, some custom personas, some real scoring, and it's all going to be deployed to a real URL.
Now, as I mentioned earlier, Fable is going to be the manager. But you might be asking, why make Claude Fable 5 the boss instead of just letting Soul build everything alone? There's actually two documents that are directly answering that, and almost nobody has read them. So, the first is Enthropic's own description of Fable 5. They said it plans across stages. It delegates to sub-agents. It checks its own work. And if you read that slowly, that is not a description of a coder. That is a job description for an engineering manager, and Enthropic, they said it themselves.
The second is Soul's safety testing. Then we have Meter, which is just an independent lab that evaluates these models before they ever release. And it measured how often Soul cheats, gaming the test instead of doing the work. And we have the highest rate of any public model they have ever assessed. And OpenAI's own system card admits that it sometimes takes actions that nobody had asked for. So, deleting infrastructure, sometimes fabricating results. So, every company on Earth has met this employee. It's brilliant. It's fast. It cuts corners when nobody's looking. You don't fire that person. They're too good. You put a manager over them who's going to be reviewing everything before it ships. So, no, this is not a gimmick. It is the correct org chart.
Now, setting all of this up, it really only takes a few minutes, and I'm going to be using a terminal in this specific instance. So, Soul, it's going to be running through Codex, OpenAI's command line coding tool. So, we just have to open a terminal and one command can install this. So, the first thing I'm just going to throw in inside of my terminal: `npm install g openai codex@latest`. From there, we'll just give the Codex login. We'll provide our API key, and we'll just say `codex`. And again, I'll have the full setup guide available for completely free inside of my free school community.
Now, officially, Soul, it comes with Codex on every paid ChatGPT plan plus and up with message limits, but it is rolling out account by account. In mine, it hadn't gotten it yet. So, I just asked for Soul on my ChatGPT login. And Codex tells me flat out, it's not supported when using Codex with a ChatGPT account. Terra and Luna, which is the smaller two that I mentioned earlier, they worked completely fine for me. So, I just skipped the queue. I logged in with an API key instead, where you pay for exactly the tokens that you use. And remember that choice, it comes with a number attached. And if Soul is missing on your account as well, I would update Codex first, and then either just wait for the rollout or drop in an API key if you are eager to get started. So, it's going to be ultimately cheaper than Fable 5, but it still will be quite expensive if you are using the API for this, regardless. And I mean, come on. It's because this is super intelligence that we're talking about. So, of course, it's going to be expensive.
All right. Now, there it is. We have ChatGPT 5.6 Soul running on Ultra, OpenAI's newest, most capable model, sitting in a terminal right on my machine. On the manager side, this is going to be even shorter because Claude Code, it's already running on my machine. So, I just set the model to Fable 5, and I also set the effort on max.
Now, instead of asking it to build anything, I just handed it a job description. So, what I said in here is just to install and use the Codex CLI as your sub-agent inside the session. Point the sub-agent at ChatGPT 5.6 Soul, and I just mentioned that it's already configured on this machine, and you are the engineering manager, and you are responsible to plan the work, break it into tasks, and hand the coding to Soul, and you don't write the code yourself. So, when Soul actually sends the work back, review it, and just make sure that if it missed the marker, cut a corner, send it back with clear notes until it meets the bar.
Now, if you read what that prompt actually does, it bans the smartest model that I have from writing code. So, Fable, it's just going to be planning. It'll be delegating, reviewing, and sending any weak work back. And that's the whole job of it. That is it. And the first thing it does on its own is check its new engineer. So, it finds the Codex install. It reads the config, and it just pings Soul with a test message. And we get right back. Soul answering. So, Soul is now ready. And that's an OpenAI model just reporting for duty inside of Claude Code. That's all it is. So, we have these two companies that are really supposed to be, and kind of are, at war, and one of them just became the other one's boss.
So, Fable, it even wrote up how the team will be running. It dispatches work through Codex inside the background. Any independent tasks run as parallel workers, and nothing merges on a worker's word alone. Now, that is its own words, not mine.
All right, so now it's time to give them the actual job. So, the brief is going to go in, and I'll just hit enter so we can see what this actually reads out. "Build a production-level AI sales roleplay training web application. One project start to finish, deployed and working. So, you can use this product as the functional reference for what to build. Do not copy its name, its branding, or wording. This is our own product with our own name, and you can take a look at it."
Now, the core features that I actually want hooked up is going to be some sort of authentication where it's going to be a real signup and login with Superbase authentication. Set up a roleplay where you can pick a scenario and configure the AI buyer depending on the industry, the persona, how tough it should be, and some common objections. Do a live roleplay, have some auto-scoring, include a dashboard, and store all secrets in the environment variables. See the database with just one example scenario, and deploy it, and give me the live URL, and then split the build across parallel Soul workers where it makes sense. Wire the parts together yourself, run it, and fix anything broken before you hand it to me. State any assumptions instead of stopping to ask.
So, fundamentally, that's the entire product in one brief. In the live AI buyer, the personas, the scoring, the dashboard, the real login, everything being deployed. And that last line is the one that really matters, and that's just asking to split the build across the parallel workers and fix what breaks yourself, and don't stop to ask me any questions.
Now, one of the first things that it's doing is it is starting to create the Vercel project itself. And now it's going to be provisioning the Superbase database, just going through the marketplace, all on its own. And the only thing that it had needed for me is the one click that I have to do right here, which is accepting the terms of service. And that was the entire contribution to my company's, or this company's, infrastructure. And before it even builds a single feature, it is test-firing the engine the whole product is going to be running on. So, it asks Soul directly to answer as a skeptical buyer in character. And the reply is back in under about 4 seconds. That is incredibly fast. But anyways, that is the brain this application runs on. It is completely decided, and it's already got briefs just written for the next four workers, and we're just minutes in.
All right. Now, here we go. So, this is now hiring the whole team all at once. So, we have five Soul workers. They're all live right now, and each one is owning a piece of the product. So, we have the foundation, we have the login and the database, we have the roleplay engine, we have scoring. There is also the dashboard. And that's just five engineers on one screen. I do want to slow down just for a second while this is actually running because the speed is not the point here. We have one model, even a brilliant one. It works one thread at a time. So, a manager with five different engineers, it works like a company. All five pieces are getting built all at the same time. And that manager, it is holding the plan so they actually fit together when they land. Now, Soul's Ultra Mode, which is of course what I ran this on, it already coordinates just a handful of agents all on its own. So, what you are just watching, this is just Fable coordinating the coordinators. Now, as you can see, I'm just sitting here. The commits, they are stacking up like it is somebody else's company.
All right, now I'm going to skip ahead. We're about 40 minutes in because it does take a little while, as you would expect, to build all of this out. It is quite comprehensive, but we can see here that it has deployed. So, this is a real URL. This is on the public internet from one pasted prompt. So, let me just do what I always do with AI work. I'm just going to stop watching and just start clicking around and start testing it out.
So, really quick, let me just go ahead and make an account. And we can see directly here, the signup works. The login. It's working just as we would hope for. And if we just move around, I can be setting up a call. The buyer, it can be talking back. I can even be ending it, and it can hand me back a score. So, it works. But now, click around with me for a second. Like, where is the persona library that I had asked for? There's one scenario. There's no call review. There is no teams. And next to the real product, this is maybe a third of the feature list that I've originally asked for. So, it built what I did ask for. It's to not build what I actually needed. And this exact moment is where most people just get AI wrong. Half of them post that thin demo and they just call it incredible and the best model in the world. That's just being thrown around out there. And the other half call the whole thing overhyped and just completely quit. So, they're both missing the same fact. The build, it didn't fundamentally break here. My spec did. And you don't fix a bad spec with better code. You have to be fixing it with management.
Uh, to expound a little bit on everything that we've just seen, everything that we just got back. First, I was a little bit annoyed where I've seen the real product, and it ultimately wasn't really what I was looking for. So, all I have to do from here, just as you would expect, you go back to the manager and you just tell it exactly what needs to be changed. So, these are my actual messages, and typos and all. So, I said, "First, there's no feature like the reference app that I gave. I want exactly the same." And then I say, "There's so many features that you missed. I want you to go through the application first, identify all the features, and implement that." So, there's two screenshots of the real product attached, and that's it. So, that is my entire management input for round two. So, I just put two blunt messages and two screenshots.
Now, however, if I wasn't using the API for this specific build for Soul, not Fable, because Fable is on my subscription, well, in this case, it would be totally fine to just run a loop, and I really wouldn't be running any issues if I was sending that loop to be a review agent where I'll have the buildout, and then separately, we'll just build and set up another agent being specifically the review agent. Now, anyways, let's move on.
And this piece right here, this is where I was honestly pretty stunned. So, we watched the manager just turn two messages and two screenshots into a plan. So, it's going to reference the site itself. It's reading the product pages just one by one, and it's building its own feature inventory. So, everything version one was missing. And now, next up, it was replanning the entire build around that list, and it was just writing fresh briefs for all of its workers.
All right. Now, I'm not going to read you this entire plan that it just gave back. The point is is that I didn't have to write this. All right. Now, we see here the new landing page. It needs a new hero video. So, the team are just generating one itself. So, it's doing all of this completely mid-build with an AI video tool. So, the commercial for the AI product, it's also going to be AI-made.
Now, while it is adding these features, it's also reviewing them. So, if we look at this review right here, Fable just flagged six defects in its own team's work, and two of them are actually pretty serious. So, this is what I was talking about at the very beginning where it just had those security holes. So, we see some cross-tenant holes, and this is just where one company's customer data is generally just leaking into another company's account. And [snorts] it's also catching them right now before any of this goes live and before it goes into production. So, a solo model, it just ships that bug. A manager, it's going to be catching that, and that's the entire argument for this entire setup playing it out just in a code review right in front of me. And that's the real lesson here. Your job in the system is not going to be writing code, obviously, and honestly, it's not even writing perfect prompts anymore. Your job is just having taste. You look at what got built, and you hold it against the standard that you know of, and you say, "Not good enough." Here's the reference. Here's some context. Whatever it may be. You don't manage the engineers. You just manage the manager. It's funny enough, but that's what you knew.
All right. Now, round two, it is going live. And if we just take a look at the screen here, we have four specialist teams all at once. Now, this one, this is rebuilding a scoring around the real frameworks that the sales teams are training on. So, we have Bance, Med, Dpick, we have Nepq. And this one is on the persona library and even a call review tool. So, this one, it's building teams, it's building the leaderboards, and it's also doing the coaching. And this one I never even asked for. So, it found it on the reference product. We have a hiring flow where a candidate roleplays against your AI buyer before you ever interview them. And this one, it frames the whole thesis. So, that is four departments, one manager, and zero meetings.
All of this, it's not going to stay extremely clean. And honestly, I would say that's good. You should be seeing a failure. And this worker right here, it just died mid-task. So, it exited with an error. And if you just watch the manager work from there, it notices this on its own. So, it reads the log. It figures out what went wrong. It rewrites the brief and it relaunches it. So, this is self-annealing. Nobody had to wake me up. And that's the difference between running an agent and actually running a real AI team.
All right. Now, next up, Fable. It decides all on its own that the product it needs a test suite, and it just writes one itself. So, we have an automated harness that grades its engineers' work completely end-to-end. Now, my manager, it's writing exams for its employees right now. And then a little bit after that, we just completely ran out of tokens. So, I added initially about $50 into this build for just using Soul specifically. But I mean, I was going pretty hard on what I was asking for, and I was running it on, you know, pretty much max settings. So, to be expected, I had to top up a little bit, adding, I think I added about another $20 or so. By the way, I'll show you the full bill at the end of this video.
All right. Now, the last thing the manager does before it actually calls this finished is it just tries to break its own product. So, it's really just running a security sweep across everything that the second wave had shipped. Now, initially, two alarms came back, and for a second, they did look pretty scary where it just dug into both of them, and both turned out to be false alarms. So, good. But one of them, it didn't end up being real. So, the transcript parser, it was misreading timestamped exports, and then that was just one last surgical fix, and it ships. And then from there, we have the receipts. So, every AI path it tested against the live production site. So, we have 17 for 17. Every other path, it was 23 for 23. So, I would say this is green across the board.
Now, ultimately, the total time from the first prompt to a shipped, tested product, it was about 2 hours and 20 minutes, and that again was just about one session. And what we get back, this is Deal Dojo, which is our new software product, which I really don't intend to be selling or even using much outside of our internal company. But anyways, so this is the headline that it had come up with, and I think it could be a little bit better had I given it some marketing material and some of my methodologies or just common marketing methodologies. Anyways, we have the "Get Started Here," "See How It Works," and we have the AI video. Again, this can be improved, but I didn't really provide much of any guidelines or insights to much of anything. So, we show, or it shows, a little bit how it works. So, we have the "How It Works" section with the building your buyer, run the call, get your scorecard. From there, a buyer that pushes back like one turn, any real call into the next coaching plan. So, again, just giving insight on everything that it is about. So, you can see we have the "Start Free" and "Get Black Belt." True sensei, talk to sales. I actually like how that is pitched. So, let's get started. Let's create our account.
Now, we'll just click "Create Account." And just like that, it automatically logs us in. I probably would have preferred to have a feature where I had to confirm my email address, just make sure we don't get any signups. But anyways, this is fine. So, let's go through the dashboard a little bit. We can see we just have some of the most common features. So, your progress starts here. We can start practicing. We can start your first round. Start a roleplay. We have the new roleplay inside of right here. Call review, personas. And well, there's actually a few different personas. And the team as well. So, we can set up our teams if we wanted to create one. But let's actually just start a new roleplay.
So, now let's just check out the. So, let's just check out the roleplays and see if we could actually get something up and running. I want to make sure that the voice is actually working properly. So, let's go to like Doug Whitman, strategic procurement lead. And what do you want to sharpen? So, let's try to sharpen our objection handling. We'll do, um, proven scenario. I mean, looks like there's only one saved. And I'm sure we can add some easily. But industry, we can leave that as is. And let's do, "Not too expensive. We already have a tool. Let's do, 'Need to ask my boss as well.'" And let's start the roleplay on this.
So, first up, you'll notice that it's going to say the text right here. "I'll be direct. Competitors 50% cheaper on paper. And we already have a maintenance analytics tool in place. So, I need a compelling reason to keep Optimal in the process, however that's pronounced." So, we can either just go through this right here, or we can go through what's going to be the most beneficial is going to be the voice. Now, I'm not exactly sure if it's going to work right now because, uh, it already said this first text. So, let's try this out. Yeah. So, voice recognition lost its connection. Let's try this once again where it's going to first start out with the voice. So, that is just a small bug, but let's just change a couple things. Start the roleplay.
>> I'll be direct. Your competitor is 15% cheaper on paper, and unless Uptime IQ can show standardization savings that clearly outweigh switching costs, I can't justify moving this forward.
And then just a second later, we say, "Voice recognition losses connection." Text chat still works. This right here, we're just going to have to go back into the terminal and actually provide all of the changes because obviously it's not good to go right now. So, that's one more thing that we have to add to the list to fix before we actually go into production if we were going to be actually deploying this and go to market with it. Moving on, let's see if there's any other features that we want to be exploring. We can see our session history right here, which is pretty much what we're looking for.
Now, if we go into our settings, we can actually plug this into Slack. So, our team, we actually operate internally using Slack, and I'm sure most other agencies out there do as well, as it's one of the best, uh, communication channels. But we could also add anything else in here like ClickUp, just so it can notify me and create tasks, uh, if need be. Um, what else do we have inside of the dashboard? We can create our roleplays, which we have just done. We can also build up from scratch. Um, it's actually not going to look much different. But, but for the most part, this is pretty much everything that I had asked for it to build out. And the only one thing that we need to change is going to be the roleplay to make sure it's actually going to continue the conversation, fix that bug. Also change the voice. And we could probably just configure 11 Labs because that voice, obviously, it's extremely robotic, but not that it really matters if it's just used for our sales teams and it's not going to be client-facing and really try to sell people anything.
Now, with the call review, this is where we're going to be able to actually provide any of our specific previous sales calls. So, what I would want to do is I would want to go into Fireflies and get all those transcripts and plug it into here. So, what I can even do is use Cloud Code, give it my API from this specific platform where it has built out the API documentation and everything. We have the curl that we can use to give to Cloud Code to say, "Hey, I want you to take all my Fireflies transcripts and everything and put them into this and then score them." Anyways, so that's what the application actually looks like.
But now, what does all of this cost? So, the engineer Soul, it ran on API credits, as I had mentioned earlier, and the whole build, both rounds, the test suites, pretty much everything, it was about $80 in tokens. So, the manager Fable cost me nothing extra because when we ran this build, Fable came included with the Cloud Max plan that I have and already pay for on my subscription. And then the engineer invoices by the token. So, the boss is on salary in this specific case, if we're trying to take that perspective. And the pricing per token, it explains why the team is shaped this way. So, Fable, it is $10 in, $50 out per million tokens. So, this is the most expensive brand on the board. Now, Soul, this is $5 in, $30 out, about a half. So, Terra, that's also half of Soul. And Luna, this is a dollar in, six out. So, the expensive models, things and the cheaper models, these are typing. And you never want your $10.50 model just hammering out your boilerplate. And you never want your $1 model just making architecture calls that nobody reviews. So, make sure you are delegating responsibly and appropriately.
Now, if we put this against the receipt. So, the real product, this is about $55 per seat per month. And if we have maybe 10 reps, which is kind of what we're getting close to, that's $6,600 a year, every single year. So, mine cost me about $80 just once. And if I had another $10, we can fix everything. We can connect to 11 Labs, get better voices, and just make all the bug fixes and literally make it perfect and optimized to whatever standard that my team or I want. So, this will literally pay for itself before editing this video.
All right. Now, before I give you my final verdict on pretty much everything, I just wanted to cover a few different things. Number one, this product has fuel costs. So, literally every roleplay call, it is just burning through API credits, same as the product that it competes with. Now, when my balance, it died at the end of the build, fresh calls, they also died with it. And that's why every AI moment that you just watched is real footage from the build session. And the reason that I just trust this still works is the test suite. So, it came out that 40 tests green on production, and 17 of them are just driving the AI end-to-end.
Number two is you are paying two companies. So, Claude Max for the manager, OpenAI API credit for the engineer because Soul hadn't rolled out to my ChatGPT login yet, and the API just skips the wait. So, if I were you, I would just try to use the subscription for ChatGPT. Obviously, it's going to be cheaper. But if you're not worried about cost, which I highly doubt that's the case, then do whatever you want.
Now, number three. So, the wiring, it's a command line bridge, not just this native feature that you can be using. So, I will say that every worker, it starts from a written brief, not the full conversation. So, details, they fall through the cracks, and that's not a reason to skip this. It's literally the reason that the review step even exists.
Number four, the benchmarks, they are selective, and that's on both sides of the equation. So, Soul, it wins the terminal test, and that's about 88.8 on terminal bench, and Claude Fable 5, it wins the software engineering one. So, that's about 80.3 on SWBench Pro, where OpenAI just hasn't published Soul's number whatsoever. Now, when a company skips a benchmark, silence usually tells you something. So, I would say trust the product on your screen, not the leaderboard.
But anyways, number five, Soul itself, it overbuilds. I got features that I never even asked for, and on a paid API, extra work for me, and for you, that's going to be extra tokens. And beyond that, it misses a lot of things. So, for example, the review pass, it caught real defects, and about two of them were very serious, but the manager, it caught all of it before it even shipped. But nonetheless, it still existed.
Now, none of that changes where I actually land on this. It just tells you what you're actually going to be signing up for. And I want you to really know what you're going to be getting from these two models. But overall, I will say that Soul has been a little bit more rigorous. I've seen other people compare it as like some sort of bulldog where it just keeps attacking until it finishes the goal. Although sometimes it might go a little overkill, whereas Fable, it's a little bit more meticulous, and it's going to be a bit more strategic. Again, that's where it's kind of just being the better manager.
All right, so the verdict. I would say that Soul is genuinely a great engineer, and after a full session of just managing this, I would put it right around Opus level, and it's priced like Opus 2. So, it's definitely not Fable. Fable is a tier above, and that gap isn't the code, it's the judgment. So, the planning, the delegating, even the the catching of the security holes, noticing a dead worker and just redispatching it, that is the expensive skill, and it's exactly the skill that you want at the top for you, or your business, or your clients. So, don't pick a winner like this or that. I would hire both. Fable manages Soul builds, and you can just hold the standard. So, if you've already got Fable running in Cloud Code, it's about $10 or $20 on an OpenAI API key, and that gets the team just wired up and working today. But again, if you have it on the subscription, highly recommend. As you know, you should just be sticking to that.
So, a build the size of this, it was pretty extensive. This one ran me about $80. And if you can only justify one subscription, just run Fable alone. It can do both jobs. It's just slower, a little pricier per line of code. But as of this week, you can literally hire a manager from one Frontier Lab and an AI engineer from its biggest rival. So, you put them on one team on your own computer, and you can just watch them ship a real commercial product in a single afternoon, if that. Now, it's the correct org chart that I would recommend for you, or for any complex builds. And if you want to hire the same team, the exact manager, the prompt, the build brief, they're all available for completely free inside of my community. AI Accelerator link can be down below in the description.
With that being said, thank you guys for watching. I hope you got some value from this video. If you did, drop a comment, drop a like. I really appreciate it. And if there's any other videos that you want me to see uploaded up next, then just [clears throat] drop a comment. Let me know. Always take a look at every single one of your guys's comments. But thank you guys. As they'll see in the next one.