21:45The Next New Thing · Listicle
12 shocking things you can make with Fable
Andrew Warner and Brian Casel tour 12 community builds from Claude Fable 5 — then share the three prompting patterns that let it run deep without hand-holding.
June 10thThe argument in one line.
Across seven independent creator tests, GPT-5.6 Sol consistently wins on cost and reliable execution while Claude Fable 5 wins on planning, design taste, and creative ambition, pointing toward a manager/worker split rather than a single best model.
Read if. Skip if.
READ IF YOU ARE…
- You're actively choosing between Claude and GPT-based coding agents for daily work and want real cost/quality tradeoffs, not vendor marketing.
- You run meaningful API spend and need a framework for when the cheaper model is good enough versus when to pay up.
- You're curious what a two-model manager/worker workflow actually looks like in practice.
- You want concrete before/after examples: browser games, dashboards, video-editing skills, and inbox/LinkedIn automation.
SKIP IF…
- You're looking for a single 'best model' verdict — the whole point of this video is that there isn't one.
- You want hands-on setup instructions — this is reaction commentary on other people's builds, not a tutorial.
The full version, fast.
Two hosts react to seven creators' independent first tests of GPT-5.6 Sol against Claude Fable 5, covering browser games, dashboards, video editing, browser automation, skill cleanup, writing, and a Vision Pro drum kit. The pattern repeats across every test: Sol runs roughly a third to half the price and executes reliably, while Fable plans, designs, and writes with more taste and ambition at two times the cost or more. One creator's identical game-build prompt cost $14.22 on Fable versus $4.50 on Sol. GPT-5.6 also ships in three size tiers — Luna, Terra, Sol — each with multiple reasoning levels. The hosts land on treating Sol as the default worker and Fable as an occasional manager for high-stakes planning.
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 – 00:36
01 · Cold open
The hosts frame the episode: GPT-5.6's benchmarks claim it beats Fable while costing significantly less, and they'll watch seven creators' real tests to check that claim.
00:36 – 03:54
02 · Nate Herk's bike-game test: Sol vs Fable
Nate Herk gave Claude Fable 5 and GPT-5.6 Sol an identical prompt to build a playable open-world bike game in the browser. Both work; the hosts debate which is which before the reveal, then Nate's verdict: Fable's build felt more fun and GTA-like, despite Sol finishing in similar time for a third of the cost.
03:54 – 07:21
03 · The manager/worker framework, and LatchLoop
The hosts name their recurring mental model — Fable as manager/planner, Sol as worker/executor — and discuss when it makes sense to switch models mid-task. Bryan introduces his own platform, LatchLoop, an AI coding-agent and general-agent platform.
07:21 – 10:21
04 · How I AI: dashboards and creative packs
A creator compares Sol and Fable on a dense operations dashboard and a creative-pack website. Sol's output is functional end-to-end with clean, neutral visual hierarchy; Fable's is harder to read with layout gaps, though both score well overall.
10:21 – 13:39
05 · Video editing, and why iterative prompting wins
A creator drags a raw conference recording into GPT-5.6 Sol, asks for five social clips, then iterates with specific feedback (more vertical, tighter, faster) rather than accepting the first pass. The hosts frame this as the core lesson of the whole video: one-shot demos look great but rarely survive real use.
13:39 – 17:24
06 · Browser automation and the video-clipping skill
One creator has Sol operate Chrome directly to triage roughly 500 LinkedIn messages by a strict filter rule. Peter Yang's clipping skill turns a long podcast link into landscape and vertical social clips with captions automatically — though both hosts are skeptical the vertical cuts are actually good enough to perform.
17:24 – 21:27
07 · Cleaning up skills, and letting the models grade each other
A creator asks GPT-5.6 to audit his personal automations repo and suggest consolidations; Fable, asked separately, suggests nearly the same thing. The hosts discuss preferring a reviewable checklist UI over open-ended chat back-and-forth for approving AI suggestions.
21:27 – 23:06
08 · Zapier MCP break, then an AI-made video nobody can tell is fake
Sponsor segment on Zapier MCP's scoped, permissioned access to over 9,000 app actions. Then: a creator shows a fully narrated video generated from a single prompt to GPT-5.6 Sol — no camera, no recording, no editor.
23:06 – 26:36
09 · Pricing showdown, and a professional writer's verdict
Direct pricing comparison shows Sol roughly half of Fable's per-token cost. A writer from Every finds GPT-5.6 plainer and less prone to Fable's tendency to overexplain, preferring it for emails and marketing copy.
26:36 – 30:00
10 · Fable's design win, Dan Shipper's workflows, and Berman on price
Fable still beats Sol on image-prompt taste using the same underlying image model. Dan Shipper describes running much of his personal and work life through the Codex app; Matthew Berman recaps GPT-5.6's per-token pricing advantage.
30:00 – 32:13
11 · Model tiers, and the Vision Pro drum kit finale
Recap of Sol/Terra/Luna's reasoning-level tiers, then Bijan Bowen's closing demo: a photorealistic virtual drum kit for Vision Pro built from a single prompt, refined once from 2D to 3D.
Lines worth screenshotting.
- GPT-5.6 Sol costs roughly a third of Claude Fable 5 for comparable coding work — one creator's identical game-build prompt cost $14.22 on Fable versus $4.50 on Sol.
- Fable produced about 90,000 output tokens to Sol's 31,000 on the same task, and the extra tokens didn't translate into a better result — Sol's build won the head-to-head.
- The recurring verdict across seven independent creator tests: Fable plans and designs with more taste, Sol executes more reliably and far more cheaply.
- GPT-5.6 input tokens run $5 per million against Fable's $10, and output runs $30 versus $50 per million — roughly half the price at every stage.
- One creator spends around $400 a day in API credits on Sol and says doubling that to run Fable 'doesn't make sense' at his volume.
- GPT-5.6 ships in three sizes — Luna, Terra, Sol — each with multiple internal reasoning levels up to 'Ultra,' giving far more cost/quality dials than picking a single flagship model.
- A creator built a five-clip social-video pipeline from nothing but a YouTube link and plain-English feedback ('make them faster, tighter, more vertical') without touching editing software.
- One-shot demos look impressive but rarely hold up: 'the one shot looks great in a video, but it's just impractical once I use it' — iterative feedback consistently beat single detailed prompts.
- Fable's outputs were repeatedly described as defaulting to a dark-mode, monospace look, while Sol trended toward a lighter, cleaner UI — one reviewer worried that consistency could itself become the next 'AI slop.'
- One creator used Sol to triage roughly 500 LinkedIn messages autonomously via browser automation, replying only to executives at 'tier one companies' per a standing instruction.
- A separate creator generated an entire narrated video — voice, face, and delivery — from a single prompt to GPT-5.6 Sol, without ever standing in front of a camera.
- Both models converged on nearly identical suggestions when asked to audit and consolidate the same personal-automations repo, which one host reads as a sign the recommendation is genuinely sound rather than model-specific flattery.
- Fable still wins on image-prompting taste, producing a simpler, more considered result from the same underlying GPT image model and prompt.
- A professional writer's read: Fable tends to 'overexplain' and drift into its own literary register, while Sol stays plain, uses natural lowercase, and one-shots marketing emails other models couldn't.
- The panel's practical default: try the cheaper model first on nearly everything, and only escalate to the pricier one when there's a specific reason to expect it'll meaningfully outperform.
Reach for the cheaper model first, save the expensive one for real stakes
MODEL SELECTION
Across seven independent tests, the pattern holds: the smaller, cheaper model executes reliably and iterates fast, while the flagship earns its price only on planning, design taste, and genuinely hard problems.
02Nate Herk's bike-game test: Sol vs Fable
- When two agents get an identical prompt and equal creative freedom, output token count doesn't predict quality — one model used three times the tokens and still lost the comparison.
- A same-prompt, side-by-side test only tells you something if you genuinely don't know which model built which output before you judge it.
- Camera control and 'feel' were the deciding factor in a game demo, not raw feature completeness — both builds had working physics and coins, but only one felt good to move around in.
03The manager/worker framework, and LatchLoop
- A cheap, capable model with lower creative ambition can be a genuinely good 'worker' — the pricier model's edge shows up specifically in planning and creative direction, not raw execution.
- The cost math only matters relative to volume: someone spending $400 a day in credits treats a 2x price jump completely differently than someone running one project a month.
- Pick a model based on the role you need filled — manager or worker — rather than assuming one model has to do both jobs well.
04How I AI: dashboards and creative packs
- A prototype's visual style consistently signals which model built it — one model defaults to dark, monospace layouts while the other defaults to lighter, cleaner UI, independent of the actual prompt.
- Functional completeness, every button actually working and not just looking clickable, mattered more to this reviewer's evaluation than visual polish.
- A design style that looks refreshing today can become tomorrow's cliché once every output from a model looks the same — 'AI slop' is a moving target, not a fixed look.
05Video editing, and why iterative prompting wins
- The most durable lesson from every demo in this video: one-shot output looks impressive in a clip but rarely survives contact with a real use case — plan to iterate with specific feedback.
- Giving an agent concrete corrective feedback ('make it more vertical, faster, tighter') gets dramatically better results than a single detailed upfront prompt.
- AI editing tools are strongest at rough-cutting and clip selection from long footage — treat the output as a fast first draft, not a finished deliverable.
06Browser automation and the video-clipping skill
- An agent that can triage inbound messages by an explicit filter rule can safely automate high-volume, low-stakes correspondence.
- A tool that makes a technically correct short clip isn't the same as a tool that makes a clip people will actually watch — production competence and audience judgment are separate skills.
- Automated highlight-clipping from long-form video is good enough today to draft several candidate cuts fast, but a human still needs to pick the one worth publishing.
07Cleaning up skills, and letting the models grade each other
- Periodically asking an agent to audit and consolidate your own accumulated automations catches drift before it becomes unmanageable clutter.
- When you want granular control over many suggested changes, ask for a reviewable list with accept/reject/defer options per item rather than an open-ended chat, which doesn't scale past a handful of items.
- Two different models converging on the same recommendation independently is a stronger signal that the recommendation is actually correct, not just model-specific flattery.
08Zapier MCP break, then an AI-made video nobody can tell is fake
- Scoped, read-only or draft-only permissions on an agent's access to email or messaging let you get automation's speed without automation's risk of an irreversible mistake.
- A single well-written prompt can now generate a complete narrated video, voice, framing, and delivery, with nobody standing in front of a camera — that alone reshapes what 'content production' costs.
09Pricing showdown, and a professional writer's verdict
- Compare models on a true unit basis, per-million input/output tokens and cache-hit pricing, rather than trusting a single project's total bill, which conflates cost-per-token with how verbose the model chose to be.
- A model that writes more plainly, with natural sentence casing and fewer 'AI-isms,' can outperform a larger, pricier model specifically on writing tasks like marketing emails.
- Efficiency-focused models are often built for scale and enterprise-serving margins, which shows up as lower cost per token even when raw output quality is close.
10Fable's design win, Dan Shipper's workflows, and Berman on price
- Even a 'losing' model can still win on the dimension that matters for a specific task — here, the pricier model's image-prompting taste was still clearly better despite losing everywhere else.
- A sensible day-to-day default is to try the cheaper model first on everything, and only escalate to the expensive one when there's a specific reason to expect it'll do meaningfully better.
- Sunk-cost thinking around 'trying the expensive model first' is a bias worth naming — trying cheap first and discarding it costs almost nothing when the price gap is this large.
11Model tiers, and the Vision Pro drum kit finale
- A model family split into small/medium/large tiers, each with multiple internal reasoning levels, gives you far more cost/quality dials to turn than picking a single flagship model.
- Spatial computing apps generated from a single descriptive prompt are now good enough to genuinely surprise someone seeing the finished result in a headset, not just on a flat screen.
- Asking for a specific correction after seeing a rough first pass ('I want it 3D and photorealistic, not 2D') got a dramatically better result than accepting the first output.
Terms worth knowing.
- Sol / Terra / Luna
- The three size tiers of GPT-5.6 — Luna the smallest and cheapest, Terra the mid-tier, and Sol the largest — each with multiple internal reasoning levels up to an 'Ultra' setting.
- Fable 5
- The flagship Claude model used throughout the video as the pricier comparison point, generally favored for planning, design taste, and creative ambition.
- Big model smell
- A term used in the video for a flagship model's tendency to over-elaborate or over-engineer a response, as though it's showing off its own scale rather than answering efficiently.
- Manager/worker framework
- A mental model for combining two AI models: the pricier model plans, reviews, and directs (the manager), while the cheaper model executes repeatedly (the worker).
- Zapier MCP
- A remote Model Context Protocol server that connects an AI model to over 9,000 third-party app actions, with configurable read-only or draft-only permissions for safety.
- One-shot
- Producing a finished result from a single prompt with no follow-up correction, as opposed to iterating with specific feedback until the output is usable.
- Browser automation
- Letting an AI agent operate a live browser session directly — clicking, reading, and replying — rather than only generating text or code for a human to execute.
Things they pointed at.
06:09productLatchLoop
23:25toolOpenRouter pricing comparison
Lines you could clip.
03:09
“This Fable run would have cost me about $14.22, whereas the GPT Soul run would have cost $4.50.”
05:54
“I feel like Fable is a better manager, and that's how I think of it. I think of it as a cofounder... Soul is just a really, really good worker.”
17:00
“It's like watching an elephant dance. It's like, wow. It can do it, but not really useful.”
22:55
“He gave me one prompt. That's it.”
25:30
“GPT 5.6 is just to the point. It's simple. It doesn't have a ton of AI-isms.”
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
00:00So we finally have gbt 5.6, soul, terra, and LOONA. And the benchmarks are insanely impressive because it basically shows that soul blows opus and fable out of the water and is significantly
00:12cheaper. We're gonna look at what many creators who had time to look at this, uh, and play with it have done with it. I wanna see what they created with it.
00:19I wanna see how it compares to Fable five, and we're gonna look at seven different videos here. And I picked the exact spots for us to look at. Presented by Zapier, the AI automation company.
00:30Brian, here's the first one that I want you to see. I want you to see what he has built with it. Claude on the left here with Fable five.
00:34I've got Codex on the right here with GBT 5.6 Soul, and I gave them both the exact same prompt. So I gave them both the slash goal, and I told them to build a genuinely fun, playable, open world bite game that runs in the browser. I gave them these other rules.
00:47I'm not gonna read off this whole prompts, but if you want to take a look, feel free. I gave them creative freedom, and I just wanted to see what they could come up with. Okay.
00:54So I haven't played any of these yet, but I'm not gonna tell you which one's which. We're gonna see them both, and then I'll reveal. So WASD to steer, space to bunny hop, q and e for air tricks, shift for boost.
01:04Okay. Let's go ahead and drop in. Nice.
01:07Okay. So we can move around. We have a city here.
01:09The camera's a little bit hard to maneuver with. You can't drive through the building, so that's good to know. Okay.
01:15Just got a little trick in the air there. So this is honestly working pretty well. I don't love the actual way that this works.
01:22Like, the top down bird's eye view, I don't think I like. It's a little bit hard to control. But, hey, functionality wise, the coins are working.
01:29The controls work. I just jumped over a building, and I can bunny hop.
01:33Okay. Cool. And it doesn't look too bad either considering this didn't take it very long.
01:37So that was the first game. Alright. And now here is the other version.
01:41Let's go ahead and see. It's pretty similar controls. So let's hop in here.
01:45Okay. So this one is more like a three d, almost more of like a POV sort of style. I can jump.
01:51I can spin. Interesting interesting differences so far. Um, but I'm in a city.
01:55I've got the coin still. Let's see if I can go try to get on this ramp and see what the, uh, the physics of that look like.
02:01Wait. Am I on the wrong side of the ramp?
02:03What do you think of this, by the way? Which one do you think is better?
02:06I I I think probably the first one was GPT 5.6, and that this one is Fable. That's what I would guess at least so far.
02:17I think so first one had what seems like to me some of the kind of remnants of GPT's design sense.
02:27GPT 5.6 is a lot better, but it it still has has a little bit of that same kind of look that's coming from a GPT model, if if I'm correct.
02:40But Let's find out. One I would guess is fable because of the the attention to all this this three d detail and stuff.
02:49Yep. That was the wrong side of the ramp.
02:52But this definitely has more of, like, that I don't know. Like, more of a GTA vibe, like an open world game that you can actually, like, I don't know, explore. Feels like a bit more fun, I think, to play.
03:01It doesn't make me nauseous. Definitely a clear winner. So I think that we've decided that.
03:06Alright. So in this exam Yeah. That one's definitely there.
03:09Which is the one I thought was the winner here was fable five. So that's on the left hand side here. This took twenty one minutes and thirty seven seconds, whereas the other side, GBT Soul took twenty three minutes.
03:20So similar on the timing here, but let's look at cost. So this Fable run would have costed me about $14.22, whereas the GBT Soul run would have costed 4 and a half dollars.
03:30Now what I think is really interesting here is the output tokens. Fable outputted about 90,000, and Soul only outputted about 31,000.
03:38So we've consistently seen that Codex and GPT are way more token efficient. And sometimes that's better and sometimes that's not. So
03:45that kinda brings me to what he thinks, and this is what a lot of, uh, the people who've messed with it have told me.
03:52Here's the difference. I feel like Fable is a better manager, and that's how I think of it. .
03:56I think of it as a manager. Soul is just a really, really good worker. Really good worker.
04:02Like, I think that having Fable orchestrate a bunch of soul agents would be amazing. I gave them each their other things that I would say, like, they are better than the other end.
04:11So let me ask you this. So, yes, we can obviously see in the Fable example, there's a lot more creativity, um, and we'll see as we go on how it does with, uh, advising.
04:19Brian Mhmm. When when you use multiple models, how do you shift between one and between multiple models?
04:26So, uh, yeah, it's interesting. I I think I'm agreeing with what he says so far, um, in relation to this. Uh, Fable has that, like, the as they call it.
04:36Yeah. Um, and that is not something that's really present, uh, in my experience so far with Soul. Um, however, um, Soul does really feel like the the worker.
04:47And I think that, uh, I I would challenge him actually that if that prompt was tweaked, I think you could probably get a closer to the fable result with Soul as well.
05:01And so if you wanted to say, I wanna be hands off as much as possible with this thing and let the AI come up with things for me, then I think you'll get the better result of, like, Fable is the the planner, and then Soul is the implementer.
05:18But I think most of us, we want to we wanna still be in the loop in some way. Right? And, uh, so if you're able to work on that plan still with Soul, um, then I think you can, uh, you can get a similar result.
05:32And Because you'd be telling Soul more specifically what to do like a boss, and that's what it means. So, like, yes, Fable has better ideas than Soul, but, uh, I believe as a human, I still have better ideas than AI. And I think that that's where where we're supposed to still be involved.
05:48Um, yeah, Interesting to see so far.
05:52Uh, my my personal opinion has so far been leading towards, yeah, Soul is the the workhorse, and if it can do 90% of things, like, yeah, maybe Fable's a little bit better, but is it, like, more than two times the price better? And for me, that's that's not true. Okay.
06:08But then when you switch so let's just quickly introduce you. You are the creator of Latch Loop. What's Latch Loop?
06:14Latch Loop is a AI coding agent platform and general agent platform. So similar to the new chat GPT work that was just announced and, uh, Claude, uh, Cowork and Codex or Claude Cowork and Claude Code. So Just final question.
06:28How do you then shift between two? So let's say let's say that you buy into this and you say Fable's gonna be the manager. It's gonna create the the task list and check on it afterwards.
06:36Sol's gonna be the worker. What's your process? And we'll go to the next video in a moment.
06:39What's your process for going back and forth?
06:41So,
06:42typically, I am sticking with the GPT models. Oh, you are? But you're not you're not getting one to create a plan and the other to act on it and then come back to the first.
06:51I'll still use probably, I'll see what I think about GPT 5.6 here,
06:56but I'm anticipating that I'll still be using it for the plan as well, and then have have separate instance of it, separate context window, go and do the implementation. Yeah. Fable has been good in my experience, but it just costs too much.
07:10And and I'm I'm spending, like, 400 ish per day in in API credits, and to then say, well, I'd have to spend 800 or something for Fable,
07:19that does not, uh, make sense for me. Okay. Let's go on to the next.
07:23Uh, this is How I AI. She's talking about how Soul, Terra, and Luna came out. Let me show you one of the things that she created.
07:30These and why I really liked Soul compared to other models. Although what where Fable did perfectly She got more creative with her tasks. I like that she didn't do games.
07:38Dense operation dashboard. It's basically like an eval for a doc scheduler
07:43app in its full design. And what you can see here is both were pretty useful. Soul on the left and Fable on the right.
07:52I just think Soul was the most unique. All of the other ones really just looked like this dark mode, monospace kinda layout.
08:02As you can see here, Soul actually has, like, a really clean kind of, like, neutral color lay with great visual hierarchy, semantic color, and this thing was functional.
08:14So, like, everything I expected to be able to click and work and assign and do, all of it actually worked. And this was just my experience across a bunch of the different prototypes is the sole ones were just a lot more functional, and that made a big difference on how I'm evaluating things.
08:36Now let's look at the fable design. Again, it's pretty good. It's actually a lot harder to read, though, and the design, I would say, is not as unique.
08:45And even some layout issues like this white space here at the bottom. Now it did do a lot of a lot of design issues with it. Yeah.
08:52I agree. I I think the
08:55the this kinda dark look is kind of the AI slop look that that we're all used to and sick of. But I have I have seen people talking about that they have found that 5.6 will so commonly do this this new design that that we're seeing here in the light mode.
09:17That while I agree it looks way better, I also wonder, like, will we will we get sick of that? Will we say that that's now that's the the new AI slop? But for now, I agree.
09:28It's definitely
09:29it's better. You're saying it is better, but maybe it's so consistent that we're gonna get tired of it too. Fair point.
09:34Here. Take a look at the creative too much in that same kind of style. It was this creative pack
09:40website. Again, both of these got fives from me. I just really preferred that Soul went ahead and had, like, a personality.
09:50Look at these placeholder images versus what Sable came up with, which I will say part. It is nice there.
09:57And clean and worked really well. And, like, I have no complaints about it. It's a good one, especially for side of sort of a wireframe style prototype.
10:06It's great. I would just say, it's not this. This is pretty interesting.
10:12It's got
10:13We'll we'll pause here because we're gonna see this over and over again. I think we'll see it from Dan Shipper in a moment. Um, let's go on to another use case.
10:20This one comes up a lot, and it's a lot of this, oh, I didn't expect that type of moments. She's the first one that I saw do this video editing. Look.
10:27Video editing. I
10:29have to do a lot of social clipping, and it's really tedious to go through and clip videos.
10:37So taking something really long and shortening it. So recently, I spoke at Cursor's event and gave this talk on the future of PM and got the recording from the Cursor team. Thank you very much.
10:47And I really wanted to make it a hype video. So all you have to do is literally drag the file in here. And I said, can you cut this video into five clips for social?
10:58And I gave some feedback. I said, want them horizontal.
11:02I want them height video cuts from various parts. I need them to be faster. I need them to be tighter.
11:10I love that she's showing how she's giving guidance and not just one shotting it. The one shot looks great in a video, but it's just impractical once I use it.
11:18Yeah. No. Well, I agree.
11:20I think one of the important things to realize in using AI is that all we're all doing is just distributing attention. And so if you can get the AI to spend more attention on these other details, either in the initial prompt or through, like, increased runs and and extra turns on it,
11:39end up with a much better result. And then I got these, like, sharp and funny height videos.
11:45Let's see if it opens up. This one's for my my talk. Yeah.
11:47I'm curious to see figure out what it means to be a product manager in the age where anybody can build anything. We have been coming up with creative ways to avoid building things That's good.
12:00Forever. Yes. PRDs, like these complicated documents where you had to describe.
12:05So, like, that would have taken me so much time to, like, find the right cute parts, clip it, cut it. I was able to drop it into CapCut, put some music, ship it on social. It's like a really cute hype video.
12:16But this is one of my favorite use cases. I'm pretty sure it can do even more color grading sound, all this kind of Yeah. This
12:23has been really impressive. What do you think?
12:26Yeah. Yeah. I I think it was pretty good.
12:29Is she gonna compare it to to any any other model of, like, a difference with it? No. She's not.
12:35She was just doing it straight up. And I I haven't seen anyone except for Nate Herc, who I'll show in a moment, who did a long video, really do a straight comparison, and he didn't even do it side by side.
12:46He just said there's another video for What we're have you done any straight up video editing where you give it a file and you tell it to edit?
12:54Um, I have not tried it, like, through, like, regular coding agent interfaces, but I I tried some, like, AI video editing products, like, over a year ago. Mhmm. And they were useful for certain kinds of cleanup, but, like, it didn't stick with with with us and how we used it, at least for for my video editor.
13:15But, yeah, it makes me curious. I I wanna get I wanna get my video editor, uh,
13:19trying some of this out and see what he thinks about it. You know where it's good? It's in the let me give my Zoom video to to GPT five six and ask it to just pull out the part of the video where the person complained so that I can show it to the team, where my meeting went sideways so I can go and analyze why it went sideways.
13:38I think that was interesting. We're not talking about Codex, the app, but
13:42I like how she's using Chrome in the app. Go with the stars and and do some stuff. And, like, I'm sorry, LinkedIn.
13:49I know I'm not supposed to do this, but I opened up LinkedIn. And I said, can you use Chrome reply to messages that are a very high value chat PRD or the How I A Podcast?
13:59Keep the bar very high. Again, I love you. Cannot deal with all the LinkedIn requests.
14:04So, like, only accept them if they're executives of tier one companies. I don't want random sets of connections. It went through and burned through probably 500 messages.
14:14It replied to people that I needed to reply to.
14:18It said thank you to people who said nice things about the podcast. Thank you to those people. I do mean it, but it just rocked through through browser use.
14:27That is so impressive. I wanna try that. I wish she had a demo of it, but I want to try that use case.
14:34Yeah. Yeah. Mhmm.
14:37I haven't tried too too much with the browser use outside of, like, having, like, the agents, like, test things in software development. But, yeah.
14:48I think it's it's really exciting to see where this is all going for, like, knowledge work related tasks now, not just coding agents. I think the big, like, unlock recently has been people realized working with these coding agents that, really, the coding agent is the everything agent.
15:05And what I mean by that is the the knowledge worker agent is basically a coding agent, but instead of editing code, it is making changes to a doc or a presentation or doing things on on your web browser. So
15:19Let's go on next. I don't even know if I wanna show this. This is essentially the same thing.
15:25Post skill that I've built
15:27that makes clips from my long form podcast videos and then post those clips to several different social platforms at once. And here's how it works.
15:36So 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.
15:48So you can see here, there's five different steps. Right? In the transcript.
15:52This 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. And then it's gonna go off and use FFmpeg, which is open source library to basically make the actual video.
16:10So you can see here, if you look at the landscape view, and let's open this, it's basically made the video here.
16:18It's even added captions. Right? Alright?
16:21And 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 Yeah.
16:30Yeah. I'm a I'm a bit skeptical of some of these things. I I think there's the the cases that you described for, like, finding some certain area in the video or, like, making some kind of, like, small cut of something you can share with somebody else, but it's not like your full produced YouTube video Mhmm.
16:47Are really useful. But, yeah, I'm skeptical because, you know, like, there's there's companies like OpusClip and and all that that are putting so much effort into And then how can we we optimize to to create these good clips.
16:59But, um, yeah. I I still it's hard to say, like, if you were to use a skill that could do this, yes, it can produce something that at first glance looks impressive, but are these the kind of videos that will actually end up getting views on the platforms?
17:16I agree. I think that every time they do this, it's impressive. It's like .
17:20It's like, wow. It can do it, but not really useful.
17:23And, alright. Let's move on to what the other thing he did. OS project.
17:27So this is where I keep all my skills and workflows like the video post scale that I shared before. So over time, this repo can easily get messy because I keep adding new skills and automations to it.
17:40So, periodically, I like to ask something like this. Take a detailed look at my repo. Based on your understanding of my workflows and our conversations, how can we improve it?
17:49What skills can we build or consolidate? Please list your suggestions in a number list. And here's what g p t 5.6 came up with.
17:57It's suggesting that I built an ops review skill. That's like a Uber skill across all my other skills to keep me on track.
18:05Let me pause here for a second. I do this for individual individual skills. Skills.
18:09I I wanna wanna now do this for all of them. I don't like doing it the way that he did here in, uh, in chat because there's too much back and forth on these. I ask to get an HTML file or an HTML page with each one as its own number, and then I want radio buttons underneath it that says do it, explain more, reject it, save for later, etcetera, and an input box underneath where I can ask questions or give it some guidance.
18:33That way there's a little bit more involved with each one of these. I don't want just a yes or no or to respond in a chat. I don't think everything fits in a chat window.
18:42Yeah. I I agree. Um, I I think I would do it the same way that you are.
18:47Like, yeah, the the way that I'll typically work, uh, with an agent is that I'll have it focus on that one thing, and then for the for the other details, I'll have it go and spawn off other tasks. And then each of those things, I I will give some additional input or or review where needed. But I think the the exercise of somehow approaching this process of like how can what are the other things that I can do to improve and and get the agents working in better for me.
19:15Yeah. Going through this exercise in some way to improve your skills, improve the automations and loops and things that you have is valuable.
19:22It's saying that I have four different sponsor skills. I should consolidate all of them, and it has a bunch of other suggestions here.
19:29Now 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 feedbacks before it just goes off and But does all of my one piece of advice is if you have a very active people 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.
19:53Right? But 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.
20:06It'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.
20:14And it has the same suggestion to try to consolidate my sponsor skills into one. So it's encouraging that both g p t 5.6 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.
20:28Pretty sure in Codex and ChatGPT, you can actually call the Cloud CI to trigger Fable. Like, just go back and
20:37You can do that. Right? In fact, OpenAI has a skill that they or repo they put together to allow you to to have codecs check the work of of Opus.
20:50Yeah. Yeah. Yeah.
20:52There's the the agent client protocol can be used to kinda connect all these agents together. I use that myself sometimes, but, yeah, I I worry if sometimes that that this can there's a distinction between, uh, it turning into a mess and and
21:09and it being useful, I think. And a second opinion. I think this was an interesting no.
21:13Actually, I don't think it was that useful. We learned really that it wasn't any more insightful than Fable. And, frankly, I would just probably run both.
21:19For something like this, it doesn't take a long time. It's not that important. Let's move on.
21:23Oh, sponsor. Zapier, real quickly, I will tell you something, Brian, that you may not know that Dan Shipper, founder of Every, gave an email address to his codex within the within the codex app, which I guess now is called the OpenAI app again or the ChatGPT app.
21:41I I don't like their whole new existing. Chat chat ChatGPT work or yeah. Yeah.
21:45He gave it he gave it its own email address, but it's like a sub email of his email address. So it's like Dan plus agent name and some numbers so that it could check messages so his team could email it in.
21:55Here's the thing. He's basically giving it access to his to his email. The way I would do something like that, and he might have already, is I would use Zapier MCP.
22:03And the reason I would use Zapier MCP is since it's his own personal email, he could say, read my email, but don't write it. Draft email for me, but don't send it. That kind of granular control from a third party tool that's trusted is really helpful, not just for him.
22:17For first team, he may never make mistakes, and so this is not an issue, but probably does. His team probably will make mistakes and could use this.
22:25And so I like Zapier MCP as a way of taking all these tools that I have. Again, it's over 9,000 tools and put it into one bundle that I can give my AI and be, uh, and trust it. If you use it, go to zapier.com/mcp.
22:38Try it out for free, and let them know you heard about it from Andrew Warner or the next new thing. Alright. Next.
22:43Oh, okay. One last one about video. This is a very impressive, uh, video.
22:49Watch this. Five point six sold this prompt, walked away, and when I came back, I got this. Okay.
22:54So you're looking at Nate, and you're hearing Nate. But Nate never stood in front of a camera for this. He didn't record this narration, and he never opened the editor.
23:01He gave me one prompt. That's it. I'm GPT five point That's cool.
23:06Impressive. Right? Mhmm.
23:08Let's look at pricing. We can see that the Soul pricing is much cheaper. It's basically half of Fable five.
23:15So g p t five one six sole is similarly priced to Opus 4.8.
23:21That's one of the things that just keeps coming up also. Lower price, um, but it seems like Fable is the one you wanna reach for when you want the really strong work and GPT five sixty
23:35say strong work. I would say, um, like, the maybe, like, the better understanding of certain nuance in in what you're trying to talk But I think, like, yeah, like, the actual implementation, like, the the reliability of the implementation, I think I would still even trust 5.6 on that.
23:52Wow. And it's gonna be more affordable at the same time. I think What do you mean by reliability?
23:58Like, so if if the code's gonna be correct and it's gonna pass the tests, like, I think OpenAI is very much designing these models of, like, how can we make this work for enterprise and what and what people are trying to accomplish with these tools. And so, like, g p d 5.6, even Soul, is undoubtedly a smaller model in size than Fable five.
24:21But, yeah, I think Open AI really realizes like, okay, how can we make these efficient enough that we can serve them at a scale? And I think that was a challenge for Anthropic with Fable five to begin with, and why they they had like, okay, well, there's mythos and they say it's dangerous, but then they they have Fable five now, but the the pricing just seems, uh, very expensive in comparison.
24:45I I wonder I wonder where this is what's gonna happen a month from now when, uh, enterprises have really had time to spend a lot of time with each of these models, um, if they will actually think Fable's worth it or not.
24:59Let's look at, um, this is Dan Shipper.
25:04I like his feedback on writing because he cares about things a lot. Fable.
25:08Both 4.8 and Fable have this tendency to overexplain, to be a little bit literary. Fable, in particular, often runs for so long that it creates almost its own private language that it ends up speaking in. And g p d 5.6 is just to the point.
25:23It's simple. Simple. It clearly expresses the thing it needs to express.
25:27It doesn't have a ton of AI isms. I use it, for example, to compose emails and, like, here's an email. Tucker, 04:30 ET on Tuesday the fourteenth works for me.
25:35Hope that still works on your end. Looking forward. It just
25:38I like also that it's so natural that it has lowercase,
25:42uh, letters, Tucker, Mike, Navin. It's all lowercase. Our head of growth, Austin, uses it to do marketing emails, and he's like, this is the first time that I can actually one shot marketing emails with this thing with any model.
25:53I'm very impressed by its ability to be such a good programmer and to have this level of, yeah, I actually wanna talk to it.
26:01It's a good writer. It doesn't overthink things. It doesn't overdo it.
26:06It's fantastic.
26:08I've gotta It makes me curious to to try it for more writing things and and see how it responds. Because historically, I think GPT models have been worse than anthropic models for for writing tests.
26:20I know Dan But is involved in, like he spends a lot of time writing. And so, um, if he thinks this, uh, yeah, I'm curious to try that. My challenge is that we don't get to see enough of these writing tests because they don't demo well, which is why we see endless video games.
26:33You know? Everybody's creating a video game. Alright.
26:37Look at this. I was trying to use it to make an image of this way of working that I think 5.6 makes This is available. GPT five six.
26:44It made this. And it's fine, but it's I don't know.
26:49It's too complicated. It doesn't look well considered. Same prompt.
26:54Look at what Fable did, one shot. Same image model too.
26:59That's what Fable looks like. Fable is just playing on a different level. They're both using the GBT image model, but the way they prompt it is so different that it makes they will make stuff like this.
27:10Five point six makes stuff like this. So if you care about designers
27:13Yeah. That's interesting. So the so they both wrote the prompts for the GBT image model.
27:18But, yeah, what what Fable came up with there, I do like better. It's more simple, and this is the thing you would want to look at versus the first one.
27:27It it tells you it tells you the thing, but it's it's a bit more complicated to to wanna visualize that. Yeah.
27:36Let's look at his personal app. It also works for your personal life. For example, I have this app right now that it just takes all the meals I've eaten.
27:44Anything that I leave in a voice note in monologue or anything I take a picture of in Apple photos. It just grabs it from my photos, figures out the macros, and then records it for me. And five point six just does this working in the ChatGPT Codex app in a loop.
27:58I do this all the time. I use it for buying stuff on Facebook Marketplace or helping me decorate my apartment. These are all the things that you can do if you have it running in a loop doing work for you that were previously not possible or it would require too much work.
28:11It just wouldn't be worth it.
28:13He's basically living in that Codex app, and he has talked about how he'll take a picture of his room and ask, uh, GPT 5.6, how would you furnish it?
28:22Show me pictures and gotten some really good responses from it. He's he's making me want to not making me want to. I am now because of him, uh, keeping that app up and running and trying to do as much of my work through it as possible.
28:35End of it. Let's take a look at this. Six.
28:37In the Chateaubiti Codex app, ChatGPT Work app. It's the gold standard. It's where I and most of the team spend most of our time working with AI.
28:46Alright. I gotta do more of that. Uh, let's look at Matthew Berman real quick here about costs and different model levels.
28:54The pricing is also much better for GPT 5.6.
28:57Not only is it less expensive, but it you've uses been saying to tokens Yep. To get to the same results. So $5 per million input tokens versus $10 for Fable, much cheaper on cash hits, and 30 versus $50 per million output tokens.
29:15This is basically gonna be the model that I'm gonna use day to day. And then Fable, I'll reach out for occasionally when I have a really big project that I wanna bring in the big the big guns for. I I know you don't agree with it, but that's that's what I'm thinking about this.
29:28I mean, I I would agree, but, like, I would give I would always give the first try in
29:325.6 probably. I see.
29:35Unless there's some specific reason I thought, oh, I I believe Fable's gonna be better with this from the start. Because I think a a lot of the problems that we have can just be solved by 5.6. And and even even if you say, okay.
29:505.61, not good. I'm gonna throw it out.
29:53I wanna try Fable.
29:54Because it is so much cheaper, you can still just try it, throw it out, and then then try Fable. The thing that is very different about GPT 5.6 versus Fable. It comes in three different model sizes, Luna, the smallest, Terra, the medium, and Soul, the largest.
30:08But even within those, you have multiple levels of reasoning. And if we choose Soul, you get all the way up to ultra, which is, you know, basically a quota burner.
30:23I always forget about those. I stick with one, and I just mindlessly continue. I like that.
30:28Last one from Bijan who always seems to, whenever he checks out a new model, has it create a drum set. It is usually a really nice drum set.
30:38This one is next level. Check out what he did. Make a vision pro app that is a virtual drum kit.
30:43I want to be able to hit the drums with my hands and play them in mixed reality. The AVP is gonna That's cool. He's doing it for the VisionPRO.
30:50Install. It started out and gave us this, which was basically just kind of like a two d thing. It wasn't necessarily what I had envisioned.
30:57I wanted something three d, so I told it that I want a three d photorealistic kit, not some on screen two d. So, apparently, this is running now in the Vision Pro.
31:07I have no idea. I have not looked at it yet. The external battery on this device is you know, some of the decision was made.
31:14Alright. So it even put an actual, like, app icon for this.
31:20So let's just see. Okay. We have sensitivity.
31:23We have kit height. Let's just okay.
31:28That is getting insane in the title. Don't tell me I can can I?
31:34Okay. So you cannot I'm gonna destroy everything inside. This is not this is this is really if someone walked up here right now, they'd be like, oh.
31:47That is cool. That is. I'm so good.
31:51I gotta try making something for the vision pro because
31:54I can imagine it's even more of a surprise when you get to to put the thing on and experience it rather than just, uh, see the the web page or the the flat thing that it generated. Absolutely beautiful. I love seeing this.
32:06I love seeing all the examples. Alright. Now that this video is done, I've got another one for you to check out right here below, and I'll see you in that next one.
The Hook
The bait, then the rug-pull.
Two hosts spend thirty-two minutes doing something most reviewers won't: watching seven other creators' independent tests of GPT-5.6 Sol against Claude Fable 5, then arguing about what the pattern actually means. The verdict that emerges isn't a winner — it's a division of labor.
Named ideas worth stealing.
05:54concept
Manager vs Worker
- Fable 5 = manager / planner
- Sol 5.6 = worker / executor
The hosts' recurring mental model for combining two models on one job: use the pricier model to plan, direct, and review, and the cheaper model to execute repeatedly.
Steal fordeciding which AI model handles which step of a multi-step agentic workflow
17:24list
Six use cases put to the test
- Build an interactive travel website
- Create a 3D retro space shooter
- Edit and publish short videos
- Add a new feature to a mobile app
- Get life and business advice
- Upgrade my personal AI OS
A structured comparison grid one featured creator used to test GPT-5.6 across varied task types instead of judging on a single coding task.
Steal fordesigning your own multi-domain model bake-off instead of judging a model on one task type
How they asked for the click.
VERBAL ASK
21:27product
“go to zapier.com/mcp, try it out for free”
mid-video dedicated sponsor read woven into the panel discussion by the host, framed around Zapier MCP's scoped read/write permissions rather than cut in as a separate ad break
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.
