The Startup Ideas Podcast - Claude Opus 4.6 vs GPT-5.3 Codex: Live Build, Clear Winner

Episode Date: February 6, 2026

I sit down with Morgan Linton, Cofounder/CTO of Bold Metrics, to break down the same-day release of Claude Opus 4.6 and GPT-5.3 Codex. We walk through exactly how to set up Opus 4.6 in Claude Code, explore the philosophical split between autonomous agent teams and interactive pair-programming, and then put both models to the test by having each one build a Polymarket competitor from scratch, live and unscripted. By the end, you'll know how to configure each model, when to reach for one over the other, and what happened when we let them race head-to-head. Timestamps 00:00 – Intro 03:26 – Setting Up Opus 4.6 in Claude Code 05:16 – Enabling Agent Teams 08:32 – The Philosophical Divergence between Codex and Opus 11:11 – Core Feature Comparison (Context Window, Benchmarks, Agentic Behavior) 15:27 – Live Demo Setup: Polymarket Build Prompt Design 18:26 – Race Begins 21:02 – Best Model for Vibe Coders 22:12 – Codex Finishes in Under 4 Minutes 26:38 – Opus Agents Still Running, Token Usage Climbing 31:41 – Testing and Reviewing the Codex Build 40:25 – Opus Build Completes, First Look at Results 42:47 – Opus Final Build Reveal 44:22 – Side-by-Side Comparison: Opus Takes This Round 45:40 – Final Takeaways and Recommendations Key Points Opus 4.6 and GPT-5.3 Codex dropped within 18 minutes of each other and represent two fundamentally different engineering philosophies — autonomous agents vs. interactive collaboration. To use Opus 4.6 properly, you must update Claude Code to version 2.1.32+, set the model in settings.json, and explicitly enable the experimental Agent Teams feature. Opus 4.6's standout feature is multi-agent orchestration: you can spin up parallel agents for research, architecture, UX, and testing — all working simultaneously. GPT-5.3 Codex's standout feature is mid-task steering: you can interrupt, redirect, and course-correct the model while it's actively building. In the live head-to-head, Codex finished a Polymarket competitor in under 4 minutes; Opus took significantly longer but produced a more polished UI, richer feature set, and 96 tests vs. Codex's 10. Agent teams multiply token usage substantially — a single Opus build can consume 150,000–250,000 tokens across all agents. The #1 tool to find startup ideas/trends - https://www.ideabrowser.com LCA helps Fortune 500s and fast-growing startups build their future - from Warner Music to Fortnite to Dropbox. We turn 'what if' into reality with AI, apps, and next-gen products https://latecheckout.agency/ The Vibe Marketer - Resources for people into vibe marketing/marketing with AI: https://www.thevibemarketer.com/ FIND ME ON SOCIAL X/Twitter: https://twitter.com/gregisenberg Instagram: https://instagram.com/gregisenberg/ LinkedIn: https://www.linkedin.com/in/gisenberg/ Morgan Linton X/Twitter: https://x.com/morganlinton Bold Metrics: https://boldmetrics.com Personal Website: https://linton.ai

Today is a massive day because Anthropic just dropped Opus 4.6. And OpenAI answered with GPT 5.3 Codex. But what is the better model? And how do you get started? And what are some tips and tricks to get the most out of them? Well, this episode is all about that. This is for the technical person who's trying to get the most out of these models, who don't just want hot takes, who want tactical sauce for getting the most out of these models.

This episode of the pod is with my dear friend Morgan Linton. Morgan is one of the best engineers I know. He was an executive at Sonos. He's invested in a lot of AI companies, and he's building an AI company of his own. He's one of my first calls when I'm like, hey, which models better? So we put the models head to head,

and there's a winner at the end. We rebuild Polymarket, a multi-billion dollar app, but we use these models. So which is the better one? You'll find out by watching this episode, but you'll also learn to become a better AI developer because you'll have these tips and tricks in your back pocket. I'm with one of my favorite people, Morgan Linton.

You might not know him, but he is just, you know, just an incredible developer, founder, entrepreneur, investor. He does it all, but today, you know, what I needed him to help me understand is Opus 4.6 just came out. GPT 5.3 Codex just came out. Morgan, help me understand. And by the end of this episode, what are people going to get out of this? Yeah. Well, Greg, thanks for having a super exciting day. It's moving fast today.

Opus 4-6 came out and then Sam Altman put together a quick tweet. I want to say like maybe 18 minutes later announcing GPT-53 Codex. And me, I think everybody else has been jumping on it, playing around, figuring out the differences. You know, all the little neat new settings that there are in each of these. By the end of this, you're going to know first how to make sure that you are running opus 4.6 and all of the little details you can change in the settings on JSON file to use some of the cool features in Opus 46, especially Agent Teams, which is probably the feature I'm the most excited about. It'll also understand why you might use one versus the other, because they

both kind of tackle different engineering methodologies. And then hopefully you'll see some cool stuff as we build some demos together that I've put together that I haven't tried myself. so I'll be trying just live with you. So we'll see how that goes. Cool. I think one of them is we're going to try to recreate Polymarket. Yes. And see which model performs best.

They're going to have both. They're going to do a head-to-head to try to each build their own version of Polymarket. So by the end of this episode, you will have a pretty good understanding of how to use the models, when to use the models, how to get started. Morgan, let's get into it. Cool. Right on. All right. All right, so I took some notes, and essentially, you know, with 5-3 codex, I'll be showing that in the desktop app on Mac because they're super excited about that.

I'm excited about it. I think if Open AI was wanting a demo to be done the right way, they would want me to do it in their app. Whereas with Opus 4 or 6, I would say the Anthropic team would want me to do it in the CLI. And so there's a few different configuration settings that you do want to make sure that you get right when you're. you're using Opus 4-6, we're trying to use Opus 4-6 today, tomorrow, whenever it is that you're jumping into use it. I've seen a lot of people online today on Twitter saying, it's weird. I'm having a problem like, it's supposed to be agent teams, but I don't see them or how do I know what version I'm running? So I thought, let's start by just giving everybody a level playing

field to know, okay, I want to be able to use Claude Code with Opus 4-6. How do I make sure I'm doing that and doing that correctly? So here's kind of the initial to-do-do-do. that everyone should have on their list. Just do an NPM update, see if that does the trick. If that doesn't, and you're running an older version, then run Clod Update. But you should see, like, as of right now, it's 2.1.32. If you see 1.something, you're running an old version. And then what you want to do is go into your settings.json, and I'll just show this here.

So if you just do like CD tilda slash dot clod So I bet that there's people who are running the old model They don't even realize it I'm probably getting bad or no idea Yeah yeah so I mean make sure you go in here CDtledslash dot clod Here's your settings dot jason

If you view this Here's essentially what you should see Now it's okay it can be model if you want to like really be specific about it, you can put in Claude dash opish dash 4-6, that'll lock it in. But because 4-6 is the newest model, you can also just put in model and just opus

and that'll work. The key thing that you want to do is, in my opinion, the coolest feature that they added with 4-6 is agent teams. I'm super excited to demo that with you. You have to make sure to turn that on because it is an experimental feature,

And that's probably the biggest confusion I'm seeing people have today with Opus 4-6 is that they are running Opus 4-6. They keep hearing about agent teams, and they're giving it prompts like, build a team of agents, do this and this, and it's not quite doing it. And that's because you do have to enable this. So you do have to add in end this ClaudeCode Experimental Agent Teams and then set it equal to one. Okay. Nothing too crazy.

Once you do that, that will make all that possible. So with that in mind, you're pretty ready to go there. Then you can just run cloud in the terminal and you're good. For people that are using the API, the one thing I did want to point out is there's a pretty cool new addition, which is called adaptive thinking. Also, just to be clear, because I'm seeing confusion on this too, this is in the API. This is not in Claude Code itself. but adaptive thinking just to show it here, you're able to essentially pick the level of effort

that you would like the model to use. This is only going to work in 4-6, by the way, if you want to use like an effort level of max. And so here's kind of the different levels. So with Max, Claude always thinks with no constraints on thinking depth. It's Opus 46 only. So requests using Max on other models are going to return an error. So if you're calling the API and you set the effort

level to max and you get an error, then you're probably not using Opus 4.6. But here's the example where you can see, if I'm calling the API, I set the model to Cloud Opus 46, and then here's where I can set the effort. And this is another thing. If you're using existing API code, you may have the model of Opus 4-5, and now you adjust the effort to max. It gives you an error. All you do is just bump the version and you're good. But this is kind of a neat thing they've added to the API with 4-6. It's worth mentioning. And then kind of the last thing I would say is just if you want to use split panes for agents, so if you want agents to show up in different panes and you're using something like warp,

just make sure to install T-Mux. You can do this with Brunstall T-Mux. And then if you do that, it's going to default to auto, which usually means in process, which means in that same terminal window you have, the agents are going to be working all together. If you want it to split pane, then you just need to update that setting and the settings.js on the split panes.

I'm not going to go into super details on that, but those are just like, I think, good housekeeping to start with for anyone using Opus, but don't worry about it. Really, all that anybody needs to do, especially if you don't even want to use teams, agent teams,

is just make sure you're updated using the newest version and that the model is Opus and it'll be using OPA. Opus 4-6. Cool. So that's that. Before I get into kind of the differences between Opus 4-6 and Codex, I thought I would actually read this because this was posted on Hacker News four hours ago, and I was reading it.

I was thinking that's like the best way to explain it. So I'm just going to read this little section here because I think they do such a good job with it. This person's saying what's interesting to me is that GPT-5 through an opus-6 are diverging philosophically and really in the same way that actual engineers and orgs have diverged philosophically. I think this really nails it. With Codex 5-3, the framing is an interactive collaborator. You steer it mid-execution, stay in the loop, course-correct as it works. With Opus 4-6, the emphasis is the opposite, a more autonomous, agentic, thoughtful system

that plans deeply, runs longer, and asks less of the human. That feels like a reflection of a real split and how people think Ellen, based coding should work. Some want tight human and loop control. Others want to delegate whole chunks of work and review the result. And I honestly, I think that says it beautifully. And I think that nails the differences. And also, hopefully, you know, everybody wants to pick a winner where it's like, oh, no, no, Opus 4-6 is better. Codex. It's different. It depends on what your methodology is. And I think what we're seeing now, not just with vibe coding, but also with like overall, like, AI-powered engineering is how do you want to work with agenetic coding? Do you want to have a totally autonomous experience where you're sending agents out to do work?

Or do you want to work with an LLM like another teammate and pair program with the LLM? And that's where you're now seeing a divergence where I think you're going to see a lot of teams using both. because Codex really is your collaborator and what they've added with 5-3 is like really good mid-execution steering. Whereas with Opus 4-6 it's probably the best of the best now

being able to say, I want to spin up three or four agents, I want them to go do stuff. Hey, and don't bug me. I want to trust they're going to do good stuff and it's able to deliver. So are you saying that there, in some ways,

it's just a preference, like, depending on how, you know, there's no right or wrong, basically. You know, not wrong to be an opus person or, you know, it just like might feel, yeah, it's just a preference. Yeah, well, and you might be both, right? That's true. It might turn out of your both. That's why, like, not to disappoint people here, but we're not going to end this with me saying, and so the winner is, it's like, well, depends on what you want to do. Everyone has a different methodology for it. I'll dive in and try to make this part fast

because I know the fun part is probably us going in and playing around with both of these and having them do a head-to-head and try to build a competitor to Polymarket and however much time we have. But I'll just start kind of going into these at a high level just so for anyone wants to know like what are the core differences?

Why is this so interesting? Just going to what that is. So with Opus 4-6, much bigger context window. So you have a million token context window here. Very strong coherence over entire documents and repos designed for, you know, like load the whole universe and reason over it. 5.3, they talk about large context, but it's not a headline feature. And I actually went back and forth of it to get it to actually give me a number. And the number is around 200,000 tokens, which is not that impressive. That's smaller than I was thinking it would be.

But that's okay. It's optimized, you know, for progressive execution rather than total recall. So that's why that's not as important. And, you know, optimized for deciding, like, what to keep in working memory. So high level, what that means is Claude is better when the task is understand everything first and then decide. GBT3 codex is probably better when the task is decide, fast, act, iterate, more of that, you know, paraprogramming, you know, mid-t task change behavior. For coding benchmarks,

Ops 4-6 is really good at code-based comprehension, refactors with like architectural sensitivity, explaining why a system behaves a certain way. And then, you know, a little less tendency of this like YOLO write code, right? Right. That's good. Which is, I think, something everybody wants. Yeah, exactly.

So, you know, that's good for everybody, but especially for vibe coders that are getting started and they may not be able to identify hallucinations. Opus 4-6 is definitely going to perform better there. But then for teams, you know, building in large code bases like me and my team are doing, that's also really important. So kind of a win for everyone there.

5-3 codex did win on SWEBBBENCH Pro, Terminal Bench. Overall, it's like scored better on coding benchmarks. So probably better end-end app generation. and, you know, Claude's kind of like senior reviewer, staff engineer, GPT-5-3, you're probably like your founding engineer, right?

Agentic behavior, Opus 4-6, this is the key one, right? It's like the multi-agent orchestration. That's probably like the bleeding edge feature in 4-6. And then with 5-3 codex, really like test-driven autonomy, build, test, modify without being asked,

but then this task steering. You can watch it. You can go in. It's like your buddy's coding and you can say, oh, wait, wait, man, wait, why are you doing this? And you can stop it and it'll go, okay. And then you can restart. You can really fix things in line.

Much harder to do that with Opus. With Opus, you'll kind of be stopping it and then starting somewhat fresh. But it has a pretty big context window, so it knows what it did. But, you know, Claude's really asking, like, should we do this? GPT-5-3 is like,

how fast can I ship? this, right? It's really, I mean, it's so cool because it almost feels like they're different people. You know what I mean? Like they have different styles. Yes, totally. Yeah, it's a good way to look at it. It's like a different, it's like a different personality type, right?

And then, yeah, failure modes, you know, Claude 4-6, it might overanalyze. It's got a much bigger context window. I can hesitate when requirements are ambiguous. And then it can stop short of full execution. 5-3 codex, could be overconfident, can lock in a flawed assumption early, but you can steer it back in the right direction if that happens.

So that's kind of a high-level overview on the two. Cool. That's helpful. Yeah. So should we just dive in? I haven't tested any of this. So this is not, I have like zero canned demos because I thought it would be more fun just to try something together and see what happens. So should we, should we try it? Yeah. Okay. So let's see. I'm going to, I'm going to start with Opus,

and I've got these prompts preloaded. So I'm giving different prompts, just like I think you said it really well. It's like you're talking to different people. And so, you know, when I'm talking to Opus, I can tell Opus, build me a team, and here's what I want each member of the team to do. When I'm talking to Codex, I can't really tell it to build me a team, but I can tell it to, think about stuff. So the prompt that I'm going to give to Opus is build a competitive polymarket, create an agent team to explore this from different angles. One team made on technical architecture, one on understanding polymarket and the ins and outs of prediction markets, one on UX, and one that just works on building really good tests to make sure everything works.

For Codex, I'm going to give it a little different prompt, but very similar. So I'll still build a competitive polymarket, but now think deeply about technical architecture. understanding polymarket and the ins the outs of prediction markets, good clean UX, make sure it builds a really good test, to make sure everything works. And to be fair, I'm going to try to pace these in and around the same time.

You're a fair guy, Morgan. I'm trying to keep it fair here, right? It's the only way to do it. Like I said, no winners or losers. It's just about letting everybody have a fair shot to play the game. Yeah. All right. So let's see.

I'm going to make different directions. So I'll do, let's just call this Opus 4-5, Jolly Market competitor. All right. So let's fire up clot in here. By the way, if you want to check when you're running, just to like really make sure that you're in a good place with the model, if you type slash model, I can see here, right,

Cloud Opus 4-6. All right. So I'm good there. I'm going to take this prompt. copy it make sure this is all copied incorrectly okay got that

I'm not going to hit enter yet making this totally fair I don't want anyone at anthropic or open AI to get upset with me so I want to be in good terms with both of them totally smart guy

let's see oh wait actually you know what I do want to create a new folder for this but we are keeping it real we're being objective neither myself or Morgan are affiliated with either Well, actually, I don't know about you. I'm not.

No, I'm not. I'm not. I'm not. I'm not. Nope. Nope. Nope. I love them both equally.

How about that? Yeah. Okay. And I'm going to try to start as close to on the same time as I can. Enter. Go. All right. They're going.

Off to the races. So what do you think is going to happen? That's a great question. Well, I know right now because I told, uh, Opus 4-5 to build using different teammates, it's going to do that. So you can see here it says, I'll build a polymark competitor by launching parallel research agents first, then synthesizing their finding to a comprehensive implementation planning code base.

This is brand, brand new, right? Like if I did this with Opus yesterday, wouldn't be possible. That's kind of the difference here is that the way that, um, the way that, um, that Codex is working is the way things have kind of always worked, right? So if you see, this is like the individual person, right? It's not saying, okay, I'm going to launch all these different agents and compare what they say. It's like, okay, I'm going to inspect the workspace.

This is your, you know, really detail-oriented, really senior, like founding engineer, like that example gave, right? Whereas over here, you can see it's already launched these agents. And now it wants to do web searches, and I'm going to let it do that. So multiple agents were asking to do web searches. So now launching all four research agents in parallel. So this is off and I've got, you know, my technical architecture agent. I've got this other agent that these are both doing web searches right now.

So one is looking at like prediction market order book matching engine architecture. So this one's learning about engine architecture for prediction markets. This one's looking at polymarket how it works, barn in particular market mechanics. And then I've got the U.S. is doing some design research, and then we've got some test research. Okay, now it's going to go to polymarket, and let's really hope the polymarket doesn't block it,

because that'll make things harder for it. Meanwhile, over here, this has discovered code. It's figured out the repo is empty, so it's going to scaffold it from scratch. And it is starting to, I'm now wiring the core market math and trading engine. So it's interesting, right? So you've got Codex,

is out here building and is like building the engine with Opus 4-6 it still has agents out there like doing research work. Yeah, you really start to see just like how different they really are. Yes. As they make progress. Yeah.

Like I said, I haven't tested it before, so we don't know how long it'll take each of these. Totally. Yeah, and I think like, I guess one question I have is like, is one, is one model better for being more of a beginner non-technical vibe coder

or, you know, doesn't really matter? Yeah, it's a good question. I mean, I think the fair answer would be probably Codex because Codex edged out Opus 4-6 a little bit

on some of those coding benchmarks and is kind of known for writing better production code, probably codex in that way. At the same time, one of the downsides, and like I said, I could only do this in a totally balanced way because they're so different. You know, at the same time for a vibe coder,

knowing when to interject and stop codex and say, oh, wait, you're doing this this way, can you instead look at doing it this way? They're probably not going to know how to do that, right? And so that's where maybe Opus 4-6 is better, where you could say, okay, spin up four or five agents and let them work with each other, right? Yeah.

Okay, Codex is done. All right. So Codex built a competitor to Polymarket in three minutes and 47 seconds. And to be clear, Polymarkets a multi-billion dollar company. Yeah, I don't think this will work quite as well. But we'll see. Let's see.

So let's just check out if it worked first. I'll let this keep running here. So, you know, it'll tell you at the end here. It actually did the testing so you can see it built a test suite. So it has an LMSR math unit test suite, an engine behavior unit test suite, and an API integration test suite. And it passed with 10 out of 10 tests. As far as what it built, it has this core LMSR market maker engine. So coherent pricing, slippage, bound loss behavior, domain trending engine.

It built a Rest API router, which is kind of interesting because I didn't tell it that it would have to build obviously any of this in any way. It figured out the architecture on its own. Clean responsive front end. All right, well, let's see. Let's see if it is, actually. So let's go here. I'll let this keep running. This has got these four agents just running away here.

And I'm in here. I'm going to do 1 p.m. test. All right, tests 10 past 10. That looks good to me. Oh, yeah. PM start. All right, it's running. Let's see.

Okay, here we go. So this looks like it has the ability. So, Greg, let's make you, we'll make you the first trader. All right? Say add. Okay. you got a thousand bucks okay there we go not bad

all right what uh what market do you want to create um well bitcoin I think as we speak is crashed to what 63,000 or something something like that yeah so I do like the I mean will BDC be about

you know be above 110k by yeah okay by desk 31 that's pretty good yeah okay so let's I mean, that's almost double. Yeah, that'd be pretty good.

It depends. It depends when you bought it, you know? If you bought it at 125K, then you're not so happy. Let's see. So then I don't even know what resolution criterion source would be. I mean, I think I know what it's getting at, but I guess you could say like, why don't we say use coin market. cap as the source and resolve

by looking at the price on the last day of December just before midnight I guess I guess like the price of BTC yeah all right okay it looks like it's okay so we've got it now

so we'll use Quaymarket cap Okay. So then you can do a yes, 50%. So what do you think? Yes, yes or no? I mean, this isn't financial advice. This is just purely educational purposes. But I think so. I think that. All right. That's a yes for Greg. Buy. Let's see. How many shares you want to buy? You've got a thousand bucks. I want to put it all. I'll put it all on this. I don't know how much it is per share. Let's see if it's 1,000, if that's right. Okay, yeah. Okay, trade executed. Okay. So, I mean, it seems like it built something, you know, as a prototype, relatively functional here. I guess that it actually has decremented. So, okay, a thousand shares was not. That ended up being, you know, about $24 that you spent. So you've got more money if you wanted to create another market. But it worked. It's not returning an error. It shows the volume here.

Interesting. All right, so let's go back. Let's see. So far's really good with that, I'd say. Let's see what's going on here. So we've got, okay, so first off, look at how many tokens. People have been talking about how token-hungry opus is, and it's very token-hungry. Each one of these agents has used over 25,000 tokens.

So let's see, though. So they finished, right? The technical research around architecture is done. Prediction market research is done. The U.X design research is done. The testing strategy is done. Now it's going to go and build. So it's writing the package JSON.

Did you see the ad that Anthropic launched about ads? Yes. I watched them all. They're hilarious. Although actually I guess Sam was not very happy about them today. I saw a tweet from Sam that was less than happy. So I found them hilarious, but I also understand his side as well. So basically, it seems like Anthropic is sort of anti-ads for now.

And the chat GPT is going to be introducing ads. Yes. And, you know, when I'm watching this and I'm seeing you're going through 25,000 tokens, 25,000 tokens, 25,000 tokens. I'm like, yeah, of course Anthropic doesn't really... Yeah, exactly. You know? Yeah, yeah. I mean, this is literally, I mean, if you add that all up,

you're talking about over 100,000 tokens used in doing this. So I think that's one of the very good things for, like, investors in Anthropic, right, is with agents and agents now being, I think, probably the new killer feature in Opus, you're going to take whatever token usage and multiply it by the number of agents. Exactly. It's actually really smart. And I wonder if that was like the thinking.

They're like, how can we get people with more tokens? Oh, we'll just like spin up agents and we'll design it like that. Or did they think like, okay, how can we design a system that is best for the use case? And then they're like, then we'll monetize it like this. I don't know. Yeah. Yeah.

Probably a combination of the two. I can tell you I've never used so many tokens in one day. as today. So it's working. 100,000 tokens is like roughly how much in U.S. dollars? I don't know because I have a, I have a Claude Max plan. So, yeah, so I'm not paying. We're not seeing it hit any limits right now, right?

So I'm not paying more than $200. I can tell you that. Yeah. My guess is it's, you know, we're talking, like in the $200, MaxPen, do you remember how many tokens you get approximately? That's a good question.

Let me fire up, let me fire up Claude and ask it. Let's see here. How many tokens do get estimate? Let's see. Okay, so here you go. Estimate.

Estimates. So 45 million tokens per month of Sonet. But let's see. What is your estimate for opus or six it's like they don't really want you to

know no they're trying to make a little harder okay yeah they're not even going to tell me actually they're just they say there's no public data it's very new opus is roughly 5x more expensive so then uh if it's 45 million that's 5x so 10 million is probably the answer about right yeah yeah um so then if we're doing you know quick math let's just say we spent 100,000 tokens.

You know, 100,000 divided by 5 million is, you know, we're going to spend more than that because look at this. We're now over 17,000 tokens on top of that in this next build. Okay. So, but still, let's say, you know, even if we use a million tokens building a competitive deployment market right now, we're still only using a 10th of what. it, but he can do. That's not terrible. No, I mean, it's $20, which is like the price of a cocktail in Miami.

Yeah, yeah, exactly. Yeah. Yeah. So let's say, but now for, as I say, I'm watching the tokens. Ah, creep out. All right, so it's building the API routes now. I have a feeling this is going to be a better end result. I was actually just going to say that.

This feels, and then maybe it's just because there were four agents that were doing all the work beforehand and now it's doing the work. It feels like we're going to see something very different when we load what it builds. Yeah. I don't think we gave it any like design, like visual design, any, you know, so do you recommend for folks to just like sort of get the MVP out, out, play around with it on local host, you know, click some buttons and then sort of update with the visual design from there?

It's a good question. I do like 50-50. Sometimes if I have something in mind, especially if I want something like on brand with something like, suppose I'm building something that is going to be in the like open claw multbook ecosystem. I would probably say, hey, I want to design a site that, you know, looks somewhat similar to or is inspired by, you know, open, open, and claw. Dot AI and moldbook.com. Right. Take a look at those sites and get inspiration.

These models are great at doing stuff like that. Cool. I'm really excited to see what this is doing

though. I think we're now like well over 200,000 tokens. They saw what I could tell.

But we're not of 10 million. We're not heading any limits. You know, we don't have to take out a second mortgages on our

Yeah, yeah, yet. Yeah, it's still going, though. I guess, you know, we can tell, like, here's an interesting thing in a comparison. Like, this is still going, why don't we say, like, the design, because the design looked kind of bland to me, right?

Yes, it did. Can you spruce it up and make it look nicer? Because, like, we may as well have codex working away, too, right? Yeah, so you didn't really give it any, like, specific, it should look like square.com. No. Now, we'll see, we'll basically see if codex, if, you know, if 5-3 has a little bit of taste. Yeah, yeah. So that's what it's saying now.

It's saying, okay, I'll upgrade the visual system without changing functionality, stronger typography, richer color direction, better card hierarchy, and purposeful motion. I don't know what that means, but we'll find out. All right. Okay. So now it's editing. index that HTML. It looks like he's going to add motion hover polish.

Okay. Totally of that. This current task for over 30,000 tokens building the front end UI. Okay, it's done. Yeah, Codex is fast, by the way. Right? I mean, that's pretty darn fast.

Yeah. So we should be able to just go here. It should have already automatically reloaded. Okay. All right. I mean, not that difference. It's not that different. I think, can I try something?

Yeah, go for it. I'm going to say, I would say, okay, thank you, but this was a minor design refresh. I'm looking for a major one. There you go, yeah. And then I'm going to say, pretend you are Jack Dorsey. Here you go. And how would he design this website to be clean, elegant, and full of interesting interactions?

Yeah, great, yeah. Jack Dorsey for people that don't know, co-founder of formerly known as Twitter and Square Block now. He's just got, he's a design guy. I don't know, he's first one guy who came to mine or first person who came to him. Yeah, that's a good one. That's a good prompt. Let's see. So I'll do a full visual re-architect, not an incremental tweak.

New layout, language, stronger typography, monochrome first palette. Interesting. Interaction-driven cards. That's a da-da. Okay. You know what's interesting is it didn't, I would have kind of hoped, and maybe we're not quite at AGI yet, I would kind of hope that it would say, let me go find some art. You told me that, Greg.

Like, hey, Morgan, can you read it? I would be like, yeah, let me go look at some articles about Jack Dorsey's design aesthetic. Exactly. I'm surprised he's not doing that. Instead, it's going like, I am assuming it knows who Jack Dorsey is, although I don't know if it actually does. It just seems like it's really just taking like this part of your question and going, oh, okay, major refresh. I'll do that? Well, can't you ask it?

Can't you say, do you know who Jack Dorsey is? Let's see. I can actually, I'm supposed to be able to in the middle cut it off, so let me see. Yeah, do you know who Jack Dorsey is? Let's see. Okay, so here we go. This is the midstream test. It's thinking about it.

43,000 tokens over here. Okay. There you go. Yes. Jack Dorsey is the co-founder of Twitter, Form, the Docton Square. Okay, with a design style that's typically minimal restraint and interact. action focused. Beautiful.

Okay. All right. So, tushay. It showed us. Yeah. Now, here's the weird thing. It looks like it's... Is it complete?

Or do we have to say... Did it complete? Like, right when I was saying that? Are you done or did you stop because I asked a question? Yeah, this is really interesting. I will say, like, oh, I pause when you ask the question. The major redesign mostly... If you want or resume, well, so that's weird.

So you ask you the question, it just stops. But like, so like, yes, of course, continue. Yeah. Okay. So that's actually some weird U.S. Like, it obviously should just continue after. Right. Yeah.

I would assume that it's such a weird thing as it said, yeah. If you want, I'll resume now. I will say I do like that you can in midstream, like kind of edit things. Yeah. Like that's how my brain works. Totally. Yeah. Yeah.

This is using a ton of tokens. It is amazing to see the detail. I mean, this should be a work of art, whatever site, this comes off. All right, this is done. So now let's see. We can go back to this. And, okay. I mean, I'm not blown away, but it's okay.

Opinions become price in milliseconds. Trade conviction, not noise. Signal market is done for fast. This is iteration with transparent pricing. Okay. I mean, I would push it more, I think. Yeah. Yeah, I guess, you know, like, yeah, go for it.

I would say, that's not the Jack Dorsey I know. I don't know Jack Dorsey. Yeah. I was looking for a Caps lock major. upgrade. That might mean way more copy, way more images.

Way more storytelling. Yeah, exactly. Et cetera, et cetera. Yeah, I'll just say seriously, take your time, go nuts. Yeah. What are credits? Famous last words, Morgan.

Yeah, I know, right? It's like, oh, perfect. Okay. That's like a signal within the opening of headcores. Like, we finally got someone. Totally. It's a whale. All right.

So, Opus has finished. Yeah. I have no idea how many credits I would use, but probably. Actually, let me ask it. How many tokens in total did you use to put all of this together, including the four agents? And then we can. Oh, it's using token stands. That's so funny.

Okay. doesn't know. Let's see. Okay, here we go. Okay, it's estimating. It actually doesn't know, which is weird, because it should know. Although, I wonder if I can actually do slash cost. Oh, here we go. Yeah, okay. Oh, it doesn't, okay, no need to monitor cost. Okay, so they really don't want you to know. Okay, it's guessing 150,000, 250,000 tokens total. Yeah, that's probably right. Okay, sure.

Okay, so here's what it's done. So first off, one really interesting thing here is, you know, Codex created 10 tests, right? Opus created 96 tests. So definitely a lot, a lot more detail on the testing side. And it's called it forecast, whereas Codex called it Signal Market, so different names. A polymark competitor is built and verified. Here's what each team member delivered. So the architecture technical lead decided modular monolith

NextGh14 app router, central limit order book, database schema, Russell API. Okay, the prediction market domain expert, binary yes-no market where yes-no is always a dollar. Okay. Seated markets across crypto politics. Okay, the U.X design lead, dark mode trading platform. pages

it is a green for yes red for no okay testing QA lead did order book tests okay so here's the test are breaking out

order book test matching engine okay all right NPN Rondeb to start the app so let's go in here oh

interesting okay I don't want to say anything actually I've already given to it what's your initial take I mean my hello

Jack Dorsey, you know what I'm saying? This is what I expected it to look like when we pushed Codex. Yeah, me too. This looks really clean. What happens when you hover over? Oh, yeah, look at that. Yeah. It's got hover states.

Hover state. It's actually got it organized. Like, where it goes sports, you know, will the next few will have over 120 million viewers? Well, AI pass the turn. touring test by 2027. Will a move in 3? It's got stuff in there.

Yeah. Yeah. This doesn't even it doesn't feel like an MVP. Yeah. This is pretty wild actually. And it created some stuff, you know, that we never talked to it about, right?

Like a leader board? Um, which it's already populated with some initial stuff. Portfolio section. Yeah. Interesting. So let's see now. So, I mean, I'm more impressed.

But it was maybe worth the 150,000, 300,000 tokens. Feeling better about it. Will SpaceX land humans on Mars before 2030? Only 8% it thinks, huh? Oh, yeah, look at this actually. Whoa. This is insane, bro. That's clean.

That's clean. I wasn't expecting to click in and actually get a well-designed page like this. Huh. So if I were to do that, I have to sign in a tray. I don't know if I were going to be able to sign in because I haven't set anything up. Let me just check. Well, you can sign up just don't have an account.

Oh, yeah, sign up. I don't know if it gets it all connected, though. Let's see, though. All right, I'm snagging. You know what, actually, I'm going to take the username, Greg. Wow. All right. Let's see.

Okay. So, yeah, it's probably because I was going to say the database isn't wired up yet. So I'm not surprised that. I would actually have to do it. I wasn't expecting to do that. So, but I get it. I mean, it's clean. this is pretty neat.

Yep. Yeah. All right, let's see. So then Ken this, all right, we've given, I don't know about you, but this is the last chance I'm going to give Codex on the design side. It's out of, it's out of opportunities here. So, oh, here, it's funny. In the end, it kind of, it's acting a little bit like data from Star Trek. When you're question, what are credits?

In this context, credits usually means, that's quite it. That's good, though. Okay. so let's see. Let's take a look and see. All right, here we go. The new version of Signal Market. Boom. Oh.

Oh, okay. This is getting a little bit interesting. Let's see here. Read the manifesto. Yeah, I mean, I don't hate it. It's definitely better. Yeah. I mean, it's got a lot going on.

Yeah. But it's different. terminal. It's different than yeah, different than any sort of prediction market app I've seen from a Ux perspective. I just feel like this is just so clean though.

This is so good. It's so fast. Yeah. Yeah. I mean, I would say, you know, like I said, I'm not going to say which one is, it's not that opus is better than code extra vice versa, but I would say in this test opus one. Yeah. In this test, Opus one. Yeah.

That's just the truth. Yeah. Yeah, but we could give it another. I mean, you know, you never know. Like, I think that what's interesting about this is, I mean, Codex built it like, I don't know how much faster. We can look at the timing on this video, but like 20 times faster or something, right? Yeah, yeah.

Yeah. Well, anything else you wanted to cover, I don't think we'll have time to do another example, but anything else you want to cover between, you know, that you want to leave people with? Yeah. Let me see if there is anything else in here. I cover the adapted thinking, oh, I guess just on the orchestration, I would say, you know, this is a feature

I'm probably the most excited about with Opus. And clearly we saw in this example, it working really well. Just make sure to look at the documentation. It's all in the docs now. And it gives some examples as well because it has this idea of like compare with subagents of like context and communication

and coordination and kind of break. breaks this down. And then it has like a sample prompt, more of the details on the display mode. There's a lot of other stuff that I didn't go into there. That's probably what I would leave people with because I think a lot of people are going to want to dive in and use agents with Opus 4-6. And they've got pretty good details on all little tweaks that you can make with it. Amazing. Well, Morgan, I can't thank you enough for coming on.

I hope people love this episode. I love talking to you. Thank you for having me, Greg. It's a total honor. You're, yeah, it's just, I love how clearly you communicate and to technical people, but also non-technical people,

and you're criminally underfollowed, so I'm going to include links where you can find Morgan and follow him on X. He talks a lot about vibe coding over there. and Morgan, any anything else you want to places that you want to leave people to go and check you out?

Yeah, I mean, I'm the co-founder and CTO of Bold Metrics. I'll just give a little plug for us. We have an AI technology that's used by apparel brands and retailers. So if you're shopping online and want to find the right size, you see a find my size button, we a lot of times power that

and have really powerful machine learning models that update and adapt over time to help people find the right size and give lots of really interesting data to lots of amazing brands and retailes that you probably all know and love. And then me and my team, you know, we're using all of this tooling. Like I had a meeting with my team this morning about Opus 46 and Codex. And I've given everybody access to both of these. And I actually have multiple teams of mine that are trying current things we're working on and are actually testing with each to

see which performs better. So, you know, the one thing I encourage all engineering teams to do is, like, and engineering leaders to do is like, let your teams loose with this stuff. Let them try it. You know, some of this stuff is really cutting edge and really perform it and gives us the opportunity to do better, more creative work. Yeah, stop listening to us right now. Yeah, exactly. Stop listening. Go ahead and X out of this YouTube or Spotify link. But actually, give us a like, a comment and subscribe. Let us know if you like this episode, Morgan, thanks again for coming on the show. This was a lot of fun.

C-N-X-A. Craig, thank you so much. Total honor.