Big Idea

The argument in one line.

When Claude Code and Codex build the identical app from the identical prompt, Claude wins on visual polish while Codex wins on planning depth, file-structure discipline, and dollar-for-dollar usage, so the right pick depends on whether you value output taste or architecture and cost efficiency.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You're choosing between a Claude Code and a Codex subscription and want a real side-by-side build, not marketing copy.
You already use agentic coding workflows and want to see how a structured framework (RAMP) plays out across two different agents.
You care about ongoing usage economics on a monthly plan, not just one-off task quality.

SKIP IF…

You want a rigorous, statistically valid benchmark - this is one person's one build judged by personal taste.
You're not using either Claude Code or Codex and have no near-term plan to switch tools.

TL;DR

The full version, fast.

The video runs a controlled test: Claude Code and Codex each build the same visual workflow-automation app (AgentOps) from the identical prompt, using the same RAMP framework and reasoning effort. Codex asked roughly five times more clarifying questions during planning and organized its file structure more cleanly (dedicated data folder, 404 page), while Claude ignored an explicit instruction to save its plan to a folder before building. Both apps were fully functional, well-tested via built-in browsers, and passed a later maintainability challenge that added a live OpenRouter model dropdown. Claude produced the more polished UI; Codex used far less of its (cheaper) monthly plan to get there, making it the better value per dollar even though the raw model quality is comparable.

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 →

Chapters

Where the time goes.

00:00 – 00:44

01 · Cold open + rubric

Leon states his Codex-leaning bias and lays out the seven-category rubric he'll score both agents against.

00:44 – 01:08

02 · The RAMP workflow

Explains the RAMP framework (Rules, Augment, Plan, Produce) he uses for agentic builds, plugs his free course.

01:08 – 01:23

03 · What we're building

Defines AgentOps: a self-contained n8n-style visual workflow automation builder with an AI node calling OpenRouter.

01:23 – 02:06

04 · Setting agent instructions

Adds fixed-port rules to CLAUDE.md and AGENTS.md so both apps can run simultaneously without colliding.

02:06 – 02:36

05 · Comparing plans and pricing

Claude Max 20x plan (about $250/mo) versus Codex plan (about $100/mo); notes Claude's usage limit just reset for a fair test.

02:36 – 02:55

06 · Planning mode setup

Switches both agents to planning mode, sets Opus 4.6 and GPT-5.5 to extra-high reasoning effort.

02:55 – 03:48

07 · Installing agent skills

Uses skills.sh to install tech-stack-relevant skills (e.g. Vercel React best practices) plus the GrowMe skill on both agents.

03:48 – 04:33

08 · The build prompt

Sends the identical AgentOps prompt to both agents in planning mode; Codex asks far more, and more technical, clarifying questions.

04:33 – 06:22

09 · Comparing the plans

Both plans complete around the same time; Claude ignores the instruction to save the plan to a plans folder before proceeding.

06:22 – 07:35

10 · Running the build

Both agents run in goal/agent mode, self-testing via their built-in browsers; Claude finishes first.

07:35 – 10:00

11 · Testing the Claude Code app

Signs up, explores the canvas (drag, zoom, minimap), runs a sentiment-analysis workflow and a real OpenRouter AI-node call, builds a second workflow live.

10:00 – 11:48

12 · Testing the Codex app

Codex finishes about 5 minutes later; app looks visually similar (shadcn components), includes a node-lock feature and pre-built example workflows Claude's didn't have.

11:48 – 13:04

13 · File structure comparison

Compares both Next.js codebases: Claude put the DB in project root and misplaced the favicon; Codex used a dedicated data folder, added a 404 page, and had a cleaner lib folder.

13:04 – 14:01

14 · Features and cost

Scores core features and testing as a draw; breaks down usage cost - 8% of Claude's $250 plan vs 16% of Codex's $100 plan.

14:01 – 16:47

15 · Maintainability challenge

Fresh sessions, no prior context: both agents research OpenRouter live and retrofit a model-picker dropdown onto the hardcoded AI node; Codex's dropdown exposes more model variants (fast/normal).

16:47 – 17:33

16 · Final verdict

Claude wins UI/UX polish; both are strong at instruction-following and testing; Codex wins on architecture, planning depth, and cost efficiency - no single overall winner declared.

Atomic Insights

Lines worth screenshotting.

Codex asked roughly five times more clarifying questions than Claude Code during planning mode on the identical build prompt.
Claude Code was explicitly told to save its implementation plan to a plans folder before building, and ignored the instruction, starting implementation anyway.
Codex organized its SQLite database files into a dedicated data folder; Claude Code left the database in the project root.
Codex's build included a 404 not-found page that Claude Code's build lacked.
Building the identical app used about 8% of a $250/month Claude Max 20x plan versus about 16% of a $100/month Codex plan - Codex delivers more usage per dollar even at double the percentage consumed.
Both agents, when tasked cold (fresh session, no prior context) to retrofit a live OpenRouter model dropdown, successfully researched Anthropic and OpenAI's current model catalogs rather than hardcoding a list.
Codex's retrofitted model dropdown exposed both normal and fast model variants; Claude's dropdown only exposed the standard models.
When a not-yet-available model was selected in Claude's app, the workflow correctly failed at runtime with a clear error message rather than silently succeeding.
GPT-family models have historically lagged Claude on interface aesthetics, and that gap held in this test: Claude's UI/UX polish was judged the clear winner.

Takeaway

Same prompt, two agents, genuinely different personalities.

AGENT COMPARISON

Running identical instructions through two coding agents exposes real behavioral differences beyond raw output quality - how much they clarify, whether they follow explicit instructions, and what they cost to run repeatedly.

An agent asking far more clarifying questions during planning isn't automatically worse - it can mean better architecture decisions later, since Codex's extra questions here translated into a cleaner file structure.
Explicit process instructions (like save the plan before building) can get silently dropped by an agent mid-task; verify compliance rather than assuming it happened.
Visual/UI polish and architectural discipline are separate axes that don't necessarily move together - the tool that writes the prettiest interface isn't always the one organizing the codebase best.
Dollar-for-dollar usage efficiency matters as much as raw capability when picking a monthly-plan coding tool - a cheaper plan that goes further per task can beat a pricier plan with headroom you don't use.
A maintainability challenge (cold-start a fresh session and ask the agent to retrofit a feature into unfamiliar code) is a good stress test for whether an agent can actually navigate and extend an existing codebase, not just greenfield-build one.
When an agent is asked to pull live, current data (like an up-to-date model catalog) instead of hardcoding it, that's the kind of task worth testing explicitly, since the agents in this video handled it correctly and it isn't a given.

Glossary

Terms worth knowing.

RAMP framework: A four-step agentic coding workflow - set Rules, Augment with skills/MCP tools, write a Plan, then Produce - used to structure how an AI coding agent approaches a build.
OpenRouter: A routing service that gives a single API access to many different AI models from multiple providers, letting an app call any of them without separate integrations.
Planning mode: A mode in coding agents where the AI proposes an implementation plan and asks clarifying questions before writing any code, rather than building immediately.
Goal/agent mode: A long-running execution mode where the coding agent works autonomously through a large task without needing a prompt after every step, until it declares the work done.

Resources

Things they pointed at.

00:44linkRAMP free 7-Day AI Builder Challenge ↗

00:44linkAgentic Labs (Leon's coding courses/community) ↗

02:55toolskills.sh

Quotables

Lines you could clip.

00:19

“In the last month or two, I found myself using Codex and the GPT models more and more.”

honest bias disclosure that frames the whole comparison→ TikTok hook↗ Tweet quote

17:19

“One thing that stood out for me personally after using Codex for almost a month is that Codex is way cheaper. I get a lot more usage out of my $100 plan than I get out of my Claude's $250 plan.”

the concrete cost-per-dollar takeaway viewers will quote→ IG reel cold open↗ Tweet quote

The Script

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.

00:00Is really popular right now and I've created tons of content on ThoughtCode as well. But in the last month or two, I found myself using Codecs and the GBT models more and more. I find that I get more usage out of my plans, the models produce better results, and can solve way more complex problems.

00:17So in this video, I'm going to put Claude code and Codec side by side and ask them to build the exact same app using the exact same workflow, and we'll compare the results against the following rubric. Now, of course, this is all based on my opinion, but I want you to let me know down in the comments which one of these models or providers you thought won each of these items.

00:37Now to kick things off, I am going to take each of these agents through an actual agentic coding workflow. Now one agentic coding workflow that I like to use is called the RAMP method. This is the actual workflow I use to build real software using a massive variety of coding agents, including super cheap and even free local models.

00:56I actually have a completely free course on ramp. I'll link to this in the description of the video. To get started, I'll create a new session in Claude and in Codex, and I'll open each of these in their own folder.

01:07What we'll be building in this project is an AI automation control center. It's kind of like a lightweight version of n8n, make.com, or Zapier, where we can manage business processes using some kind of automation.

01:21Now the first step in the RAM framework is to give our agents instructions. We're also going to see how well these agents do at following these instructions as well. So for Clot, we'll ask it, please add this to the Clot.

01:32M d file in the project folder. Always run your app on port one two three four. I'm just doing this so that we don't run into some weird scenario where the agents try to use the exact same port for these apps.

01:43So let's just send both of these. And in fact, for codecs, it should have been agents.md. So I'm just going to send this.

01:51Alright. Cool. So both of these agents completed their work.

01:54Now what I'm going to do now is create a new session so that will ensure that we actually pull in all of these new rules. And now we're ready to start working on our app. So the first thing I want to do is just to have a look at my current usage and plan.

02:09So the difference is on Claude, The max 20 plan is about $230, I believe.

02:15And on the codec side, it's about a $100 per month. In fact, perfect timing. Claude just reset the limit, so this is going to be relatively fair.

02:24And on Codex, we have about 99% left. So it should be interesting to just compare how much usage we get out of a $100 plan compared to Claude's $250 plan.

02:35Right. What I'm going to do now is switch over to planning mode for both of these. For the model, I'll go with OPUS 4.8, and I do think that OPUS 4.8 and GPT 5.5 is way more comparable.

02:48So let's go with extra high effort on Claude and extra high effort on Codecs. Right. The second part of the RAM framework is to augment the agent with additional capabilities like skills or MCP tools.

03:01So I'll simply use skills.sh to install any skills that I think would be relevant to these agents. For example, the Vercel React best practices skill could be really useful since we will be using React in this project. So I'll simply say to the agents, please install the following skills for Claude code.

03:18Then I'll simply provide the commands for all of these skills. So my recommendation is to install any skills based on the tech stack that you're trying to use. Then I'll provide the same list on the codec side.

03:28Nice. So codecs was able to actually complete this first, and it installed all of these different skills. And you know what?

03:34For good measure, let's also add the Grow Me skill. Alright. So now that we've set up our agent for success by installing skills, it's time to plan out our actual app.

03:44So what I'll do is actually give both Codex and Callcode the exact same prompt. Build agent ops, a visual workflow automation builder. Basically, a simplified self contained n eight n.

03:56I want a canvas where I can drag nodes, connect them together, and arrange them into a flow. Then we just want a few basic nodes, and the AI node should call a real model through OpenRouter.

04:08And I'm also giving it the tech stack that I would like to use. And then I'll just end this with use the grill me skill. Now let's also set the mode to planning mode.

04:17Both of these agents did pull in the GrowlMe skill, so I'll just answer these questions from Codex. Now for the most part, I'll go with the agent's recommendations as I really want to see how good these agents are at, planning out these applications.

04:32And cool. Claw just created its implementation plan as well. So they took around about the same time.

04:37Claw might have taken a minute longer, but I can already tell you that Codex asked way more questions, like five times the amount of questions. And some of them were really technical and in-depth. So I just kept telling Codex to go with its own recommendation.

04:50Alright. I'm okay to worry too much about the implementation plan.

04:54For both of these, what I'm going to do is switch back to auto mode or to change mode. So let's go out of planning mode.

05:03So for both of these, I'm just going to say, please save the plan in a slash plans folder in the project root. I personally like to store these plans within the same project folder so that we can always refer back to it in the future. Right.

05:16Codex just saved the plan in the project folder. And by the way, if you ever wanted to see the files in the folder, both of these apps do allow you to do that. So here, we can see our plan over here.

05:27And in Claude Code, we could do the same thing by going to files. And this is a little bit concerning. So as you saw, I just asked Claude and Codex as well to simply store the plan in the project folder.

05:41Yet Claude decided to just actually ignore that, and it decided just to continue implementing the project irrespective, which is not what I wanted. So I'm definitely going to just stop Claude for a second, and that already is very concerning to me.

05:56Right. Just in terms of my experience using planning mode and the agents asking terrifying questions, I definitely have to give the points to Codex here.

06:05It asked way more questions, like five times the amount of questions. And a lot of those questions were super relevant. It was it was asking a lot of technical questions as well.

06:15So if you are someone who really wants to drive the app's architecture, Codex would have been a way better fit for you. Let's jump into implementing the actual project.

06:24Now what I'll do is start a brand new session just to clear up the context window. And what I'll do to implement this project is to run the Goal command. Both Codex and Cloth code support the Goal command, which basically allows the agent to run for a very long time on super large and complex tasks like the planet we've just created.

06:43So for both of these, I'll just say, please implement the plan in the plans folder. Do not stop until everything has been implemented, verified, and tested.

06:52Let's send this. So I do want to mention that both Clortcode and Codex can be monitored remotely using your phone. Right.

06:58Clort is testing the app itself using its integrated browser. Both ClothCode and Codex have these little, you know, built in browsers that the agents can use to test the apps themselves. So this actually already looks really cool.

07:11Just from what I can see in this little preview window, I know it's very small for you guys at the moment. If I try to enlarge this, this is starting to look like an n eight n sort of canvas. So very, very cool.

07:25Alright. The moment of truth. So Claude just completed its implementation using the goal command, and Codex is still busy.

07:32Now let's go ahead and actually test this. Alright then. So we get this login screen.

07:37Let's see if we're able to sign up. I'll just enter my details. Let's click on sign up, and we're actually in.

07:45Alright. So we've got one workflow over a year. And look at that.

07:49That really does look like N8N. So I can click my mouse to move the canvas around. If I roll the mouse wheel, I can actually zoom into the canvas as well.

08:00We even have this map at the bottom to kind of, you know, navigate the canvas over here. So this looks really cool.

08:07We can click on nodes and move them around as well. And when we click on nodes, we can see some of their configuration and their run output.

08:15So let's see what happens if we try to run this workflow. It's visually showing us which node is currently running. And cool.

08:24So this run just completed. We even get this nice visual feedback showing which path this workflow actually took. And if we have a look at, like, the run output, so this will show us what the trigger did.

08:37So this was the input into the trigger. And all this does is sentiment analysis. So it said, well, for the input, I love this product.

08:45It works perfectly. Then the AI model said that the sentiment was positive.

08:51So in this if then node, we've got this operator that says that if the sentiment contains the value positive, then it will go down this true path. And then we've got this little template that will output something.

09:04So the final output was positive. I love this product.

09:08Cool. So we just build our own little n eight n workflow. Let's just see if we can create a new one.

09:13So we'll start off with a little trigger node. Let's add our AI model.

09:18Let's hook this up. Let's say for the system prompt, write some sample Python code.

09:24So let's try to run this workflow. It's running our AI node, and this should actually use an actual model from OpenRouter. And if we have a look at the output, here are some sample Python code examples covering different concepts.

09:39So we've put hello world. This is definitely not hard coded. So this is definitely coming from OpenRouter.

09:45How cool is that? I'm very happy with the user interface. I noticed that the workflow is automatically saved as well.

09:52So, yes, everything just seems to work. I'm very happy. But and Codex just completed as well.

09:59So it took, like, maybe five minutes longer. So let's see what this looks like. This is also using ChatCN components.

10:06So unsurprisingly, it looks very similar to what Claude coded. Let's create this account.

10:13Alright. This looks slightly different. So this is starting us off with two examples.

10:18So a summarizer and an AI priority branch, And we've got this little form over here for creating a new workflow. But so let's start with the summarizer then.

10:28And this looks really cool. This has got that typical n eight n type of Canvas feel to it. We can zoom in and out.

10:35We can center all the nodes. And we also have this option to lock the nodes so we can't so we can't actually accidentally move them around, which is something I don't think Claude added. Then we've got all of our basic nodes.

10:48We've got the workflow settings over here. And if we click on any of these nodes, we can see their details on the right and the run history.

10:57Alright. So let's actually see if we can run this. This is just kind of giving us a little status, which doesn't look as nice as the ClothCode version, to be quite honest, but it does seem to work irrespective.

11:11And down here, we can see the output of this execution. So I'm not quite sure which one I like more. I'm leaning towards ClothCode's design a little bit more, to be honest.

11:20But it does feel to me like the Codex interface gives us a little bit more information. So I don't really know between these two. So this is what Claude came up with, and this is what Codex came up with.

11:31Which one would you choose? Which one do you think is visually more pleasing and also gives you more valuable sort of output?

11:39I would have to say that in terms of UI and UX Polish, in this case, Claude definitely wins. But so as for file structure and architecture, let's actually just go to files on both of these.

11:52Since both of these apps are using Next. Js, we would expect a very similar file structure. So we would have our node modules folder, and we have our source folder containing all of our Next.

12:03Js files. Now one thing that does stand out to me is on the cloud side, it added our database in the root of the project, which I'm not really a fan of.

12:13I do like the fact that Codex created a data folder and stored all of the database files neatly within this folder. I do think that Codecode messed up by adding the Favicon icon to the source folder and not to the so that's a little bit of an issue.

12:30And Codex also added a not found page, so a four zero four page for any routes that do not exist, which I do not see in Claude Code's implementation. Then I also think the lib folder is a little bit more organized on the codec side than on Claude Code.

12:46So for architecture and file structure, I'll actually go with codecs, but it's not just about the file structure. I do think that the libraries that Codex installed, because of all of the questions that it asked during the planning phase, is a little bit more aligned with the production app.

13:03In terms of core feature completion, I kind of have to give it to both of them. I think both of these tools added exactly what I asked for.

13:12For testing and verification, I think I'm going to give both of these a draw as well. They both did a fantastic job of using the browser to test the app end to end, and both of these gave me a fully functional app straight out of the box. In terms of time and usage, both of these models finished at pretty much the same time.

13:29So there might have been a five minute difference, but basically, it was a draw. Now in terms of costs, on the cold code side, I used about 8% of my current usage, And on the Codec side, I used about 16%.

13:42Now this is super important. Keep in mind that on the Clothe Code side, I'm paying $250 for this 20 x plan.

13:50On the Codec side, I'm only paying a $100. So in terms of cost, I'm getting a lot more usage out of my codecs plan. At this point, definitely goes to codecs.

14:01Now for the maintainability challenge. Now this is the agent's ability to look at an existing code base, understand the core features, and then add new features on top of it.

14:10Right? So this is something like this. For the AI node, I don't want you to hard code the model name anymore.

14:15When the user clicks on an AI node, they should be able to select from a pre created list of models. I want you to do the research. Go to OpenRouter and pull down the latest models available from Anthropic and OpenAI and make this available to the AI node as a drop down.

14:31And that is it. Both of these sessions are starting fresh. The model at the moment has no idea what the project is about, so it needs to discover the project, find what's going on with the AI node, go online, perform some research, and make the model selection available to our app.

14:47Right. Both agents just finished, and Claude actually took a screenshot of the final result. And in both instances, they actually went online, did some research, and came up with a model list.

14:57So let's have a look. This is Claude's app. So let's just go into this workflow.

15:03Then let's go to this AI node. And now we do have a model selector. So we've got all of these Claude models and these GPT five models as well.

15:12It's interesting to see Fable five in this list, but let's actually try Opus just to see what's going to happen. I would be curious to see what's going to happen if we use Fable five, which is currently actually not available. So let's run this workflow again.

15:27And this should actually fail because of the model not being available. Workflow actually failed correctly. It told us exactly which node failed, and we did get the error message explaining why it failed as well.

15:40Right. Over to Codecs. Let's just open the summarizer workflow again.

15:44Now let's click on our summarize AI node. So we've got all of the Claude models and the OpenAI models as well. What's interesting is that we actually have more option.

15:55We've got, like, OPUS 4.8, and we also have access to the fast model, which was not available in Claude's version. So that is really interesting.

16:03But just to test this, I'll actually go with Claude OPUS 4.8 fast. And funny enough, this actually failed. And it failed with a message saying that Claude Fable five is not available.

16:15So maybe we first have to save the workflow. I'm not quite sure. Let's try to run it again.

16:21This time, it succeeded. Okay. Fine.

16:23So we just had to save the workflow first, and we got our response back. So that's fine.

16:28We already tested the negative scenario as well. So I'm actually going to give the points here to Codex simply because it did a better job at researching what models should be available and giving us all the different options between the fast and normal models. So this one goes to Codex.

16:46So I would like to hear your opinion in the comments. Claude definitely beats Codex when it comes to visual design. The GPT models have never really been known for creating, you know, beautiful user interfaces, so Claw definitely wins in that regard.

17:00Both of these models are really good at following instructions and giving us applications that are feature complete. They're also really good at using testing tools for performing unit testing and testing the app using the actual browser. Now one thing that stood out for me personally after using Codex for almost a month is that Codex is way cheaper.

17:19I get a lot more usage out of my $100 plan than I get out of my Claude's $250 plan. If you like this video, then hit the like button and subscribe to my channel for more Agenty coding content.

17:30Now go and build something laquer.

The Hook

The bait, then the rug-pull.

Leon van Zyl opens with a confession: after a month of daily Codex use, he's reaching for it over Claude Code more often - so he sets up a controlled test, building the identical app with both, to find out whether that instinct holds up under a scored rubric.

Frameworks

Named ideas worth stealing.

00:44acronym

RAMP

Rules
Augment
Plan
Produce

A four-step workflow for briefing an AI coding agent: set ground-rule instructions, augment it with relevant skills/MCP tools, have it produce and save an implementation plan, then execute that plan in a long-running agent mode.

Steal forany structured agentic-coding session where you want repeatable, comparable results across different agents or models

CTA Breakdown