Has Codex CLI Finally Overtaken Claude Code?

Over the last week, I've been seeing a lot of talk  online about how OpenAI's competitor to Claude   Code, Codex CLI, has become really good, and I'll  be comparing both in this video. But to show you   what I mean, this person over here, they said they  needed to mock up some complex interactions in a   graph model, and they gave it to Codex CLI and hit  90% coverage 15 minutes later. Claude never got   past more than 10% on Opus. This person over here,  they said that GPT-5, high with Codex CLI, is   really good when it comes to Swift UI development.   It one-shots insanely complicated asks,   and there's been a couple of other people talking  about it as well on Twitter and also on Reddit if   you scroll down. There have been some big updates  to Codex CLI recently, one of them being that   ChatGPT Plus users now get 50% increased usage  limits, and some people have also been saying   that Codex CLI crawls their codebase to a degree  that they've never seen before from Claude Code.    And with a lot of the other discussion that's  happening, I think my expectations here are   pretty high. So I'll be using both Claude Code and  Codex CLI to implement three different features

across three different production code bases.   This video is not sponsored by anyone, I don't   accept sponsors because I don't want to bias my  videos in a particular direction, but it is made   possible by the people who do buy my AI tools  and products using the links and coupon codes   in the description down below. The first thing  we'll be doing is editing my AI News application,   Tensor AI. It's 100% free, and basically helps you  stay up to date with the latest AI News. Recently,   it became the number one app for the AI News  keyword on the US App Store, so there's a link   in the description to download it if you're  interested in it. But basically, we'll be   improving the recommended notifications to use  embeddings instead, because I added embeddings   for user interests and also the articles as  well. Next, we'll be editing my advanced,   highly customizable frontend for LLMs called  MindDeck to add MCP servers. There's a coupon code   down below if you're interested. And then finally,  we'll be doing something related to SwiftUI,   because this person said that a GPT-5 high with  Codex CLI is really good for SwiftUI. And this   is my speech-to-text application, HyperWhisper,  and I basically want to make some improvements   to design. There's a coupon code down below for  that as well. Both MindDeck and HyperWhisper are   single licenses, so you only have to purchase it  once, there's no recurring subscription. Firstly,

you want to make sure you have Codex CLI installed  by using npm install, and then make sure on the   latest version. And then you can just run Codex,  and then link it to your ChatGPT account, and then   you should see a welcome to Codex. Now, Codex CLI  does use an agents. md file, so I basically cloned   the claude. md file to be an agents. md file as  well. And I'll be running the same command in   two different folders that have the exact same  codebase. Anyways, I'll use my tool HyperWhisper   to transcribe my instructions by saying, hey,  basically I want you to add the ability where

when recommendations are made to a user based  on their interests, like the notifications are   sent to a user, then instead of using GPT-5 Nano,  it should use the article embeddings and the user   interest embeddings instead. I have a Supabase  function that already does this for the For You   feed. I want you to be able to make a similar  function that gets articles that are relevant   to a user after a particular created at, and then  uses those to recommend articles for the user.    Likewise, you should also update the roundup that  happens daily and weekly, and do roundups using a   brand new Supabase function that uses those  embeddings as well. Basically, I want you to   remove ChatGPT or GPT-5 Nano completely from the  codebase when it comes to making recommendations.    And now it's written everything out over here,  so I'll copy this over and paste into Codex CLI,   and now I'll press enter on both and we'll see  how they perform. Alright, so both of them are   done for this particular task. Claude is still  going around in circles, and that's because of   a bug in Supabase when it comes to migrations. So  let's quickly look for the code. Codex CLI's code   is on the left, and Claude Code is on the right  over here. Now, the solutions are pretty much   the same for the Supabase function. They both do  16 comparisons when it comes to distance between   the vectors and choose the greatest value. But  the Codex CLI one is slightly faster because it   handles the nulls, which will appear when they  don't have enough interesting embeddings. So I   actually prefer this one over here, and because  it also has a filter article IDs, which I didn't   realize that I wanted, but is useful. But despite  that, Codex CLI doesn't seem to have written the   second function that I hoped for, whereas Claude  Code did write a second function. So maybe the   implementation is slightly different here. And  now comparing the solutions and notifications   themselves, for Codex CLI, there does seem to be  a small type error over here that I can fix quite   easily by moving this. And the rest of it looks  pretty good. Codex CLI did implement a filter   articles ID in the Supabase function itself,  whereas Claude Code does the filtering after using   a set over here. So I think Codex CLI is slightly  more efficient, but they both get the same job   done at the end of the day. And as for creating  the roundups themselves, Codex CLI does use the   same one function that I wrote before. Whereas  Claude Code uses the other function that I wrote.    So it wrote two Supabase functions. And I actually  don't know which one I prefer over here. It's   quite nice that Codex CLI is efficient, and it is  not making multiple unnecessary functions that it   may not need to. Whereas it seems that Claude Code  did actually follow my instructions more clearly,   and did make a separate function which has  different logic, with a minimum and a maximum   match count. But I think for this particular  step, I do prefer Claude Code solution. It is   still nice to know that Codex CLI does seem to  be quite efficient, when it comes to thinking   about the best way of implementing something,  especially when it comes to performance. But yeah,   I think I'll actually be mixing the two solutions  together in this particular case. And now next up,

I'm going to be editing MindDeck over here.   allowing users to add their   own custom MCP servers, and use from the existing  MCP server templates. And basically, in MindDeck,   you can have many different chats in parallel,  to many different LLM providers. You can use any   provider that's available on Open Router, which  is literally hundreds of providers in parallel.    There are nested subchats as well. And everything  happens locally. So your API keys that you enter   are stored locally on your device. There is  no backend. And you can see this yourself by   going to Inspect, and then going to Network. And  then if you say something like, Hi to your chat,   then you can see your device makes requests  directly to the OpenAI endpoint. There is no   backend that's like funneling all your requests,  which means that's privacy first as well. And if   you want to, then you can import your chats from  ChatGPT. And I personally use it every single day,   across literally thousands of different chats  that I have. I'm always adding it to it,   and there's a coupon code linked down below  for the one-time purchase that it is. Anyways,   I'll be using HyperWhisper to describe the changes  that I want. Hey, so basically, I want you to add   a new section of the website, kind of like the  models section, which allows users to add their   own MCP servers. There should be some pre-built,  pre-added MCP servers, mainly the Tavily   one, whose documentation I'm about to give you.   But you should allow the user to add their own   MCP servers with their own headers and URLs and so  forth. And those MCP servers are injected into the   OpenAI responses API that is being used. I'm going  to give you the documentation for that as well,   of how that should be combined together. You  should also add a model for this as well,   to store the MCP servers that the user has added  in Dexie. Annoyingly enough, when you start a   new Codex session, then it switches back to GPT-5  Medium, so you have to switch back to GPT-5 High,   and I'll do that. And I'll also get it to search  online, because I'm quite interested in how   good at searching capabilities are. At least, we  have a recent paper from Salesforce AI Research,   called MCP Universe. GPT-5 is really good at web  searching, compared to Claude 4. 0 Sonnet. Claude   4 Opus is not on here, but I imagine it wouldn't  be much better than Claude 4. 0 Sonnet. If OpenAI   Codex did have a planning mode, then I would use  a planning mode right now, but since it doesn't,   I won't use a planning mode for either agent. One  thing I do want to mention when I'm watching these   models work, is that Codex has a really strange  way of searching the internet. Firstly, it uses   Google over here, and then later it switches  to DuckDuckGo for some reason. And then later,   it tries to do it via Python over here, and using  like requests. And yeah, it's really interesting   watching it go. Okay, so it seems that both of  them are done. Codex CLI did use more tokens as   well, it used 130,000 tokens, whereas Claude Code  used 72,000 tokens. So I expected more tokens,   it probably got more context from online or  something, so I'm hoping the solution will be   better, but we'll see right now. So I'll run both  of them. Here is the one from Codex CLI first,   and it has an MCP tab over here. It has a list of  MCP servers. So it has Tavily, twice. It has an   add MCP server over here. So let's actually check  if the MCP server works properly. So I'm going to   have to add my API key. And yeah, it's quite nice.   It added an icon URL over here as well, and it   lets me delete some of these default servers too.   But now the sidebar has disappeared, so that needs   to be added back. So now that's loaded, we can  have a search online and check. So we can say,   who won the first FIFA World Cup, search online.   And then using the responses API, it should search   online hopefully. And I should really add a  thinking thing. So did it actually get the   information online? We can check the response. And  yeah, you can see the MCP tool call was made over   here. I wish I had updated the UI as well, but I  should have told it in my instructions to update   the UI. But you can see the MCP tool call was  made using the API key that I provided to Tavily,   and then it searched with Tavily, and then got  the information like that. So overall, I am really   impressed with the solution that I gave. Although  I wish, just like in the model section, you have   the sidebar over here. It had the sidebar in the  MCP tab over here. And it like kind of copied the   design over, because it seems to have come up with  a slightly different like, padding, color scheme,   and design. But yeah, as for the underlying logic,  it works really well. And now this is the one that   Claude Code gave us, so we can go to MCP servers.   It did copy the design over, like the design is   very similar to the one in models over here. It  did add two Tavily servers as well. I'm not sure   why it added two. And there's a nice on button and  off button. I wish it allowed me to delete them,   because there are two of them. But I guess, can  delete them somewhere else maybe. And Claude Code   does seem to have a better distinction between the  built-in servers, and the custom servers as well,   compared to at least Codex CLI solution. And it  has a tool approval mode, which is really good,   because actually, Codex CLI did not implement  a tool approval mode. So I wonder if they added   some front-end logic for tool approval as well.   But anyway, make sure the server is active,   update server, and let's try a new chat, and  enter our OpenAI key. And now we can open it   up in the Inspect Element, go to network over  here, press new chat, and then enter in the same   message about the FIFA World Cup, and then see if  the responses API actually searches online. So it   has been provided with MCP tool over here. And  if I search the event stream, if I search MCP,   then it actually is using the MCP tool. So you can  see it's being returned in the event stream over   here. And yeah, overall, both of them implemented  a working solution, which is really good.    I do prefer the design of the one that Claude  Code gave when it comes to the MCP service page,   because it copied over the existing design, and  it considered things such as the tool approval,   which Codex CLI did not consider, even though  Codex CLI did use more tokens. But something that   I do like about Codex CLI is that it gave a nice  questions for you section over here. And it said,   like, here are some next steps that I can quickly  take. And this is a thing that I noticed with   GPT-5 in general, whenever it ends a message,  it usually has some kind of, like, follow-up   question or just something that I can do later  on. Okay, now we'll be moving on to HyperWhisper.

There is a coupon code down below for lifetime  access to this if you do want it. But basically,   when you see it right now, the design is kind  of, like, basic in the sense there is a fast   Fourier transform being applied to your voice  as you're speaking. And there's kind of, like,   a rainbow that's, like, swashing back and forth  as you're speaking. And I kind of want a design   that looks slightly more like this over here. So  there's kind of, like, waves, and these waves go   up and down, and there's, like, a neon, like,  glow around the waves or something like that.    Basically, I'm going to pass it the screenshot,  and then I'm going to see how well each of the   models do when it comes to updating the UI for  this. Okay, it turns out you can't just paste an   image into Codex CLI. You have to attach this,  like, image command over here and then give it   the file path. So I'm going to do that right now.   So now I'll just type out the prompt and say, hey,   so basically I want you to update the audio  visualizer for this application. Currently,   I'm using a fast Fourier transform and showing  the peaks and the waves as the user is speaking.    I want to update the design to use this kind  of design in this image instead. It should be   like neon waves with a glow surrounding them.   They should be able to go both up and down,   not just up. And, yeah, see what you come up with.   So now I'll copy over this prompt to here as well   and then press enter over  here. So this was slightly annoying because you   can't just drag and drop an image. It has to be  there with, like, the initial prompt. And it seems   that someone did open up a new pull request on  GitHub adding the feature to drag and drop images,   which hopefully goes through. But, yeah, I  think overall Codex CLI does not have as good   functionality built in by default. It's, like,  very limited in the amount of features that it   has right now. Okay, so both of them are done.   Claude Code took 54,000 tokens whereas Codex CLI   took 50,000 tokens. So I ran about the same. But  the tokens would be priced differently if you are   using the API instead of having a monthly  subscription. So that's worth taking into   consideration as well. Anyway, we can use Xcode  to then run both of them. We'll start off with the   Codex one. All right, so this is what the Codex  one looks like. So if I press this then allow to   my microphone. It looks pretty interesting. It  does have the neon effect and it does kind of,   like, move and react to what I'm saying. You  can see it, like, kind of dies down as I stop   speaking. But I don't like these small bumps that  are going across it. It's strange. But I think the   effect is very cool. I'm actually quite interested  or excited to see what Claude Code came up with.    And then press Control-Shift-I. And yeah, Claude  Code did not do as good of a job certainly when   it comes to SwiftUI. I think I'm going to hand  it to GPT-5 for this one. So I think the tweet   that I mentioned earlier where this person was  saying that GPT-5 High for SwiftUI was really   great was actually correct. So I think I'm going  to iterate on that design with GPT-5 High and make   it look better and then update the application.   And of course, if you want to use the application,   then you can download it using the coupon code  and the link in the description down below. Like   with the other application, your requests go  directly to OpenAI. So any post-processing that   happens on the audio or the speech-to-text of  the audio itself is done by OpenAI. But you can   also download any of these models offline. So if  you go to the model list over here, then you can   see these are all different whisper-sized models  and you can download bigger and smaller ones and   multilingual ones depending on your own use  cases. And then you can create modes out of   them. Like I have a Japanese mode over here. And  of course, you can switch between the modes and

so forth. I'll continue to add more features to  it. And I think I will continue to use GPT-5 for   this particular use case because it seems good  at SwiftUI. But I think for the other use cases,   when it comes to Next. js, especially like I did  with MindDeck, I will continue to use Claude   Code. It's quite interesting that each model  kind of has its own strengths and weaknesses.    But I think overall, I will continue to be using  Claude Code because just the CLI experience is   still better right now. They have more features.   Things just work out of the box. In the sense like   the images thing, I had to do this like crazy  workaround. Sometimes when it asks me, do you   want to approve permissions to use like curl or  something? And I select always, then it doesn't   stay as always. The model keeps changing back to  medium every time I start a new Codex session. So   I think Codex CLI does need a lot of improvements  still. I think it will actually be really good   in a couple months from now. But I think we're  starting to see where different coding models are   better suited towards slightly different things.   So some people will have multiple coding agents.    They might use GPT-5 when it comes to some  kind of design work and they might use Claude   Code for bigger tasks or they might want to like  optimize their costs in some way. I don't think   there will be an overall one winner, at least not  for the next year or so. I think each model will   still have its own strengths and weaknesses. But  for me, my overall conclusion is that when I'm   editing a web project on Next. js then I will be  using Claude Code. When I'm editing any Swift or   Swift UI applications then I will be using GPT-5  High. And I'll probably also be using GPT-5 High   to solve any bugs that Claude Code could not solve  and to make any performance improvements because   it seems to be able to write more efficient  code from the start. But yeah, hopefully that   video wasn't too long and it was interesting. If  you do like this kind of stuff then do subscribe   because it means that I should be making more of  this and it will let the algorithm know this is   a good video as well. And as I've said earlier,  I've never accepted a sponsor on this channel but   I still need to be able to pay like OpenAI and  Claude Code and everything else to make videos   for you guys. So if you do want to support the  channel then you can buy any of the AI tools   that I make and use myself in the description down  below. There should be some coupon codes as well   that give you a bit of a discount and you would  be supporting the channel as well. Thank you.