Claude Code VS Codex VS OpenCode

Key Takeaways

Today, we're going to do a comparison of the three major coding agents out there, which are Claude Code, OpenAI's Codex, and Open Code. The comparison will be a combination of my personal preference or vibe, you could say, hard facts like features missing in one tool versus another one, and also some general public sentiment and opinions. We're going to analyze multiple aspects like ease of use, default settings, model variety, model exclusivity, tooling, efficiency and speed, and also stuff like limits and transparency. Now, if you like this video, let me know by hitting a like button and subscribing, but now let us get right into it. >> [music] >> All right, so let us get started with the comparison. The first thing I want to tackle is the ease of use and also the default settings. So, which of these tools comes with the best default settings? And of course, this is already a very subjective thing. There is no correct answer to this. It's a matter of preference. So, what I'm going to do here is I'm going to go to my tutorial directory, and I'm going to be playing around with these three tools and showing the different uh settings. And in my case, I have all of them installed. I have Claude, I have Codex, and I have Open Code on my system. And I deliberately logged out of Claude Code to show you the onboarding process because I think it's very convenient. So, when you're logged into Claude in general in your browser, you can run Claude Code on your system, and it's going to ask you right away, what style do you want? Do you want dark mode, light mode, whatever? I can choose that. Now, it opens up my browser to sign in. This is super simple, just authorize. And then it asks me for some default settings. First of all, the warning here, and it also asked me if I want to use the recommended settings, which means I can use stuff like shift enter for new lines, which is also possible in the other tools. But basically, with the recommended default settings, I'm actually quite happy when it comes to Claude Code. Again, this is a matter of personal preference, but for me personally, Claude Code wins this because both Open Code and Codex come with a setting that allows you to do everything. So, if you don't change the config file, if you don't set up different settings than the default, open code and codex basically just go and execute whatever they want to execute. I'm going to show you that. I have codex with the default settings here. If I go into my tutorial directory, open up codex, and tell it to do stuff, to code something like code a minimal CLI calculator in Python. Uh it's not going to ask me if I approve what it's going to do. It's just going to go and write the code. And it's also going to run commands, and it does basically whatever it wants without me having to approve anything. As you can see, it just creates files. It doesn't ask me at all. Now, this is not a huge issue because, as I said, you can just change the config file and it's going to ask you for permission for everything. But, I personally like Claude Code's default settings more because I don't like to do full vibe coding. I don't like to just put in a prompt and then have it do everything. I want to approve every step of the way. I like to leverage the power of AI coding because I don't want to type everything myself, especially when it comes to front end stuff, but I don't like to have uh just a model do everything it wants to. So, if I do the same thing in Claude Code, build a simple to-do CLI app in Python, you will see that already when it tries to do the very first thing, it's going to ask me for permission. I mean, for LS commands, not because that's a just a read-only command. That's not a big deal. But, whenever I want to do anything here, it's going to ask me for permission. Now, at the same time, I can open up open code here. Open code has the same problem, I would say, as codex, but in my case, I just easily set this up to be different. Here, you can see in Claude Code, it asks me now, "Do you want to create a to-do file?" And also, if I want to, I can do shift tab and allow it to do everything, which sometimes can be useful. But, other than that, I have to constantly approve everything it does. Now, for open code, as I said, the default settings are the same as for CodeX. So, the default setting is also that it does whatever it wants, but you can easily change that. So, that's not a deal breaker. If I go here to dot config open code and I open my open code JSON, this is all it takes to uh to to make it ask permission for bash commands and edit. So, that's not a deal breaker. But, if you evaluate the default settings, in my opinion, Cloud Code has the best default settings. I just most comfortable with this one. Uh when it comes to open code versus CodeX, I would say that the overall experience of open code, how you select a model, because what I have here is I also have GPT-5.5 connected from my OpenAI subscription. I would say that this is a very convenient onboarding process, but I wouldn't necessarily say it's much better than CodeX versus like that open code is much better than CodeX, but in my opinion, Cloud Code wins the default settings. Then, maybe a related point, the design of the terminal user interface. I already made a video on this one. For me, there's a very clear winner. If we're just looking at design, we're not looking at functionality or powerful models or anything like that or tooling, just how it looks, just what the terminal interface looks like and how convenient it is to do stuff like change the models um and take a look at different settings that we have. For me, 100% open code wins the race. It just has the most beautiful UI. You have a really convenient way to look at stuff. You have a very beautiful terminal interface. Also, when you start doing stuff like code a simple snake game in Pygame, then you can see this whole interface here, the animations, everything just looks way more polished. So, my ranking when it comes to the mere design, to the TUI design, would be open code number one, Cloud Code number two, and CodeX number number three. It just feels the least polished to me personally. Yeah, so here you can also see again how open code looks. I think the code preview is very convenient, too. I like this one the most. But, yeah, all of this is very subjective and design-oriented. Let's get to something that's actually relevant in terms of functionality and probably also the biggest uh the biggest selling point of the proprietary ones, which is the model exclusivity. So, in general, it is possible to use Anthropic models in OpenCode. If you type {slash} models and you go and connect any provider, you can see that we can choose Anthropic here. We can provide an API key and also use Opus and everything. However, you cannot use your subscription. This is not true for OpenAI. You can use the ChatGPT Plus or Pro subscription here. That's nice, but overall, if you want to use the Claude Code subscription, you have to use Claude Code. They don't allow you to use it with any other tool. So, that is on the one hand, I don't like this about Anthropic. So, for me, that's a big minus point for Anthropic, but also it makes their tool more interesting because that's the only tool that you can really use to use your subscription unless you want to invest a lot of money into uh API tokens with Opus, which is, yeah, going to make you poor very quickly. But, in general, the exclusivity is probably the biggest selling point of Claude Code and also to some degree of CodeX because even though you can use the models in OpenCode, probably the tooling is optimized in Claude Code for Opus and for the Claude models and in CodeX for the OpenAI models. That is just the thing that's always going to be the case. The people that create the models and create the toolings for the agents that use the models, they're going to probably do the best job when it comes to building the environment specifically for these models. So, I would say that this point always depends on the current state of the model market, if you want to call it that. Right now, in my opinion, Opus 47 is the best model out there. So, if you want to have the best experience, the most powerful agent, in my opinion, you would use Claude Code just because the model is the most powerful one. If you're watching this a month from now, 2 months from now, maybe the GPT model that is currently out there is the most powerful one. Maybe a completely different one is the most powerful one. In this case, that IDE or that agent, I should say, is probably the most powerful one and the one that wins this race. In my opinion, right now, this point goes to Claude Code just because it is the most exclusive, even though I don't like this as a philosophy, I don't like their business model when it comes to that. But just considering what I want to be using as the developer, as the engineer, in my opinion, Claude Code wins the race here because it has the most powerful model and the easiest way to use it and the best way to use it is by using Claude Code. But then we also have the opposite point, which is model variety. And here the winner is very clearly Open Code because if I go to Code X and I say model, you can see I can only choose my GPT models here. If I go to Claude Code and say model, I can only use the Claude models. If I go to Open Code and I type {slash} models, you can see that I can use a bunch of free models, too. So, these are free models that I can use with Open Code by default. I have OpenAI models here and with control A, I can connect a bunch of different providers. I can go with Google, I can go with ZAI, I can go with all the different providers here. I can connect so many different. I also think Open Router, uh and that is definitely the tool that you want to use if you want to have the variety of models available. You can also use local models here. So, Open Code is by far the best choice if you don't want to rely on a single type of model. So, if you don't want to use the Claude models, if you don't want to use the OpenAI models exclusively, you want to be using everything, maybe you focus on open source models, maybe you have some simple tasks that you want to use MiniMax for or something like this, that is where Open Code excels. So, at this point I want to briefly talk about the sponsor of this video today called Q and Q is a very useful tool when you're working with AIs and want to de-risk the process of trusting them. So, let me show you a very simple example. What Q is, it's a browser extension for Chrome. So, you install it, you log in, you don't have to pay anything to try it. So, you can just download it, add it, and then you can talk to your favorite AI. In my case here, I can talk to ChatGPT and ask it a question. Now, for demo purposes here, I'm going to use something where ChatGPT is going to give me the wrong answer. So, my question is going to be, I want to wash my car. The car wash is 50 m from my home. Should I walk or drive? This is a famous question that confuses AIs because they say walking is the better choice, even though of course, in order to wash your car, it needs to be at the car wash. So, you need to drive even if it's just 2 m. So, what you do now is you send a prompt and here it tells me, "Walk there, then drive the car through the wash." And here you can see Q and A is comparing the model answers to give me a better answer. And in this case, it found one. It found an enhanced answer. I can view it and it will give me the summary of the different models. And here it says correctly to decide whether to walk or drive, you need to clarify the goal. Since you cannot wash the car if the car is not at the facility, the answer depends. So, that is a balanced answer compared to this one here. You can see here that Gemini correctly identifies at the bottom here that you need to drive if the car itself is what needs the service. Obviously, otherwise it doesn't work. Claude and Grok, for example, tell you to walk because you should burn calories and not fuel. And ChatGPT does the same. So, you can see the problem when you're interacting with ChatGPT or Claude or whatever, it can hallucinate. It can give you a wrong answer. This is now an extreme example, but you can also ask for side effect of drugs. You can ask for comparison what is the best tooling, what AI agent has the best tooling, and they might come up with wrong answers, whereas other models can have a more objective answer, a better answer, and Q and A will automatically recognize that. And if it doesn't, if it doesn't trigger something, you can still force a comparison manually. And the best thing is you don't even have to use a specific tool for that. You don't have to use Q and A. You can just use ChatGPT, Claude, and it will automatically appear here as a button for comparisons, and it will automatically notify you when the answer is not trustworthy. So, if you want to give it a try, you will find a link to it in a description down below. You can just add it as a Chrome extension and it will show up where you need it. Also, with the code Neural 90, you will get 90 days of QE for free and you don't even need to provide a credit card. As I said, link is down below. So, next up, let us move on to tooling. And by tooling, I mean the features that the respective agent offers. So, some agents will have features that the other agents are missing and for me, personally, this is the biggest deal breaker when it comes to Codex. I know most of you guys won't care about this. Codex has a huge fan base. A lot of people think that Codex is better than Claude Code and there might be aspects where I agree, but overall, I will never use Codex until they fix this because for me, this is a deal breaker. And what I'm talking about is the undo command, the undo tool or the rewind command, call it whatever you want. You have a session here. For example, we did a to-do application. What I can do with Claude Code is I can say {slash} undo or {slash} rewind. And what's going to happen is I can go back in time. I can restore code and conversation or just code or just conversation and there you go, the file is gone. That is a feature that I also have in Open Code. So, I can say undo here and I can go back and the file was never created. We can also go here to the tutorial directory and see that only the calculator is left from Codex. But in Codex, I don't have that. I don't have a rewind. I don't have an undo. There is no such functionality here. And we can even see that there's a GitHub issue here. Please make undo back and a lot of people want this, but for some reason, they don't implement this. I cannot imagine that this is so complicated, but this is for me the number one reason I will not use Codex because I don't want to be using Git every time I make a change and sometimes I will have multiple prompts before I get the final result that I want to push on GitHub, like that I want to actually commit and and save. And in between, maybe I have some progress and then it makes a mistake and I cannot reverse that mistake. And if I just tell it to reverse it, it's going to make mistakes. So, I'm not a big fan of that. For me, the undo feature is very important. So, there Codex absolutely loses me and this is the number one reason I will not use it. But also a small extra point here for open code. It's not really a huge deal because it's not the feature that I would like to see, but we do have in open code a redo command and we do not have this in Claude code and definitely not in Codex. So, we cannot undo the undo in Claude code, but we can also not entirely do it in open code because if we take a look at the tutorial directory here, if I list the files, you can see that the snake game did not return. The chat history was redone, so we have the same context here, but the file is not there. So, that's not a proper redo like a Git a Git pull or a Git checkout to a different branch. This is actually just a redo of the conversation history. Nevertheless, for me, a big minus for Codex because we don't have an undo feature. Now, a pro that I see with Claude code here is that there is a voice input mode that we do not have with Codex. We cannot do {slash} voice or something like this. We can also not do that in open code. There's no voice input. There is, I think, a plugin, but out of the box it doesn't come with that. Uh for Claude code, I can say {slash} voice and this now disabled it because I had it enabled. With the voice mode enabled, I can now hold space to record and I can give it instructions. Please code a snake game in Python using Pygame. So, I just talked into the microphone, hold held space, and now I can just commit that. I can also type in addition to that. And you can see that this is going to now execute it. This is a very convenient thing. I haven't used it that much, but I think I'm going to start using it because that's a very convenient thing if you have a sort of vibe coding session, especially when you're doing something not super important like front-end design or something. That is a very nice feature to have. Uh Maybe there's also a way to connect this via the phone, I don't know, but Open Code misses this feature. I think there's a plugin as I said, CodeX definitely misses this feature, so that's a big point for Cloud Code. Another feature that we only have in Cloud Code is a marketplace, so you can say {slash} plugin or {slash} marketplace is also possibility here, and you can discover plugins, you can install plugins, you can look at marketplaces, you can also add custom marketplaces. And this is just a very convenient way to manage the skills and the plugins and the MCP servers and everything from within Cloud Code. Not a huge feature, not something that I need to have, but it's a nice extra that we have in Cloud Code. And in general, there's a pattern, I'm not going to go through all of it because I made a Cloud Code crash course where I showed you all the different features, but to me it feels like Cloud Code has the most stuff to offer here, the most interesting features. We also have the work tree feature, which is basically you can say cloud and then {dash} {dash} work tree, which is going to work with Git work trees. Also not a huge deal because I can do the same thing with CodeX and Open Code by just using Git work trees manually, but in terms of like tooling and usability, I think Cloud Code has a lot of small things that make it easier to work with. I do also think though that Open Code has a lot of features. We do have a lot of commands here, we can fork sessions, session management in general I would say is very powerful in Open Code. We do have a lot of features here, we do have a lot of community response, so the community consistently develops this. I think Open Code is also very strong when it comes to tooling. CodeX is very minimal. I think CodeX doesn't have a lot of things that we can do. I don't like the tooling in CodeX, you can see I'm already a little bit biased against CodeX, don't hate me for that, I know it's right now probably the most popular one, but for me Cloud Code and Open Code probably have the best tooling. And yeah, CodeX is the weakest in this category. Now for the next point, I would like to combine a couple of aspects and analyze them together. I would like to take a look at efficiency, speed, limits, and transparency. These are of course different points, but they kind of belong together because we have a trade-off when we use Codex versus Claude code versus open code. And this is based partly on my own experience, partly on the consensus opinion of people on Reddit, Hacker News, and so on. And the general sentiment seems to be that Codex takes a lot of time to respond, especially when you use high reasoning levels, but usually gives you deeper, more correct answers in a one-shot way because it actually takes a look at your code base. It actually tries to understand before it does stuff. And Claude code usually gives you a quicker answer. It's better for prototyping and getting results quickly, but it doesn't think as deeply. And this is not just because of the effort level, but this is this seems to be something that is specific to Codex that it thinks longer and deeper before actually taking actions. Now, of course, it's hard to quantify this stuff because I think there is some truth to it that Codex thinks longer and deeper and actually doesn't just jump into coding. Nevertheless, I personally feel like I get better one-shot results with Claude code. It's it's just my experience. Whenever I want something to be done and I want to trust the agent as much as possible, I'm going to use Opus 47 and not Codex. But I do get the point I do see the point also myself that with Codex you really get this long thinking time and it actually looks at all the components and things. I just personally don't like the GPT models when it comes to coding. They do a lot of stuff that I find odd. They make design choices and they make code style choices that I really don't like. But in general, you could say that Codex goes deeper, takes longer, Claude code does it quicker and in a more prototyping way. And I think that open code with the GPT subscription basically does the same as Codex, but of course, the harness, the tooling is going to be more optimized in Codex. So, if you want to get the best experience for in-depth coding, you probably want to go with Codex for a one-shot prompt, but Open Code will be quite similar when you use this model. Now, when it comes to limits and transparency, I think everyone will agree that the tokens are very quickly gone when you use Cloud Code. It is very transparent. If we go and say {slash} usage, you can see how much of the current session or the current week you have wasted or already used, and that's nice because you can always see how much more you can go or how much further you can go before you have to use extra or something like this. But in general, it's very easy to run through your tokens with Cloud Code for very little work, like two prompts that are a little bit more detailed and the tokens for this session are gone. You have to wait, you have to use extra usage, you have to upgrade, you have to use a different account, whatever. But with Codex, you have more time, I think. You have more tokens. It feels like you can do more before you run out of tokens. But also, it's kind of not clear to me when I actually run out of tokens. It's a little bit intransparent. And of course, with Open Code, it dep- ends on the model that you use. So, this is less about the coding agent, more about the model and the plan behind it. I think that you can get for 20 euros or 20 dollars with Codex, you can do more prompting and get more problems solved compared to Cloud Code. Cloud Code actually wastes your tokens faster, but I also personally think the quality is better. Now, to try to make this a little bit more objective, even though it's not really going to be possible, especially not with one example with the sample size one, but what I did here is I had a test prompt. This is the prompt that I entered. Basically, code a Sudoku browser game. The game should have a button for generating a Sudoku and a drop-down for selecting a difficulty level. And yeah, basically, I gave it to stack Python back end, HTML JS, um CSS front end. Nothing fancy. This was just a prompt that uh yeah, needs to use some intelligence to get it done properly, especially with good design and everything. And I ran this prompt through Cloud Code, Codex, then also Open Code, which is Open code with Codex, so with GPT, and then open code with Deep Seek, MiniMax, NeMo Tron, Quen, and uh this is a different directory. And now we can take a look at the different results. So, this is the Claude one. We can choose a difficulty level. We can say new game. It's going to generate this based on the difficulty. I can also go with expert here. Uh and we can type stuff. So, for example, here one will be flagged as impossible because I already have a one here. Maybe if I go with something like four, no, also not possible. Maybe with something like six, that would be possible. And basically it's a very simple UI. It probably works. I can also solve it and get to the correct answer here. Uh I can check if it's solved or not. Maybe I can I can also add some mistakes here. But it's a very basic result here and it works. Then the second thing is what I got from using Codex. So, Codex in Codex actually. And here I have the same idea. I can generate a puzzle. I can also uh type in stuff like two. For example, here we also have a notes feature that I didn't ask for, but it's quite convenient because I can go say notes and then I can type uh stuff like this in here. It doesn't count as an actual attempt. It just gives me the ability to write down some numbers. I can erase. I can check. In this case, no rule conflicts found. Uh it doesn't have a solver, which is fine. I didn't ask for it. But yeah, this is probably equally good. Uh I think I like the user interface of this one a bit more. But overall, I would say the quality is roughly the same. Then we have this one here. This is basically open code using Codex. And again, I can go here hard generate Sudoku. There you go. I can type stuff in here. Um maybe let's go with something that actually makes sense like five. Um and I can play this. I can check the board. I can clear the entries. I can also show the solution, which is also interesting. Uh yeah, probably I would say equal quality. I like the design of this one more, but this is probably just a random noise thing. Overall, the important thing for this particular project is I don't see a huge difference in the quality of code when it comes to open code versus Codex, but also to be fair, it probably also really only shows when you work on a large code base and you do some massive refactoring. And there, to be honest, I haven't used Codex or open code that much. And there the couple of times that I used open code and Codex, I didn't see a huge difference. I still like Claude code most, but I didn't see a huge difference in quality when it comes to open code versus Codex when using the same model. But what's interesting here is I also have results from the open source models, from the free models that we can use with open code, and this is what they look like. So, we have here MiniMax 2.5, I think. And basically it also works, so I can type a number like three for example. The UI is good. Um I can also use the keys here. I can use backspace to clear. So, good enough for something like this. Then this is what I got from Quen 3, so just go and generate. It allows me to do three mistakes for whatever reason. I'm not sure um what's what the logic behind this is, but that also works. Maybe if I go with four here, that's Actually, it gives me some animation if this doesn't work, even though it's probably okay. Uh yeah, I don't like this one too much. This seems buggy and odd, so that's not a very good performance. Then this is what Deep Seek gave me. That's I think version four, if I'm not mistaken. Doesn't work at all. This is complete garbage, so that's the actual actually the first one that failed entirely. And then this is what Nemo Tron gave me. Basically, it didn't even succeed in setting up the project properly. I ran the command that it wanted me to run, but it didn't work. So, yeah, but what's interesting is that if we take a look at the different code bases, for example, if I go to the Claude one, you will see that it uses the Flask framework, which makes sense if you tell someone to use a Python back end, you're going to go with Flask or Fast API or something. So, you can see it used Flask and the interesting thing about the GPT models, about Codex, is in both Open Code and in Codex, it decided to use raw HTTP in Python. So, an HTTP server, basically. If I go to test one Codex and then to app, you can see it actually uses http.server and serves everything manually, which is kind of odd. Maybe it took that too literally that I wanted to have a Python back-end and it thought that Flask is not allowed to be used because I also specified don't use external packages for the Sudoku solving and Sudoku generation, but that doesn't mean that you're not allowed to use Flask. So, for the for the web stuff. And same thing happened with Open Code because it seems to be a model thing. So, this doesn't prove anything. It doesn't prove that the harness is not important or anything, but you can see that for for the most part, I think it's the model that really matters and the harness can be the extra thing. Overall, the performance is pretty good with Claude, Codex, and Open Code when using Codex. And then of course, finally, we have the sympathy points and the open source aspect and here obviously the winner is Open Code. Open Code is open source. It is community-driven. It is open for everyone and the least points in my opinion go to Anthropic. Not only are they not open source, which is fine, they are actively blocking Open Code from using their subscription. So, the philosophy and sympathy points definitely don't go to Anthropic. I don't like OpenAI as a company either. So, Open Code definitely wins this one, which makes it overall my preferred, like my favorite tool, but not the one I use. So, there's a difference between my favorite tool, which is Open Code. I like the tool the most. I like using it the most. This is the project I support the most, but for practical purposes, if I actually want to get work done as of right now, I'm using Claude Code. This is just the sad reality as of right now. They have the best model, they have it exclusive, they have the best harness for their model. So, when I need an agent that I can trust with big refactoring tasks, I'm always going for Claude Code as of right now, but I hope this changes in the future. If at some point the GPT models, and some people think that that's already the case, but if it obviously becomes the case that the GPT models are more powerful than the Opus models or the Claude models, I would probably go to Open Code and use the GPT models there. So, my current workflow is I use Claude Code for most of the engineering, and when I run out of tokens, I don't use extra tokens, I use Open Code with GPT. So, yeah, this is basically my very biased and subjective evaluation comparison of Open Code, Claude Code, and Codex. The main reason I don't like Codex is the lack of tooling. I don't like the way that the harness is structured. I don't like using it. Um, and also I'm not a big fan of the models. So, for me Claude Code wins this overall, but the sympathy points definitely go to Open Code. Let me know in the comment section down below what you think, what your favorite tool is, and if I missed something that's worth highlighting. So, that's it for this video today. I hope you enjoyed it and hope you learned something. If so, let me know by hitting like button and leaving a comment in the comment section down below. Besides that, don't forget to subscribe to this channel and hit the notification bell to not miss a single future video for free. Other than that, thank you very much for watching. See you in the next video, and bye.