Gemini CLI vs Claude Code vs Codex Compared! What AI CLI is Best?

So in this video I'll be comparing cloud codec and Gemini CLI across a couple of different domains and seeing how they perform. And those are coming up with ideas, planning, integrating new features into existing projects. And the reason I'm doing this video now is because over the last few weeks, Codeex CLI has gotten quite good because they've been pushing updates like crazy. And also Gemini CLI, they've been pushing a lot of updates as well. And I haven't tried it since it came out, so maybe it's gotten quite good. Anyways, just to be clear, this video is not sponsored by anyone. I don't accept sponsors on my channel because I think they've bias your video or channel in a certain direction. But this video is made possible by the people who buy my own products using links and coupon codes down below in the description.

Anyways, for coming up with ideas, I basically want to know how I should integrate cloud storage into my application mind deck over here. And that's basically a LLM front end where you can compare many different LLM providers in parallel and you can bring your own MCP servers and add your own models and use any model available in open router and so forth. Basically all the data is stored locally on your device. All your API keys that you bring as well are stored locally as well because you can go to inspect and then you can see over here if you go to application you can see it's all stored locally on an indexed DB over here and basically I want to add a cloud option where all your data is encrypted and uploaded to cloud and then you can sync between devices basically and you have some kind of like master password that is only stored locally and used to encrypt your like data and if you lose your master password then you lose all your data. So you should keep that safe. But basically, I'm not sure whether to have my own database or whether to use like an S3 bucket. I also want to allow users to bring their own storage solution in case they don't trust Mind Deck. But I'm not exactly sure the best way of implementing it and the pros and cons of each. So I'm going to ask all of the providers, all of the coding agents to look through the codebase and then like suggest which is the best idea going forwards. So I have cloud code running on the left with Opus 4. 1, codec CLI running in the middle of GPT5 high and also a Gemini CLI running on the right to have Gemini 2. 5 Pro. So basically use my application HyperWisper. I'm going to index this codebase. So I'm going to select the file and it basically allows me to tag like the files just by using my voice and now I'm going to say can you tag db. ts. Basically I want you to look through the codebase and come up with a solution to this problem. I want you to add a like kind of cloud storage option for users. And I'll press option space to stop and it will transcribe it. And you can see over here it's going to tag the file as well by finding in the codebase. So when I press enter over here, you can see I actually found the file and then tagged it. So I'm going to copy this over. Actually I'm going to switch over to planning mode over here because codeex I mean cloud code does have a planning mode. And then I'm going to give the same prompt to GBT5 high in codeex and then in gemy as well. So it'll be interesting to see how they all compare and I think if they all suggest the same solution then I can be pretty sure that solution is the best solution as well because it's as though the LLMs are all reaching some kind of consensus. So you can see that cloud code and Gemini CLI is done and Codex CLI took the longest. So I'm going to read through all these solutions and then basically comment on them. So before asking the question, I had a feeling that S3 was going to be the best solution. But after reading through all of them, I definitely know it is because they all agree that S3 is the best solution. But I think that codec's plan here was probably the best because it actually included more options such as like incremental sync with per record objects that the others did not discuss and then also using pre-signed URLs which Gemini CLI did talk about briefly over here. But Codex CLI did provide a short-term, medium-term and long-term solution with like adding compressions to backups and uh testing connections and so forth. So I did actually quite like this. And as for cloud code, it did acknowledge that S3 is a good solution, but it did say there's larger sync operations because it wants to encrypt the entire database and uh sync that. Whereas codeci did actually recommend a incremental sync with per record objects. So basically each object record is synced like separately such as each chat group or each folder and so forth as like with the database or every table. And I think that this solution is probably the best out of all of them when it comes to option two here. So I think for this particular use case, I'd actually put codeex CLI number one when it comes to coming up with ideas because it did seem to take the extra step. I'll put cloud code number two and then Gemini CLI number three over here. But I would say that Gemini CLI and also cloud code are closer to each other than they are to codeex. But one small issue that Codex CLI did have is they got quite eager and started implementing the solution which I just deleted to make sure it's fair uh for the next step. Uh whereas Gemini CLI and claude code did not start implementing the solution partly because cloud code is in plan mode over here which prevents it from being able to write any code. Anyways

I'm going to copy the solution over make a new chat for all of them and then use a right model and then uh basically describe that solution. So seeing how all of them planned that solution. Hey, so basically I want you to implement S3 like cloud storage. The user can either use our own S3 buckets. So I just described what I wanted for about 2 minutes and now it's transcribing everything over here and I enter the text and basically you can read through it if you want to, but I'm not going to go through the entire thing. I'm going to paste it into here in plan mode. I'm going to write plan over here. And even though Codex CLI does not have plan mode, I saw someone say online that it's really good at following instructions. So if you tell it to plan and not make changes, uh then it will just plan. Uh so plan don't make changes. Whereas uh cloud code kind of forgets its instructions after a couple of steps and it starts to edit files that you told it not to edit. So whilst we're waiting for those plans, I'm going to start planning something else. And basically right now in hyper whisper you can either use your own local models by downloading the models over here uh locally on your device or you can bring your own API keys and use any cloud models for transcription and post-processing. So there's open AI and froic Google gemini and you for whisper you can use open AI gro fireworks AI and you basically set it when you have the mode over here you can see like provider gro then I can choose any of these gro models and basically the advantage of this becomes that instead of paying a monthly fee you only pay what you use and you only use API keys or the providers that you will trust with your data like there is no backend over here it doesn't go to the hyperp servers because there are none it basically goes directly to these providers over here. Um, and I want to add more providers because it makes more sense to like have a greater range of providers. And there's deep over here which I want to add. And they have some speech text models such as like Nova and I think they have like a medical model as well um that I remember seeing online. There's also Assembly AI over here who have their own speech text model which can be pretty cheap and has like slam one which is a brand new one and a universal one and some streaming models as well. And then there's also 11 labs over here which have their own speechto text model too. And I basically want to add all three of these providers for the languages that they support so that anyone can use them in the application. So we'll see how they all perform when it comes to planning this approach. So I'm going to say, hey, basically I want you to add 11 Labs speech to text as one of the providers when it comes to transcription. So I'll press enter here, and then enter here. So basically reading through this quick comparison of the plans, it seems like Claude Code and Codeex CLI did do pretty well on here. Gemini CLI did do the worst because it just did not consider nearly as many things. Um, I think realistically I'd probably combine the plan of plan one and plan two together in some way to make it even more comprehensive and then just kind of disregard plan 3 because there are some additional things that cloud code consider that Codeexi did not and vice versa. So, I think I'll combine those two plans together. One thing that I do really like is it's a device on boarding via a QR code. So, you can basically really easily connect another device which I do like the sound of. I had never considered myself. Um but I think that can come up in a later edition. It shouldn't come up now. Now it seems that plans for hyper whisper are done as well. So I quickly used 03 to compare them and plan one is cloud code, plan two is codec and plan three is Gemini CLI. And basically you can see that plan two across these different categories are the most comprehensive. uh you can see what I excels at over here whereas plan one like has the most amount of like ready to go swift tasks as Overy puts it. So honestly I think I will just combine plan one and two again over here. Plan three is like basically useless can be discarded. So whilst we're waiting I think that claude code and codeex CLI are basically at the same level when it comes to planning out something. And then Gemini CLI is basically number three again over here. And there are things that cloud code consider that codeexi did not and vice versa. I would say that codeexi did think for longer and it did look through more files when coming up with a plan which I think can be quite good. So I think it's probably worthwhile using both of them to like find gaps in each other's plans or find things that did not consider and then maybe combine the plans together in some more comprehensive way. One annoying thing

that I do find whilst implementation is happening is that cloud code like basically completed it all at once and I said are you done and I said it's done and codeex CLI like stops after a while and says phase one is done do you want me to continue to phase two and then I have to say yes continue are you done if not continue so you have to like keep prompting it to continue with a really long plan okay so

it seems for mind deck they're all done but before looking through the code I'm actually going to try each solution to see if it works straight out of the bat. So, going to the cloud solution. I'm going to press I over here to open it. And basically, when trying it out, I can't actually find where the setting is. It doesn't seem to be anywhere. So, I'm going to ask Claude where it is. Okay. So, I managed to get it to add the cloud storage option tab over here. And this is what I came up with. So, there's mind cloud and it has a master password over here. And this isn't set up because I need to add like authentication account system where they like connect subscription or something. Um, as for custom S3 storage, I'm going to add everything over here and then I'll see how well it works. So, I have all the details. I'm just going to press test connection. And it says storage not configured for some reason. And how about if I press setup custom storage and I get an error here as well. So, yeah, I think it makes sense. This is a pretty difficult solution to oneshot. But I'm interested how codec CLI did here. And yeah, here it is. It didn't

really fit it in with the UI. So it has like a status ideal operations conflict thing happening over here. And then when I go to settings over here, it has the same mistake of not actually adding it to the top the storage configuration. Okay. So now added storage and it has Cloudflare R2 AWS free and then S3 solutions and I'm going to give it the same path and I'll just use a booker name as a master password. Press test connection. uh fail to execute on this. Uh I wish it actually integrated in with the design better rather than adding this random

thing to the top. And now in Gemini CLI solution, it also does not have the storage tab here. So it's interesting that all three of them failed to add the storage tab. Okay, so basically because there are way too many files that were changed, it's pretty hard for me to read all of them. So I'm basically getting codeex CLI to make a comparison and also claude code between

claude code to make a comparison between them. So basically it describes Gemini solution as being like really basic works for small data sets whereas codec solution seems to work really well for large data sets with unlimited jotting. It seems production ready. It marks its own solution as being feature complete and advanced. Uh it does have encryption streaming web chunking. Uh Gemini did not implement that despite it being in the instructions as well. And yeah, basically Codeex's sync engine is considered production ready with retries. It's using web workers in the background for background syncing instead of just basic intervals. And basically the summary of the recommendations is that cloud code actually prefers uh codeex solution because it's the most robust, scalable and production ready with comprehensive error handling and monitoring. And this is like an analysis of the technical debt. But it does say like uh codeex's solution is the most complex to understand. But honestly looking at this I think when I go back and clean up the code to implement this feature it would kind of be like halfway between these two solutions maybe the uh codeex solution because I think the codec solution may be unnecessarily complicated uh whereby maintaining it in the long run and making changes to it might be quite difficult. So I think I'll settle for a nice in between the codeex and claw code solutions and it seems that codeex is not really coming up with a comparison between the all the solutions because I told it to compare it and it just did a basic feature implementation comparison which is largely the same across all of them and I asked it to do it again. I like changed instructions and then it basically just listed the files that were changed instead. Uh so we're relying on the previous comparison uh that cloud code gave us. So yeah, I think the main reason for this difference is that despite all of them being given the same instructions, uh GPT5, especially Hide does a better job of following instructions. And they even say on their website, like GPT 4. 1, GPT5 follows prompt instructions with surgical precision, which enables its flexibility to drop in all types of workflows. However, its careful instruction following behavior means that poorly constructed prompts containing contradictory or vague instructions can be more damaging to GPT5 than to other models. as it expends a reasoning token searching for a way to reconcile the contradictions rather than picking one instruction at random. And I have seen some people say this as well on Twitter where they tell Claude code not to edit a file and then 3 4 5 minutes into a long coding task it decides to edit the file for whatever reason whereas codeex just absolutely avoids editing that file because like GBT5 is better at instruction following. I think that if you do have really good instructions or you have a really thoughtful plan, then you will probably get more out of this, which I can see in this case because it implemented a better solution that was more to the spec of what we described earlier. But as for UI, I think that cloud code was better because it feels as part of the application. It doesn't add some kind of random thing to the top which doesn't fit in with the UI and it kind of follows the same principles that are being used in other pages in the application. So I think what I will likely do is I will use a codec solution but then I'll get like clawed code to like critique it and then also change a UI around and then fix any like bugs and also implement some testing as well.

Okay, so now I'll be checking the solutions to hyper whisper to see how well they work. Now basically starting out with the cloud code solution over here. If I go to API keys, I can see added a bunch more API keys over here. So like 11 labs, deep gramm and so forth. If I now get myself an API key from assembly AI for example, then I can go and paste it into here. And now I'll check if assembly AI's like transcription works. So I'll quickly just switch out the model from uh this to assembly AI. Choose best over here. I guess I would have to double check the model names because it seems as though maybe it may not be correct. I'll have to double check that. And I'm going to try transcribing something. Hey, so this is just a quick test to see whether this transcription works. And it's transcribing and postprocessing. And wow, it actually does work. And yeah, so I made the transcription over here. So it's right over here. Use a best tier and so forth. And I wonder if the languages work. I would have to check that later, but I guess we're checking the base transcription for now. As for the cost, I guess it barely costs anything for now. I think there is a minimum transcription of about 10 seconds that you do have to do. So you will be build for 10 seconds regardless. Uh, but yeah, it's super cheap. Anyway, let's now check how the Deepgram works. So, I already made a Deepgram account before and basically when I signed up to Deepgram, I got $200 in credits. So, I can make an API key and start using that and just call it Hyper Whisper Test. Hey, I'm doing a quick test on Deepgram. Uh, let's see if this works. And it failed again. So, it did not implement Dgram properly for some reason. I would have to check and it is making the request but there's that invalid query parameter. So I guess it didn't read through the documentation properly. Uh but anyway let's now check 11 labs and for some reason it is now stuck on transcribing over here. So if I go back to Xcode uh basically what is happening it just keeps saying not found 4041 labs API error. So I think it basically found the wrong endpoint online somehow. But yeah I'm going to

quickly check codeex and see how the codec solution compares. Hey, just checking whether OpenAI's codec solution works here. So maybe it did not integrate it. Oh, it did mean it seems to integrate it properly. Uh but I have this permission popup uh which I can fix later. Uh so let's also try deepgram Nova free. Save changes. Hey, I'm just checking if this works properly. Transcribing again post-processing. And wow, it actually did manage to implement deepgram gram properly. Um, and it did not face the same error with the vocab like last time uh that claude code had with. And let's also try 11 labs now. 11 labscribe. Save changes. Hey, just checking whether this works as well. And there's a network connection error. So maybe it's to do with like the API endpoint not being found or something like that. But yeah, this is good because I managed to implement both deepgram and assembly AI properly. I would have to check if the languages do match up. But unfortunately, it did put the API keys in the wrong section. It put it under post-processing instead of transcription, but that should be a

pretty easy fix. Now, I'm going to ask Gemini to build and then run its solution. Now, about 10 minutes in, Gemini is just going around in circles and like filling up its context window without achieving anything. So, I'm just going to give up on it for now and then basically rank them again. So, Gemini

CLI comes in number third again because it's just worse overall. When it came to hyper whisper, it did implement two of the free API endpoints properly, whereas Claude code only implemented one of the API endpoints properly. But I feel like claude code implements things or designs better. It has like a better overall understanding of the codebase in the sense it puts the API keys in the relevant section. It follows a design of the rest of the application and makes it feel as though it was part of the application like the developer had made it themselves. Whereas Codex CLI does seem to add like more comprehensive solutions and does follow its own instructions clearly. It does feel like it adds something to where it doesn't belong whether that's design-wise or it just doesn't feel like it was part of the application because it just kind of feels jammed in there. So it actually put these two neck on neck for different like reasons. Basically I would use cloud code when I want something to integrate with the wider application better and I would use codec cli when I want like more precise or better instruction following. I noticed that claude code does not follow the instructions or initial plan that you gave it as clearly over long horizon tasks. Whereas codeex cli does seem to do a better job at that. But I think if your instructions are poor or your plan is just like not fleshed out or if your plan is contradictory in some ways then it may not implement as good of a solution. Whereas cloud code doesn't seem to mind if your plan is slightly ambiguous or it's somewhat contradictory in some way because it will still try to implement a solution that works for that particular codebase. So I think when it

comes to integrating things into existing code bases as I mentioned in my last video I will continue to use cloud code to actually implement the features themselves like write most of the code but I will use codeex CLI to then check cloud code's work to make sure that it implemented everything that was required because it is better at instruction following and checking when things have been done according to their instructions and when they haven't. So honestly I think I will continue to use cloud code for most tasks. It just does seem to do a better job. I do still use codeex CLI to do any swift UI related tasks and also to make sure that cloud code is on the right track and keeping it in check because sometimes cloud code says I successfully implemented the plan or whatever and then I get codeex CLI to check and it didn't actually complete the entire plan and then I give those instructions back to cloud code as I mentioned last video and then it completes a plan and then claims to have completed the plan and I just go back and forth and yeah this is basically where the landscape is right now. I think Claude Code is still the best as of earlier September, but Codeex CLI does seem to be catching up slowly. But I think Codeex CLI would take the top spot if it was better actually testing its own solution by always running the builder command had other features that cloud code has like sub agents or background tasks. And it would also be good if it was better at design and also better integrating a new solution into existing projects. The landscape is rapidly changing and I will come back to this in a couple months when things may look a bit different. But for now, if you do want to check out the applications that I'm making, then there are links and coupon codes in the description down below. But as I said before, this video is not sponsored by anyone. I don't accept sponsors on my channel, but it is supported by the people who buy my applications and products using links and coupon codes in the description down below.