AI Coding Agents Compared: Codex, Claude, MiniMax, Hermes and Antigravity

different harnesses. And I'll tell you why, because I think in the agentic world, not only does model matter, but but harnesses are going to matter also. So, what I wanted to do is go over the ones that I've been testing recently, and then uh go over where I think the general are in terms of capabilities and capacity and uh my personal reviews of each of them. So, let's get started on why I'm actually looking at harnesses, right? So, uh a couple months ago, the I was using Open Claw for pretty much everything, and I was using uh Anthropic models inside of there. While I was in Puerto Rico, they actually banned us, and everyone got kicked off of uh Open Claw with Anthropic using their subscriptions. So, I'm a big fan of using subscriptions, and that's the other thing I wanted to show you guys today. I'm going to show you each of the different subscriptions that I use, their pricing, and also what you get for that price. Because it's wildly confusing out there. People didn't even realize that uh OpenAI's pro plans, their their pro max plans, their biggest plans, they do have subscription limits, and some people didn't even realize that. They They thought it was just completely open. But now, you notice here I created a thumbnail, and this was using this this harness right here is actually Codex. So, uh this is the Codex harness, and in my opinion, well, and it's not just my opinion. You can look at the other harnesses. Uh they've all been modeled after what was done in Codex. This is the right way to build a harness in terms of uh look and feel, right? It is I mean, the fact that uh Claw the Claude code desktop app now looks like this, um the Minimax code app looks like this. Anti-gravity now looks like this. Like they all look like it now. And they did not before. The in particular Cloud Code desktop was particularly terrible looking in my opinion up until recently. Uh, this one has massive massive changes. So, let's take a look at each of these desktop apps. The only one that I don't have installed, I use the Cloud Code app. command prompt. So, you'll see the Cloud Code command prompt back here, and I just run, you know, dangerously skip permissions to be honest with you most of the time, uh, because the work that I'm doing is similar to the way Open QA works where I give it permissions to this particular machine, and it goes out there and does the work. So, that's the way I run this typically. And why am I not using their desktop app? Even though they made it, uh, look far far better than it used to. Well, the reason why I'm not using their desktop app is because it started hallucinating wildly on me. So, the Cloud Code desktop app seemed to really still have problems, and I'm not willing to investigate it because I tried the Cloud Code desktop app once, and it was horrible and unusable in my opinion. And then I tried it again, and I'm like, "Ah, they finally fixed it." It looks so much better. Looks like Code app. Uh, they finally fixed the way the way it looks and the way the user interface works. And then after 2 days it just started hallucinating like crazy. Uh, is it the fastest in the world? No. We'll talk about which one I think is the fastest. And and in fact it was pulled up first. So, spoiler alert. The thing that I think is the fastest is Code app with GPT 5.5, and I am using it for high right now. Uh, but I I actually should dial it down to high. This is where I would prefer to keep it, but for I I had it up at extra high whenever I was building this image. This is for sure the fastest, most accurate, in my opinion, of these artists. And this is the one where I'm paying $100 a month. Now, you notice I have So, I have the Claude stuff in my browser. All right, so this is pulled up. This is Claude AI new settings usage, right? That's where this this lives. Okay. That's where the usage meters live for Claude. So, however, for Codex, I could not find a web interface to find the usage meters. So, for the usage meters for Codex, it's settings, usage remaining, and down here it will show you. And this is at uh this is the $100 plan, and I will tell you what So, up until a couple up until a week ago, I could never hit their 100% usage even with try because they were doing that double promo. And then I was really nervous cuz I had it down around 30% usage, and I still had uh days and days left on my on my thing, and it just reset. So, it just magically reset out of nowhere and gave me back 100% usage you had. Now, it's just here's your token quota, and here's how much you've used, just like the other models. You know, just like what you would have with Codex. And I was very nervous that the usage that they were going to give you would be so neutered that it's unusable. But, they also gave, whenever they made this change, they gave you 20,000 tokens. And I've been using it the most aggressively I've ever used MiniMax, and I am still at 6,000 tokens. And you can see right now I have a ton of my quota for the $5 limit. But, realize I was I was taxing this thing and hitting it as hard as I could. The weekly limit is set to unlimited right now, which is just awesome. Uh and and in terms of being impressed with a model and a harness, I went into Minimax Code thinking it would be absolutely terrible, and it is actually pretty good. You know, it is as far as a harness goes, it seems to be as capable as any other, if not more capable. So, to be honest with you, I was shocked at Minimax Code as a harness. So, I've only been playing with this for like 3 4 days now. You can see I only have like three task histories, whereas if you go to Codex, I mean, this is, you know, all kinds of projects from all different periods of time, right? So, Codex is the thing, and I have a $100 Codex subscription. Right. Here with Minimax, the subscription that I have on that is this this token plan is if I go to token plan, I have the $20 a month plan, uh but billed annually. So, realize with Claude, sorry, with this one, with Claude, so, with Claude, I have the uh $20 plan billed annually, basically, and with Minimax, I have their $20 plan billed annually, which brings down the cost to under, you know, way under, you know, way under 20, and where I wanted it moved it moved to. But, otherwise, it is not terribly bad for uh for something that is $100 a year versus, you know, something that I'm paying $100 a month. Realize the difference in price between these two is this MiniMax M3 subscription, which can power uh two Hermes agents at the same time, plus powering the uh MiniMax code, right? And it can it it can power all of that at the same time for $100 a year. Versus this Codex subscription is $100 a month. It's 110. Now, do I think uh and by the way, up until today, I had Hermes on the DGX Spark, uh one of my machines. That was pointed at at uh ChatGPT 5.5, right? And then uh I had Hermes on this machine pointed at MiniMax 2.7, and I just upgraded that to uh you know, MiniMax M3 uh whenever I put whenever I installed the Hermes agent desktop. And so, what we're looking at here is Codex being, you know, massively massively more expensive. But, it is especially with ChatGPT 5.5, it is incredible what you can do with with these tools. So, my personal opinion is uh MiniMax is the biggest bang for your buck. But, realize that I used to describe MiniMax 2.5 and MiniMax 2.7 as the hold my beer model, which uh meant that it would it would, you know, that's a saying that they say in the country. And what what's going to happen is if something is going to happen after that, but it's probably not going to be what you wanted. Right? So, that's the description of it for me. Uh that's one of those things where I came from the country and And would be a thing, you know, that you know, you'd hear hold my beer and then someone's going to get into a fight or someone's going to try to pull something out of the ground that they shouldn't. Like all kinds of chaos would ensue, but something would happen, right? MiniMax was uh MiniMax 2.5 in 2017 really had the capability to do things, just not correctly. And I would say with M3, they're getting much, much closer to being able to do a gigantic task properly. It's just it still requires multiple prompts. Whereas with OpenAI, usually it's a single prompt quality results. Very rarely do I have to prompt it again. And that's specific to ChatGPT 3.5, by the way. Uh and with earlier models, it really didn't work well with And the the reason why I moved most of my tasks over to Hermes Agent is because Open Claw updates are just off. Today's update was the first update with Open Claw where I did not have to actually run And even then, I still ran I still believe it I really expected it to be broken after the update. And for the first time in a month and a half, the update didn't randomly break something. And every other update has randomly broke something. So, pretty wild to think that we got, you know, that kind of bugginess, you know, from Open Claw. And it's just consistent, whereas Hermes Hermes Agent though, uh everyone's saying that this thing's great and like there's no problems with So, I've never I I had one instance where OpenAI stopped working within it, and it turns out it was a bug on their end. They opened a ticket with them and they fixed it. I I expected it to be a bug on Hermes' side, but no, it was actually a bug on OpenAI's side, which kudos to them, by the way, for fixing it so quickly um and allowing you know and allowing the patch to go through whatever needed to go through for Hermes to be back up and running quickly. Uh but it definitely created a chaos situation where everyone's agents broke off once they were on Open AI, including my DGX Spark. So, that's how I knew it was specific. It wasn't Hermes, it was Hermes plus Open AI because I have one agent unit one agent that was using uh Minimax on this machine itself things ready and the other one on the DGX Spark was using Open AI 4 now. Now, in terms of Hermes agent itself, where would I rank it? I can tell you it seems far slower than uh it seems far slower with any, you know, Open AI uh you know, Minimax, any model that any other arcs, including the desktop app uh versus like if you compare Minimax codes response times to Hermes agents response times, it's night night day. It is like Hermes is so much slower. In terms of the quality of the output, I don't think it's any better than Minimax agent would be, but again, this is preliminary um because Hermes agent I've only been using inside the desktop app for today. Whereas but I've been using Hermes agent outside of the desktop app for quite a while and I can on Telegram and I can tell you it's it's slower. Like it's slower than any of the other it's slower than you know, the only one that rivals it in terms of speed is definitely Antigravity, which I'll get to next. But my personal opinion on these things is Hermes agent is definitely worth experimenting with. I don't think it is as great as people's hype. I do think that it's better than open call in that in that it doesn't break every time you update it. Uh but I will say the initial boarding I had uh the Hermes agent desktop app I had to have I spent like an hour or so getting everything so that I could just get the switcher so that I could switch between models. Uh and it showed all my models. I took like an hour, you know what I mean? And what's funny is I actually had offloaded the task to Claude uh code because I was frustrated with how slow Hermes desktop was to respond to me. And Claude code's response came back as Hermes agent was actually fixing. With with Mini Max M3 Hermes agent was actually able to fix it. And so here's anti-gravity. And if I go in here to find out your usage, by the way, anti-gravity's another one where I could not find it on the website, but in order to find the usage if that was available you go in the settings and then you go into models and you can see right now it looks like this. This is not how it is not how it looked earlier today. These were all like I used all of it the quota. But uh you know, I had the end of my work day and then and then here. So uh so like last night this was all used up and into this morning it was all used up. And so now because I had reoccurring tasks that were running as I was working through the day. So I had stuff that was still in process. But now everything's freed up and uh one big thing is switching models before. So these have different you can't see it here. I I should probably send you some messages so that you can see. But I'll just describe it for you real quick. So all of these Google models have one quota and then the Claude models have separate quota. They the bars go down separately. So when you pick a Gemini model you only have one. Like you don't actually have individual photos for all of these. They all go down together. But the Claude models go down separately. And if you had to switch contacts before between Gemini and Claude, it couldn't do that. Um and I don't know when that changed or if it changed because my subscription changed. So I don't know which one caused it to change. But I can tell you now I was able to today switch between Gemini and Claude. It had all my contacts and switched providers and then just finished the task because I ran out of Gemini credits. So that's pretty cool to see. You can just switch the model and say continue on this and boom. And I strongly recommend plans over individual tokens because I don't want to have to worry about or think about the prompts that I'm giving, clearing sessions. I don't want to have to think about that as I mean you do have to think about clearing sessions a little bit and freeing up memory context, but it's not as crazy as if you're paying per token. I feel like I would over analyze what I'm sending to the AI if I was paying per token and I strongly recommend of all of these plans, I would say that in terms of ranking them, if you can deal with multiple prompting, the minimax plan is the best. Like that $100 plan will actually get you uh get you a ton of usage. Now, if you want frustration-free best for any tool that you could possibly want to use, whether that be Hermes, whether that be Open Claw, Open AI is your choice today. If you're locked into the Claude ecosystem or you want very specifically Opus, the latest Opus, uh the latest Opus, then you're going to have to go with that. Like you don't have a choice if you really like Opus and Sonnet other than to pay for a cloud subscription. I really debated about never paying into a topic another dollar after the open flaw thing, but I really, really do like Opus 4.8. Uh I There was definitely model degradation for a while there, but it feels like it's back. Feels like it's just as smart as GPT 5.5. And each of them have different quirks of different things that they make mistakes on. And it's so good to have both of them to AB test. I can tell you I barely ever hit the limits on Codex. This week was the first time I thought I was going to hit it, and then it man magically just reset to zero today. So, I would strongly recommend people into the $20 Codex plan was just not enough. I was hitting it all the time uh within within open flaw. And so, I had to bump up to the $100 plan. I originally had the $20 Codex plan. Oh, well, I originally had no Codex plan. I was using free tier. They used to give you a ton of usage on the free tier, which is what got me hooked on the Codex app. So, in terms of being hooked on one of the apps, the Codex app is the one that I'm the most hooked on in terms of usage. And you can see it has the most. And this is after deleting a bunch of things from the projects because it was just too much. So, my personal opinions on this, again, to kind of summarize, uh anti-gravity with that with that pro plan is incredible. Uh Minimax is just incredible across the And M3 is really impressive as long as you can deal with multiple prompting because you are going to have to do that. Uh Codex as an app, in my opinion, is the best, but the $100 plan is what you're going to have to go with if you really want to be able to use it, especially powering agents and and within the Codex app itself. The other ones just don't give you enough usage. Slot code, it's going to vary. I I find their desktop app to be garbage and maybe they fixed it, but I'm not giving it another shot after giving it two shots.