AI Coding Tools Are Getting More Expensive: How to Adapt with Models, Harnesses, and Skills

Baird

Yeah. Probably more important. And you know, I, I think a lot of it from my perspective, and I think you've been feeling a little bit too, is just, you know, we've kind of known it's coming because AI is been cheap and everybody's always like, oh, it's getting cheaper and cheaper. And, you know, people that really look behind the scenes are seeing. It's cheap because AI companies are trying to get you hooked on their product so they can then jack the prices up because

Brendan

Right.

Baird

you can't spend billions to trillions of dollars on infrastructure and then. Charge 20 bucks a month or whatever and expect unlimited usage. And so, you know, this week we're starting to see some of that come due just in, you know, if there's Claude Code might be coming out of the pro account at this point. And so that's definitely something where people that are on $20 a month accounts are gonna be. Feeling some pain pretty quickly. I've seen that like Codex is not planning on doing that. But at the same time, OpenAI does not have the best track record of telling you the whole truth. So it'll be interesting to see how that develops. It feels like a let you know philanthropic go first and then see how it goes and then, you know, change their model at that point and you know, between that and then

Brendan

from behind a little bit too,

Baird

for sure.

Brendan

they're, they're playing catch up now, so

Baird

Yeah.

Brendan

a little more willing to let it go longer probably.

Baird

For sure and, and I think at this point they might have more infrastructure than Anthropic does, but at the same time, I just saw that Anthropic signed a deal with Amazon, and I think they're putting like, what, 20 billion, 50 billion in infrastructure with them as

Brendan

Okay.

Baird

deals. Amazon is investing in them, but they're also putting a bunch of that money back into infrastructure. So yeah, it's hard to see where, where it'll go. But I definitely have been expecting the AI's gonna get more expensive. You know, you're not gonna be able to, and it was actually called out pretty interestingly from the cloud team, the Claude Code team, talking about the fact that when they created Claude Code and when they created actually the plans themselves, they didn't. Plan on Claude Code. They didn't plan on cowork, they didn't plan on agents like all of this stuff that they've created. But the plans were really for chat, you know, just straight up chatting without all of this ongoing long running usage. So it's definitely like, it makes sense. You kind of knew it, but. It's also a pain because you gotta pay for so many services these days because you can't just run one thing and do it successfully, and you also want to keep up with it. And so it's kind of like, how can you get around it a bit? You know? What are the alternatives these days? I mean, what have you been seeing on your end?

Brendan

Yeah, I mean, you, you shared an article and we'll link to it in the show

Baird

Yeah.

Brendan

that kind of kicked this off, but also even before this happened, you know, you and I are working together. We're together in parallel, I guess, to build a couple products that we'll talk about probably later if they come to fruition. But, you know, the, the point being though that we're both trying to get stuff done, we're trying to get, use these tools. both are on the cheaper plans right now, at least for personal use. So, you know, we're, we're trying to maximize essentially what you were just talking about they're potentially gonna take away. you know, I, I had seen even before, like, the last week when I'll, I'll just say, like, you and I were both using the, the Get Shit done GSD plugin or library. I'm not exactly sure what it's called. Um, but, you know, it's … My take really quick just on that is, like, it's

Baird

Yep.

Brendan

like, a, a Ralph loop or anything that's gonna basically try and, like, do the work without you doing a lot, but also it's gonna burn a lot more tokens that way.

Baird

Yep.

Brendan

And, you know, I hadn't been using any of those things before this week, and on the $20 a month plan, like, doing the little bit here and there, like, I never really hit my max. I would never hit

Baird

Sure.

Brendan

usage.

Baird

Yep.

Brendan

And then I used that on, like, a couple nights ago for the first time. And just in the planning stage, which is pretty thorough and I think it's pretty

Baird

Okay.

Brendan

of it,

Baird

Yep.

Brendan

like, I hit within, I think, 30 minutes maybe,

Baird

Yeah.

Brendan

just with the planning stage, I hit my 100% limit for the five hour mark

Baird

Yeah.

Brendan

I haven't burned through, like, my entire week's worth yet but I could definitely see where you would.

Baird

Oh yeah.

Brendan

so yeah, I mean, it definitely, there's a lot of things there, I guess. It's like to get, actually get a big project put together like this on the cheaper plan. I, I knew it was probably gonna be a problem anyways, but it's like this realization now that, A, yes, it is, you're not gonna get very far and because it's part-time, it's okay

Baird

Yeah.

Brendan

but if this was my, know, day-to-day thing, like, that would be a nons- non-starter there. So, but yeah, you, I think you, you've gotta almost, in some ways, a little more interesting setup that you've been running though because I'm only doing Claude Code

Baird

Yeah.

Brendan

you, you still use Windsurf, so maybe

Baird

Yep.

Brendan

bit about what you've been seeing.

Baird

Yeah, it's actually, you know, it is funny 'cause I'm essentially paying, I think $60, no, maybe it's $50 a month right now. So like half of a cloud max plan at this point. And. I've actually found an interesting kind of setup of, you know, I have windsurf, I've got Claude Code, and then I also have Codex that I run too. And I found it's been interesting because recently, um. Windsurf, you know, they've got all sorts of models that you can choose from, just like Cursor does. But they have Kimi K 2.5 and 2.6, which 2.6 just came out and they actually are free right now and based off of the benchmarks, 2.6 is supposedly rivaling opus. I don't know if it's actually 4.7 or 4.6, but it's been interesting because what I've found and just kind of to back up a little bit with the Get Shit done. I started using it a little while back because it's really interesting how, you know it goes through and it does, you know, you discuss your features like, you know, you essentially like create the roadmap. Then you discuss each phase, you plan each phase, and then you execute it and verify it. And so, and the cool thing is, is it like produces artifacts for each of those pieces.

So you have, you know, all of that memory baked in. And so I've had a lot of success kinda missing all of the, uh. You know, oh, it built it, but it didn't hook it up correctly and all that type of stuff because it has that context that it can keep going back to and it kind of keeps you on target. Like for me, I tend to side quest as we laugh about all the time, I do it when I'm, you know, using, you know, coding agents where it's just like, oh, but what if I add this feature? And no, what if I add this feature? And which you know, is cool and it's powerful, but it's also that like. You're making progress, but you're not actually making progress to your like, overall goal. You know, like you're doing little things here and there, but it is better if you like add it to the backlog, which you can do and get shit done. And so kind of going back around what we were talking about. I found an interesting kind of mix where you can use either Codex, you know, on, you know, for 5.4 on medium or high reasoning, or you can use Claude, whether it's Opus or sonnet on high, to do a lot of like the planning and stuff like that.

And even some like the more complex features like let it kind of churn on that. But for. And I actually wouldn't even say all the planning, like for the research stage. I actually let Kimi do that instead because it's free. So I'm running, you know, it can do all the tokens at once. It can run through it, and I haven't hit any limits yet there. So to save some of like the. Bigger horsepower, so to speak. I kinda let the other models do some of the more intricate stuff, but then let theoretically a comparable, but the open source model that runs for free do a lot of that, like churning type work. And I found a lot of success in it where it's like. I will hit my limits still on Claude Code pretty frequently. And even on Codex, after they changed their, like it was up until April, they were like, you get two X usage, just all the time that dropped off. So like I'm hitting my limits both in Codex and Claude Code. Pretty frequently. Like it's usually like two hours before my window's even up, I'll hit it. But before you'd have to stop, you know, you'd just be like, okay, I can't do anything until I've got my usage back.

Or obviously you can upgrade and pay more, but it just doesn't make sense in this context. And so having that other tool that has a free model, and I mean, the nice thing is Windsurf also does have, you know. Opus, Sonnet, you know, any of the GPT models. So if I have a bigger problem that I need to do in that window, I can still just use their usage as well, which is pretty reasonable. So, but I, I have found it also interesting because they're all different in what they're good at. You know, like, it, it's kind of a. I have seen Kimi do some actually pretty cool things and it actually does a pretty good job. But I've noticed with the new model, because its focus is actually on agent work supposedly. You're, I think you could run like a hundred or 400 agents with it. Like that's what it's kind of designed for. I actually noticed it sitting there like chatting to itself constantly and it like almost gets stuck in loops because it just sits there and like thinks about it and it's constantly like, oh wait, but what about this? Like, then two minutes later it's like going back in the direction it came from. And so it's also interesting to see how models are.

Expanding their reasoning, which is actually what I've heard about Opus 4.7 too, is like people are burning tokens faster in that because its reasoning produces so much more that your token usage has gone through the roof because of that versus the actual output. And so there's a lot of these like. Dynamics that are changing. And I, I think it's interesting and like windsurf is not the only way to do it. Like you can actually, you know, hook up, um, it's a GLM, uh, from z AI has a plan where you can actually hook that into Claude Code. And there's a couple others like that where you can run, like, you can actually even run like Gemma four. On your Claude Code, but it's super slow. I tried it one day and it took like 10 minutes to actually produce anything. And if you didn't have the token or the context window high enough, it just wouldn't even do it. So, so where I'm getting at there is that there's a lot of like the open source and alternatives that I think are gonna start being more and more. Useful as the big players are not able to actually keep up at a lower price, you know, to do their models and, and honestly, like I'd be interested in your take too.

I mean, are you feeling like Claude Code is still as good as it used to be?

Brendan

You know, that's a good question. I guess it wasn't really until 4.6 came out, you know, a couple months ago or, or maybe even 4.5, I guess, when, in

Baird

Sure.

Brendan

when I really started pushing it more to its limits, I guess.

Baird

Sure.

Brendan

I, I have been using it for a long time, obviously, and we've talked about that a bunch, but

Baird

Yeah,

Brendan

I, my, I guess I, I struggle a little bit because, like, it's definitely way better than it was a year ago,

Baird

sure. For sure.

Brendan

that's not even comparable, I, for, for what I've been using it primarily, it is more things like just me out, like the command center

Baird

Yeah.

Brendan

I've talked about over and over, and I'm still using … And actually, like, I keep pushing that even more,

Baird

Sure.

Brendan

even, I've got that, like, coding things for me now, like, not directly, but long story short, like, it, that, to me, that output probably is fine whether it's Opus 47

Baird

Sure.

Brendan

or, you know, anything else. So that, that one's a little bit harder to gauge.

Baird

Yeah.

Brendan

when

Baird

Yep.

Brendan

write the code with it, you know, I'm, I'm not looking at it close enough to really say, like, it's not doing a good job, it's getting the things done that I want it to. So I, I probably haven't seen as much as maybe you have.

Baird

Yeah, and it is funny though too because I mean, that's kind of a good point is that a, a lot of times there's the expectation of like, oh, I need to use Opus 4.7 or Sonnet four point. Like it really comes down to gauging what you're actually doing and being more thoughtful about. What is this like? Is this a complicated task? Is this novel, or is this something that's like a quick bug fix or like a simple feature that probably is very well represented in the training data? Like

Brendan

Right.

Baird

on the complexity, you can scale back on the model and still get good enough results. You know, it doesn't have to be perfect, but it's kind of a, you know, because we don't have to write the code anymore, there's more and more. Kind of thought that can go in. I mean, if you have unlimited funds, of course you can just run Opus 4.7 all the time and pay for the, you know, even the API and who cares. But if you have a budget, if you're trying to do this, you know, in a reasonable price, it's like. Uh, just putting a little bit more thought and care into what models you use. Especially as, you know, agent harnesses have become kind of, the, the harness is almost more important than the model, and I've seen a lot of that too. And so, like Codex is getting better Claude Code, like that's where I'm a little iffy, like comparing Codex to Claude Code to, you know, even windsurf cascade and all that like. It feels to me like Claude Code has actually gotten worse on the harness side of things.

It just feels like things have gotten a little weirder from that perspective, just in my personal experience, like maybe other people aren't experiencing that, but I'm not as, like, I'm not feeling that Claude Code's quite as far ahead as it used to be, which has been interesting to see. But it also is like I'm starting to see past just the.

Like there's that, you know, model plus harness mixture. That's just interesting to see. Plus like, what are you actually doing with it? Because it varies so much based off of how complex of a problem you're actually trying to solve.

Brendan

Right. Yeah, that's a good point. And, I mean, I do have access to Codex now, so I, I think … I, I wouldn't mind testing it a little bit. Like, I, I have a few things where, like, I will run a, uh, review, a pull request review or whatever, and then I'll … Some, sometimes I'll do it, I guess, a couple ways. Like, one is I will actually run both of them in parallel just to see,

Baird

Sure.

Brendan

do they find the same things or s- you know,

Baird

Yep,

Brendan

So far, like, honestly, they're pretty similar. Like,

Baird

sure.

Brendan

I have not, never seen anything drastic where or the other finds something the other one didn't that was, like,

Baird

Yeah.

Brendan

right?

Baird

Yep.

Brendan

Um, but then there's also times where I'll be like, "Okay, this was written with Claude, so I'm gonna use Codex to review it, " that sort of

Baird

Sure. Yep,

Brendan

Um, but I haven't really pushed Codex that hard on the actual coding stuff. Like, I've really just stuck with Claude Code

Baird

sure.

Brendan

partially out of being lazy, I guess, because I've got it set up it's e- while it's not that hard to really get Codex in the mix, it's just like, it's enough of a barrier where, which is also, I guess, something, know, it, it's kind of like that stickiness, right, that we,

Baird

Yeah.

Brendan

talk about with products all the time. It's like

Baird

Yep.

Brendan

by Claude, or sorry, Anthropic building all the tools, and even if Codex and or ChatGPT, I can't even say their name right, OpenAI is catching up, know, it's like, well, but Anthropic does have that built-in base now

Baird

For sure.

Brendan

using their products, the switching cost is getting higher now where it's not, not just switching the chat interface, you're switching all the harnesses and all the scheduled jobs and all those things, you know. So not exactly sure where I'm going with that, but it, it's a, I guess just another

Baird

No,

Brendan

of it all,

Baird

it, I mean, it is a good point and it's actually something that I've noticed myself just in, you know, with the rise of AI in general, it's like, and I, I see it too. With how people use it. Like even people that are not technical like us, it's like they find something and you know, at first it was chat, GPT. Everybody used chat, GPT, and like at work, everybody used ChatGPT for, you know, up until. The end of the year essentially,

Brendan

Yep.

Baird

and then all of a sudden as everybody kind of realized that, oh, Claude's out there and it's actually good, then all of a sudden everybody mass migrated over to Claude. But I don't see a lot of that like curiosity of like, Hey, I'm gonna poke around. Like me personally, I've used Claude as my daily driver for probably two years now. From pretty early on, but I still use chat gpt for some things. I actually use Gemini quite a bit too to do stuff at work. 'cause we've got it through our Google Workspace account. So I actually. So I try to push on all of them to see like what is good at what. And some of it too is like, because I don't wanna invest a hundred, $200 a month in one of the tools, it also then puts it like a artificial constraint essentially of. How can I achieve what I want to without investing all of my money in one place? You know? And then I can kinda see like if I use Codex Windsurf, Claude Code altogether, like what does that look like? Does that work great? Maybe it doesn't sometimes, you know, and maybe it would be easier in one place. But I do like the, and like what you were just talking about, the. Build it in one, review it in the other. It does.

I mean, no, they're not radically different in terms of their output, but I do find sometimes like one's a little bit more diligent about like how it reviews or how it tests or things like that. And I honestly, another thing I just popped into my head with that is the other thing I'm finding that's really starting to be. Kind of a game changer more than the models or even the harnesses are skills and like I've, I think, have we talked about superpowers before at all?

Brendan

I don't think we have on here,

Baird

Yeah. Like super.

Brendan

know what it is.

Baird

Yeah, we don't go too deep into it, but like superpowers has a ton of different skills and one of 'em is like test driven development. One of 'em is like systematic debugging and stuff like that. And having those skills is actually making a bigger difference than anything because it's actually using frameworks and different repeatable patterns for all of my projects. Like I install it for everything. And so. It then picks up those skills that actually works in Codex, it works in Claude Code, it works in Windsurf, all the same thing. So it's not like you have different skills in each, it's you can share 'em across, and that kinda lowers the barrier of like, okay, I'm gonna go from Claude Code to Codex. And so it's been

Brendan

Yeah.

Baird

interesting to see that. And obviously skills are popular, so it's not like this is a secret, but knowing that you can use it across platforms actually changes how you interact with them a little bit.

Brendan

Right. And that's a good point, because I, I also noticed that with the get shit done

Baird

Yeah.

Brendan

plug in whatever skill, like, it,

Baird

Yep.

Brendan

you could pick your model right there too. So it's the same kind of idea, which is a really good point. And, and I definitely would agree, like, feels like the power … Well, yes, these models are gonna have to get smarter to, you know, if they wanna replace all of us or whatever. But even if they … Let's just say they never even get past where they are today, so much you can already do and, but, like, leveraging the skills, to your point, can get so much more out of it. And I, I think that's just something, you know, maybe this is a good place to start wrapping up is, like, if you're someone that is conscientious, you're, you don't wanna spend the max plan or whatever or, or maybe eventually these plan, you know, those max plans might get to

Baird

Oh, they're gonna get more expensive.

Brendan

tokens too, right? And so, like, there is definitely something here that you wanna start thinking a little more about, not

Baird

Yep.

Brendan

things at these AIs, but also just, like, how do you leverage them? How do you use the right, you were talking about earlier, like, using the different models. And even, like, I've seen Conductor, which is a harness that I

Baird

Yeah,

Brendan

you know, I've, I've seen that one switch to, like, Haiku for writing

Baird

sure.

Brendan

Git Commit, because, like, why use Opus

Baird

For sure. Yep.

Brendan

th- that's side note quick, that's funny because then it's like, it tells you that Haiku wrote it, like, the whole commit, and I'm just like, "Whoa, that's no."

Baird

Hiq is just taking credit for the whole thing.

Brendan

Pretty much. It's like, okay, step over 4.7, I took care of this.

Baird

It's like I, I haven't been updated in two model versions, so I, I better take some credit for some stuff.

Brendan

Right, right. Like, I know how to code.

Baird

Yeah.

Brendan

But yeah, I mean, that, that's just, like, I, it's a really good point. There's, like, the skills and leveraging them and especially, like, the getting, the right skills together, I guess is another

Baird

Yeah.

Brendan

You know, there,

Baird

Yep,

Brendan

that … And this doesn't have to be things that are coming out of, um, superpowers or any skill tool. Like, you can make your own. Like, that's something that I, I'm gonna start doing a lot even for our show here, just, like, building skills that are, like, a

Baird

yep.

Brendan

essentially to, like, the automations that I'm already doing manually anyways,

Baird

For sure.

Brendan

I can just leverage a skill, which is just an automation, right? Like, that's something that I've been doing the dawn of me programming was automating things that I didn't wanna

Baird

Yeah. Yep.

Brendan

This is just another version of that, really, so … Yeah.

Baird

page where they have all the things that you can just download and install really easily. So it's an easy way to see a lot of what's out there. But yeah, to your point. Building skills is literally just a markdown file. Like it's just creating a markdown file. And you know, most of the models now have a skill to create skills, so you don't even need to do it by hand. You can just talk to the AI and it'll do it for you. So, but yeah, it's definitely something to look into.

Brendan

Yep. All right, cool. Should we wrap it up there?

Baird

Yeah. Sounds good.

Brendan

All right. Well, yeah, thanks for watching and, uh, you know, feel free to like and subscribe on YouTube. And if you're listening to the audio version, you know, feel free to rate us on Apple Podcasts. I know we have a lot of listeners there, so, uh, again, thanks for listening and we'll talk to you next time.