How I Decide When I Should Use Claude or GPT or Something Else...
BDBen Davis@bmdavis419 · 45.9K subscribers
- Views
- 21.4K
- 21,394
- Outlier
- ×1.0
- Median 21.4K
- Length
- 21:26
- Published
- Aug 7
- Friday
- Comments
- 85
- 690 likes
Script
How it's written
- Words
- 4,966
- Words a minute
- 231
- Reading time
- 21 min
- Segments
- 15
The hook
0:00–0:25With how absurdly fast the AI world is moving right now, I wanted to do another update to the models and harnesses that I'm using every single day. I spend way too much time with these things, burn way too many tokens. I have very strong opinions on stuff like GPT versus Claude, the new open weight models that have come out, and even the really cool stuff from like xAI, the new Grok 4.5 model is very, very good. While unfortunately, this will go out of date within a month, I'm sure, just because
Structure
With how absurdly fast the AI world is moving right now, I wanted to do another update to the models and harnesses that I'm using every single day. I spend way too much time with these things, burn way too many tokens. I have very strong opinions on stuff like GPT versus Claude, the new open weight models that have come out, and even the really cool stuff from like xAI, the new Grok 4.5 model is very, very good. While unfortunately, this will go out of date within a month, I'm sure, just because of how fast this stuff moves, at least as of August 5th, 2026, this is what I'm using and why. And the best place to start, of course, is going to be with the models. I kind of want to try and explain the axes on which I think about these models, because I do actually use a pretty wide range of models in my day-to-day work. The two most common ones for me to use are GPT 5.6 all and Claude Fable 5. I have dedicated videos on both of these. Most of my opinions have not changed too much since then. The one that really has changed the most is actually Opus. I've really soured on this model, and I think most people have as well. Did a full dedicated video on that as well, so not going to waste too much time on that. And this diagram I drew out, I think is the best way I can kind of explain how I'm thinking about these things. On the Y axis, I have more vague to very specified, and that's talking about the prompt. When I tell the model to do something, how specific do I need to be in order to get out what I want? And then the other one is brute force versus beautiful code that scales.
Watch from 0:00Brute force is more just like, okay, I need this to solve a problem, don't really care how it solves it. The code quality doesn't matter. I don't really intend to merge it into production. It's just doing something on my computer or doing some test or data analysis or whatever. Quality is not the point there. Solving a specific problem is the point there. On the other side of things, this is where the actual output that matters is code. Code that is going to go into a production service with end users that needs to be robust and scalable and make me not hate myself when I hit the merge button on the PR. It's code that I would actually want to read. And I have this bookended with Fable on the more vague prompts plus really beautiful code comes out. And then on the other side I have 5 6 Luna as the kind of brute force hyper specified prompt, kind of like the ultimate small model. And then a ton of stuff in between. The only model I've been using that I didn't include on this diagram is Kimmy K3. It's a very very good model. The reason why I've kind of stopped using it is because I've been getting a lot more into like local model stuff. I actually just got and set up some DGX Sparks, which I'm very excited to talk about that soon. And DeepSeek just released a new checkpoint for DeepSeek V4 Flash. It's uh I believe it's the 0731. This model is absurdly cheap, really fast, and is about on par with something like Opus 4 6. It's an absurdly good model that punches way above its weight class. Luna's in a similar scenario here, but the reason why I use both of these because on paper Luna is better than DeepSeek V4 Flash.
Watch from 1:25But the difference is since DeepSeek V4 is open weight, you can get away with a lot more on that thing. Luna's still going to be subject to the normal safeguards that any frontier model API is going to have. V4 Flash will pretty much just do whatever you want it to, even if you're hitting it over API from a hosted provider in the US. There's still not really all that many safeguards, and if you just tell it, "Hey, can you go try and hack my site to find security vulnerabilities?" It will happily go off and do that. When something like Fable will give you a lecture and then turn into Opus 4 8 and fall apart. You really can't use the frontier for that kind of thing. You can use the open weight, so I included it here. I put it down here because right out of the box it doesn't write amazing code. It's functional, but it's not beautiful. It's not nearly at the quality of something like Fable. I have it more towards brute force because again, you really need to give it a lot of uh examples and direction on what the code should look like in order for it to be good. Right out of the box it's just not there. Same thing for Grok 4 5. This model did surprise me a lot. It is really fast and really cheap and definitely worth trying out, especially in something like Cursor or even the Grok build CLI. That's one of the most impressive CLIs I've actually tried from any of the Frontier Labs. I included Opus 5 here to just have it as a reference, but in reality, I have not touched that thing since the first couple days of using it. It looked so promising and then fell over so hard so fast it is not a very good model that I would not recommend trying. The only Claude model in my opinion that's worth using right now is Fable. But luckily, today's sponsor is worth using. I want to do something very different for this ad. I just need you to bear with me and trust me on this one. It caught copy-paste the URL for this video into
Watch from 2:50your coding agent and tell it to go look at the ad spot for today's sponsor Encore. Because Encore is one of the most impressive infrastructure as code platforms I have ever seen and the easiest way to show this off is to just have your agent go through and build something out with it. The example I did here was I told it to build a dependency radar with Encore and SvelteKit. Every 30 minutes sync the top 50 NPM packages, store everything in Postgres, use Encore APIs, cron jobs, pub/sub, put together clean front end, add in searching and individual package names, research the APIs, keep the scope simple, and make the whole app run locally. And once that generation finished, I ended up with this. It is a full app that has the compute for an API, database, pub/sub, cron jobs, could have object storage, but this specific one didn't need it. And all of this is being handled directly by Encore. I really wish I had more time in this ad to explain everything that they do and genuinely just ask your agents about them. They work incredibly well with agents because all of it is infrastructure as code. It's kind of the next generation of Terraform where instead of a gigantic hellish Terraform file that you don't want to ever deal with, everything is defined within normal TypeScript the way you would be used to. Like you're creating a new SQL database and this is how you define a SQL database. That's it. It's normal code in your app. So the app is defining the logic, not some giant config that's bolted on top of it trying to define the shape in some awkward way. Huge companies like Groupon and Ashby and many more are already using these guys to ship their apps because unlike most infrastructure as code platforms where they lock you into some like baby cloud, these guys are
Watch from 4:22built to orchestrate AWS and GCP directly. So, when you define your SQL database, you're actually defining a real Postgres database on AWS. Same thing with all of the other services. I keep scrolling and finding more things I want to talk about but I just I have to cut the ad here. Trust me, go give Encore a try at david7.link/encore. And now the thing I really wanted to talk about here, which is the differences between GPT 5-6 Sol and Claude Fable 5. This is one of the more interesting ones because everyone I talk to has very different opinions on this. I know a lot of people who are entirely fable pulled. They cannot stand using 5-6. They find it to be way too literal, does not write good code, runs for too long in weird directions. It's just not a good model versus Fable just produces this beautiful code. It understands what you're saying. It's just nice to work with. I kind of fall more in the middle where if you look at my token usage over the last week, it is primarily Codex, which is just 5-6 Sol. You can pretty much proxy these two. Anything in Codex is 5-6. Anything in Claude Code is Fable. And then anything in Pi is either DeepSeek or 5-6 Sol. If you look at the breakdown here, 2 billion on 5-6 Sol, about 600 million on Fable, 100 million on Opus, 100 million on DeepSeek, and then just random stuff like this that just doesn't really matter. In my day-to-day usage, the models have sort of diverged into two buckets for me. The first is writing code for projects that ship to end users and run for a long time. That is pretty much entirely Fable's domain. If I'm writing application code, I want it to be good.
Watch from 5:52I want it to be Fable. But there is an entire other class of problems that I run into every single day to do my job and other random things like just use my computer where 5-6 Sol excels. It is kind of my general-purpose model for everything from computer use. A lot of the way I interact with my computer at this point is just going into the Codex desktop app, telling it to use computer use to go off and do something, and then letting it run. I also use it a ton for managing and transferring stuff between all of the different computers that I'm working with. I have my Windows machine, my MacBook, all the stuff in my home server. 5-6 Solver Codex manages basically all of that for me. I'm also using it to manage and run everything on my Sparks right now. I'm not actually manually doing anything like setting up Deep Seek V4 Flash in order to run. I kind of just point the URL for the model and some reference repos at it and I'm like, "Hey, can you just spin this up and run it and then add it to my pie config and tell me it's ready?" And then it'll do it. It's a great model for just telling it to do something, letting it go off and do it. It's more efficient, it doesn't lecture me, and it is very, very capable. It You're just going to have to bear with me on this one cuz I know that this is a very strange diagram, but I want to try and explain why people like Fable so much and why I like it so much. The way I kind of visualize agent runs in my head at this point is kind of like a journey. So, imagine you're starting right here and then you need to end up here. Your code base is in this state over here. And the way we interact with this has changed over time. Like if I was making this video a year ago and I was using this example, I would say that like you start here, so your code base is in this state, and then you were describing the steps that it takes along this line to
Watch from 7:19reach some desired target. That's when you would be in the editor and cursor using like the sidebar view, change this file in this way to do this thing. With the big boy frontier models, that's kind of changed. We're at the point you to use the effect CLI for this. I want you to structure it in this way. But, these are more vague guidelines in order to get it to what I'm actually telling it to do, which is put the code base into this state, add this feature, remove this feature, change this thing. Whatever you're trying to do, you're describing the end point where you want it to end up. And in a perfect world, the model would start here and then it would make the long journey over here and we get to the target. But, that's not how reality works And prompting natural language, all of that stuff, is a very lossy medium for interacting with these things. We've done a lot over the last year and a half to improve this journey, {quote unquote}, with the harnesses we're building with. A great example of this is stuff like lint rules and check commands and that kind of thing, where the model would start here, it would go off, it would be kind of going in the right direction, but then it would make some dumb weird changes. Old models are very susceptible to this. It would start veering off down the wrong path. But, something like a check command would act almost like a block, like a wall here.
Watch from 8:51So, it would run into this wall and realize, "Oops, I need to go back this way." And it would bounce it back on track. And you would just keep doing this over and over again and ping-ponging it back into actually hitting what you want. That's still the case. We're still doing a lot of that. Things like skills and lint rules and check commands and even guardrails within the harnesses themselves, these are all just trying to keep the model on the right path to reach your desired destination. The problem you run into, though, is how do you properly describe what this target is? Like, how do you describe exactly, "I want you to end up in this place." In a perfect world, you could just like, if we actually use this map example, the target is at these exact coordinates, go to these coordinates. Then that is a verifiable thing where it can just straight shot. It doesn't really matter what happens, it can zigzag around, but as long as it ends up in the right coordinates, it's good. And it can check that it got to the right place, it's all good. There are surprisingly large number of tasks that can be described this way. Anything that is extremely verifiable fits under this category. Like, all the crazy math proofs that coming out of the last few weeks from the labs, those are very verifiable. Like, it's if you're telling the model, "Hey, can you just find a contradiction to this statement?" That's a clear yes or no. There's a clear end state there, and it can just work around in whatever direction it wants to, and eventually verify that it hit the exact right target. The thing is, everything you tell the models to do can't perfectly be described. If you could describe it perfectly, then you would just be doing it. The best way to describe the desired end state of like a code change is to just give it the end state code and then like what was the point at all. If you're telling you to add a button to the home page, that can
Watch from 10:08be interpreted in wildly different ways. Using the map example again, say we have the target here and you're telling it, "Okay, I need you to end up at the foot hill of this mountain." And you have the mountain floating right here. But the foot hill of the mountain could be anywhere like in this range. That's kind of a broad thing. Same thing with like add a button to the home page. Like even just add a log out button to the home page. There are a lot of implicit assumptions that go into that where you would assume that if you're adding a log out button to the home page, that would go in the top right hand corner of that page. The model needs to have the baked in understanding and discernment to figure that out from just a few words that you give it. The reason why Fable is so magical is because it is the best at intuiting kind of where this target is and it keeps itself on the right path. Or sometimes even it will go to a target that you didn't intend for it to go to, but ends up actually being better. That's the magical feeling that it has that nothing else can quite replicate. 5 6 is incredible at quote unquote crossing the ocean into like call anything in here is a successful code change. Like it works, it runs, it functions, we're good, but it might not be exactly what you wanted. And if you just imagine the straight line that we had from before, if you have a straight line here, this is the perfect change doing exactly what you intended for it to do, that starting angle right here matters a lot. And that starting angle is how well it interpreted what you wanted to actually do. Fable's really good at getting that right. 5 6 is really good at getting over here, but it's not always great at figuring out exactly what you want and can sometimes start going down very much the wrong path, even to a subtle degree. The longer a task is, the more important it
Watch from 11:38understanding where you're trying to go is because you just make a slight correction here, you just change that starting angle a little bit, you suddenly end up down here and you get something you really do not want. I've had this happen so many times with 5 6 where I will let it run for hours at a time, spin off tons of sub agents and it will successfully output something that works. It's successful, quote unquote, but it's not at all what I wanted. It's too complicated. It adds too much random stuff. The code itself looks disgusting. And as a result, you can't trust the model as much. But the shorter the distance you go, the less that matters. Like, if it's a very specific change, and you were just starting here, say, and you just need to get like from here to here, 5 6 is amazing for stuff like this. There's not much room for error, even if it's slightly misinterprets, like it's still going to hit it, and it's going to cross that distance perfectly every single time. And a lot of the models that I have down here, like V4 Flash, Grok 4.5, 5 6 Luna, all of these models do really, really well on shorter time horizons. The places where they start to fall apart is just going this full distance. 5 6 and Fable are really the only two that can go that full distance, that can run for absurdly long periods of time. But it is changing and increasing. Like, I would say a couple months ago, even the best models probably couldn't go further than that. But we're already at the point where even something like V4 Flash running on my Sparks in my closet can get pretty damn far. I've been watching it orchestrate sub agents, check its own work, reason through different problems, and figure out, "Oh, I shouldn't have been doing that," and correct itself down the right path, and get to a pretty good desired end state. It's improving a
Watch from 13:10lot, but the only model that feels like it can effortlessly do this journey right now is Fable. I think that's a combination of it being a really big model. It's a really good pre-train, and Anthropic is amazing at doing pre-train runs. I hope Open AI has an answer for this soon. Even things like Kimmy are getting closer to being an answer, cuz that's a really big model. That model does well at going this distance. But, yeah, I don't know. Like I said, this was a very weird, convoluted way of explaining this, but hopefully that's kind of helpful for how I think about these things. Cuz this also heavily informs the way I think about the harnesses that I use day-to-day. And that's the other thing I wanted to talk about in this video. What am I using to interact with the model every single day. The two main services are Codex and Claude Code. The Codex CLI and desktop app, and then the Claude Code CLI. I wouldn't don't use the Claude desktop app, it's not very good. The Codex desktop app is overall very good, and it kind of just feels natural and like home to me. The one that I probably should have added into this is actually Cursor. I've been using them more, especially like their background agent stuff is really nice. To me, Codex, Claude Code, Cursor, and Pi are all really, really good ways to actually run the core model. But, in this diagram, I have it set up kind of weird here, where like I put these two down here deliberately to just make a separation, because I have Pi up here as an agent that controls other agents. Because when I did my custom Pi setup, a big thing I added to it was a sub-agent feature that can call off normal Pi sub-agents the way it would work in a Codex or Claude Code.
Watch from 14:43But, the other type of sub-agent it can spawn is a Codex sub-agent and a Claude Code sub-agent. So, it will use the app server for Codex and the SDK for Claude Code under the hood to do a generation there. So, I could have GPT 5.6 Sol running in Pi, that's what it's interacting with. But, 5.6 Sol running in Pi is using Fable under the hood to write plans, review its code, even do some implementations, just orchestrating that thing. We're getting into this weird scenario where it actually makes a lot of sense for agents to control lots of other agents. This is the best way to use Opus 5 is I don't want to deal with it directly, but Fable is perfectly capable and patient enough to deal with Opus 5, and it doesn't burn your rate limits as hard. So, you can just kind of tell Fable, "Hey, use Opus 5 and keep a close eye on what it's doing. Be very, very specific with what you tell it to do. It's kind of stupid." It'll do a good job with that. It can control those other agents. I really like Pi for this. Like I showed earlier, I do use a pretty large amount of Pi at this point. The main reason why it's not nearly as high as something like Codex is one, I really like having a real desktop app for a lot of this work. The Codex desktop app is just kind of in a league of its own as far as on machine normal desktop apps go, to where I don't really feel the need to use all that much of anything else. The only other main interface I use for this stuff is T3 Code. Obviously, like I'm close with the Oh, yeah, yeah, yeah, whatever. But, the reason why I use it is because they genuinely do have the best remote experience. T3 Code is not a normal harness like a Codex or Claude Code or Pi. The agent doesn't directly run in T3 Code. T3 Code is an interface on top of
Watch from 16:10Codex and Claude Code. So, if you go into my T3 Code instance, you'll notice one, it's running on a URL here and that URL is a Tailscale URL that's going into the Linux box running in my home server so that I can send off prompts whenever I want to from the website and then just shut my laptop and they'll run in the background. T3 Code is the best interface for this by far and basically all my Claude Code usage here and probably like half of my Codex usage here is running through T3 Code. I should mention briefly too the subscriptions I'm using for this stuff. Right now, I have two Claude Max plans entirely because of Fable. The limits on Fable are pretty brutal and I burn through one sub way too fast. I haven't burned nearly as much this week cuz I was out on vacation with family last week so I wasn't using these things as much, but on heavy work weeks, I burn through two subs very, very quickly. I only have one Codex sub just because they're so generous with the resets to the point where I just don't really need another one and the model so efficient that I haven't run into any issues. In the future, I might get to a point where I will need to have a second one possibly. I have all this stuff set up through Vi Proxy. It's an open source project that I have running on one of my Linux machines and since all my devices are always Tailscaled together, it exposes an endpoint and that endpoint allows me to hit like the Claude Code API and it will under the hood load balance between these two accounts. And to be clear, I am only using this with Claude Code itself. I'm not using this to stick the Claude Code sub into some other harness. You can technically do that. It's totally possible by using it.
Watch from 17:41I wouldn't recommend it cuz they're very trigger-happy on banning and TOS violations. They don't want you using Claude anywhere outside of Claude code itself. But I think this here, just using it to split between two different subs, so I don't have to constantly log out and log back in, is generally fine. I also like having a web page I can go to to just quickly check on my limits, cuz these do burn down pretty quickly. One last thing I did want to mention before I close is that I have started a Substack. The main reason why I'm doing this is because one, I found that for all these videos, I think things through much better in writing. I just wanted a place to dump more unfiltered thoughts and breakdowns of this stuff. Like I have a companion article to this video on like the harnesses and models I'm using. This is not a replacement for videos. You do not need to go to this or subscribe to this. It is totally just supplementary extra stuff if you're interested. And I do have it behind a paywall. I really was rolling this around a lot in my head of whether or not I wanted to do that. The main reason I just wanted to have it be more brain dumpy and not have to worry about it as much. I want to be super clear, this is not replacing videos. Videos are not going anywhere. I'm not going to paywall those, ever. Videos and the info and things I'm doing, I do not want to be paywalled at all. I want that all to just be out there and available for anyone to get use out of, cuz people do seem to, which I do genuinely really appreciate. It is really cool that I'm able to do this and just talk about the random stuff I spend way too much time thinking about and have people enjoy it.
Watch from 19:06Really do appreciate that, and I'm not going to take that away. This is just an extra place for me to dump thoughts. I'm probably going to do this for a lot of my videos. I have one in drafts for my general thoughts on open weight models right now. That's another reason why I kind of wanted to have this like a little more gated off is because I'm going to get spicier in there. Open weight models are kind of becoming a politically charged topic, which really sucks. I want to avoid that like the plague, but I don't think we have much of a choice. And I really really like open weight models and local models and stuff like that. I think it's really dope. So yeah, hopefully this is kind of helpful to see where my head is on these things. Again, 5-6B Fable 5, Luna, Flash, Grok. Those are the main models I'm using. They're all fantastic. Just experiment a bunch and try pushing these things harder than you think is possible. We're really getting to a point where these models can do things that are far beyond anything I imagined would be possible for years. It's accelerating, it's scary and exciting and weird at the same exact time. I don't know. I I'm just enjoying the ride. It'll be what it'll be. If you enjoyed this video, make sure you like and subscribe and I'll talk to you again soon.
Watch from 20:25Full transcript
- 0:00
With how absurdly fast the AI world is moving right now, I wanted to do another update to the models and harnesses that I'm using every single day. I spend way too much time with these things, burn way too many tokens. I have very strong opinions on stuff like GPT versus Claude, the new open weight models that have come out, and even the really cool stuff from like xAI, the new Grok 4.5
- 0:20
model is very, very good. While unfortunately, this will go out of date within a month, I'm sure, just because of how fast this stuff moves, at least as of August 5th, 2026, this is what I'm using and why. And the best place to start, of course, is going to be with the models. I kind of want to try and explain the axes on which I think about these models, because I do actually use
- 0:41
a pretty wide range of models in my day-to-day work. The two most common ones for me to use are GPT 5.6 all and Claude Fable 5. I have dedicated videos on both of these. Most of my opinions have not changed too much since then. The one that really has changed the most is actually Opus. I've really soured on this model, and I think most people have
- 1:01
as well. Did a full dedicated video on that as well, so not going to waste too much time on that. And this diagram I drew out, I think is the best way I can kind of explain how I'm thinking about these things. On the Y axis, I have more vague to very specified, and that's talking about the prompt. When I tell the model to do something, how specific do I need to be in order to get out what I want? And then the other one is brute
- 1:23
force versus beautiful code that scales. Brute force is more just like, okay, I need this to solve a problem, don't really care how it solves it. The code quality doesn't matter. I don't really intend to merge it into production. It's just doing something on my computer or doing some test or data analysis or whatever. Quality is not the point there. Solving a specific problem is the point there. On the other side of things, this is where the actual output
- 1:45
that matters is code. Code that is going to go into a production service with end users that needs to be robust and scalable and make me not hate myself when I hit the merge button on the PR. It's code that I would actually want to read. And I have this bookended with Fable on the more vague prompts plus really beautiful code comes out. And then on the other side I have 5 6 Luna as the kind of brute force hyper
- 2:07
specified prompt, kind of like the ultimate small model. And then a ton of stuff in between. The only model I've been using that I didn't include on this diagram is Kimmy K3. It's a very very good model. The reason why I've kind of stopped using it is because I've been getting a lot more into like local model stuff. I actually just got and set up some DGX Sparks, which I'm very excited to talk about that soon. And DeepSeek
- 2:28
just released a new checkpoint for DeepSeek V4 Flash. It's uh I believe it's the 0731. This model is absurdly cheap, really fast, and is about on par with something like Opus 4 6. It's an absurdly good model that punches way above its weight class. Luna's in a similar scenario here, but the reason why I use both of these because on paper Luna is better than DeepSeek V4 Flash.
- 2:50
But the difference is since DeepSeek V4 is open weight, you can get away with a lot more on that thing. Luna's still going to be subject to the normal safeguards that any frontier model API is going to have. V4 Flash will pretty much just do whatever you want it to, even if you're hitting it over API from a hosted provider in the US. There's still not really all that many safeguards, and if you just tell it, "Hey, can you go try and hack my site to
- 3:11
find security vulnerabilities?" It will happily go off and do that. When something like Fable will give you a lecture and then turn into Opus 4 8 and fall apart. You really can't use the frontier for that kind of thing. You can use the open weight, so I included it here. I put it down here because right out of the box it doesn't write amazing code. It's functional, but it's not beautiful. It's not nearly at the quality of something like Fable. I have
- 3:32
it more towards brute force because again, you really need to give it a lot of uh examples and direction on what the code should look like in order for it to be good. Right out of the box it's just not there. Same thing for Grok 4 5. This model did surprise me a lot. It is really fast and really cheap and definitely worth trying out, especially in something like Cursor or even the Grok build CLI. That's one of the most
- 3:52
impressive CLIs I've actually tried from any of the Frontier Labs. I included Opus 5 here to just have it as a reference, but in reality, I have not touched that thing since the first couple days of using it. It looked so promising and then fell over so hard so fast it is not a very good model that I would not recommend trying. The only Claude model in my opinion that's worth using right now is Fable. But luckily,
- 4:13
today's sponsor is worth using. I want to do something very different for this ad. I just need you to bear with me and trust me on this one. It caught copy-paste the URL for this video into your coding agent and tell it to go look at the ad spot for today's sponsor Encore. Because Encore is one of the most impressive infrastructure as code platforms I have ever seen and the easiest way to show this off is to just
- 4:34
have your agent go through and build something out with it. The example I did here was I told it to build a dependency radar with Encore and SvelteKit. Every 30 minutes sync the top 50 NPM packages, store everything in Postgres, use Encore APIs, cron jobs, pub/sub, put together clean front end, add in searching and individual package names, research the APIs, keep the scope simple, and make the whole app run locally. And once that
- 4:55
generation finished, I ended up with this. It is a full app that has the compute for an API, database, pub/sub, cron jobs, could have object storage, but this specific one didn't need it. And all of this is being handled directly by Encore. I really wish I had more time in this ad to explain everything that they do and genuinely just ask your agents about them. They work incredibly well with agents because
- 5:16
all of it is infrastructure as code. It's kind of the next generation of Terraform where instead of a gigantic hellish Terraform file that you don't want to ever deal with, everything is defined within normal TypeScript the way you would be used to. Like you're creating a new SQL database and this is how you define a SQL database. That's it. It's normal code in your app. So the app is defining the logic, not some
- 5:37
giant config that's bolted on top of it trying to define the shape in some awkward way. Huge companies like Groupon and Ashby and many more are already using these guys to ship their apps because unlike most infrastructure as code platforms where they lock you into some like baby cloud, these guys are built to orchestrate AWS and GCP directly. So, when you define your SQL
- 5:57
database, you're actually defining a real Postgres database on AWS. Same thing with all of the other services. I keep scrolling and finding more things I want to talk about but I just I have to cut the ad here. Trust me, go give Encore a try at david7.link/encore. And now the thing I really wanted to talk about here, which is the differences between GPT 5-6 Sol and Claude Fable 5. This is one of the more
- 6:18
interesting ones because everyone I talk to has very different opinions on this. I know a lot of people who are entirely fable pulled. They cannot stand using 5-6. They find it to be way too literal, does not write good code, runs for too long in weird directions. It's just not a good model versus Fable just produces this beautiful code. It understands what you're saying. It's just nice to work with. I kind of fall more in the middle
- 6:39
where if you look at my token usage over the last week, it is primarily Codex, which is just 5-6 Sol. You can pretty much proxy these two. Anything in Codex is 5-6. Anything in Claude Code is Fable. And then anything in Pi is either DeepSeek or 5-6 Sol. If you look at the breakdown here, 2 billion on 5-6 Sol, about 600 million on Fable, 100 million on Opus, 100 million on DeepSeek, and
- 7:00
then just random stuff like this that just doesn't really matter. In my day-to-day usage, the models have sort of diverged into two buckets for me. The first is writing code for projects that ship to end users and run for a long time. That is pretty much entirely Fable's domain. If I'm writing application code, I want it to be good. I want it to be Fable. But there is an
- 7:21
entire other class of problems that I run into every single day to do my job and other random things like just use my computer where 5-6 Sol excels. It is kind of my general-purpose model for everything from computer use. A lot of the way I interact with my computer at this point is just going into the Codex desktop app, telling it to use computer use to go off and do something, and then
- 7:41
letting it run. I also use it a ton for managing and transferring stuff between all of the different computers that I'm working with. I have my Windows machine, my MacBook, all the stuff in my home server. 5-6 Solver Codex manages basically all of that for me. I'm also using it to manage and run everything on my Sparks right now. I'm not actually manually doing anything like setting up Deep Seek V4 Flash in order to run. I
- 8:01
kind of just point the URL for the model and some reference repos at it and I'm like, "Hey, can you just spin this up and run it and then add it to my pie config and tell me it's ready?" And then it'll do it. It's a great model for just telling it to do something, letting it go off and do it. It's more efficient, it doesn't lecture me, and it is very, very capable. It You're just going to have to bear with me on this one cuz I know that this is a very strange diagram, but I want to try and explain
- 8:23
why people like Fable so much and why I like it so much. The way I kind of visualize agent runs in my head at this point is kind of like a journey. So, imagine you're starting right here and then you need to end up here. Your code base is in this state over here. And the way we interact with this has changed over time. Like if I was making this video a year ago and I was using this example, I would say that like you start
- 8:45
here, so your code base is in this state, and then you were describing the steps that it takes along this line to reach some desired target. That's when you would be in the editor and cursor using like the sidebar view, change this file in this way to do this thing. With the big boy frontier
- 9:10
models, that's kind of changed. We're at the point you to use the effect CLI for this. I want you to structure it in this way. But, these are more vague guidelines in order to get it to what I'm actually telling it to do, which is put the code base into this state, add this feature, remove this feature, change this thing. Whatever you're trying to do, you're describing the end point where you want it to end up. And in a perfect world, the model would
- 9:31
start here and then it would make the long journey over here and we get to the target. But, that's not how reality works And prompting natural language, all of that stuff, is a very lossy medium for interacting with these things. We've done a lot over the last year and a half to improve this journey, {quote unquote}, with the harnesses we're building with. A great example of this is stuff like lint rules and check
- 9:52
commands and that kind of thing, where the model would start here, it would go off, it would be kind of going in the right direction, but then it would make some dumb weird changes. Old models are very susceptible to this. It would start veering off down the wrong path. But, something like a check command would act almost like a block, like a wall here. So, it would run into this wall and realize, "Oops, I need to go back this way." And it would bounce it back on track. And you would just keep doing
- 10:13
this over and over again and ping-ponging it back into actually hitting what you want. That's still the case. We're still doing a lot of that. Things like skills and lint rules and check commands and even guardrails within the harnesses themselves, these are all just trying to keep the model on the right path to reach your desired destination. The problem you run into, though, is how do you properly describe what this target is? Like, how do you
- 10:35
describe exactly, "I want you to end up in this place." In a perfect world, you could just like, if we actually use this map example, the target is at these exact coordinates, go to these coordinates. Then that is a verifiable thing where it can just straight shot. It doesn't really matter what happens, it can zigzag around, but as long as it ends up in the right coordinates, it's good. And it can check that it got to the right place, it's all good. There
- 10:55
are surprisingly large number of tasks that can be described this way. Anything that is extremely verifiable fits under this category. Like, all the crazy math proofs that coming out of the last few weeks from the labs, those are very verifiable. Like, it's if you're telling the model, "Hey, can you just find a contradiction to this statement?" That's a clear yes or no. There's a clear end state there, and it can just work around
- 11:17
in whatever direction it wants to, and eventually verify that it hit the exact right target. The thing is, everything you tell the models to do can't perfectly be described. If you could describe it perfectly, then you would just be doing it. The best way to describe the desired end state of like a code change is to just give it the end state code and then like what was the point at all. If you're telling you to add a button to the home page, that can
- 11:38
be interpreted in wildly different ways. Using the map example again, say we have the target here and you're telling it, "Okay, I need you to end up at the foot hill of this mountain." And you have the mountain floating right here. But the foot hill of the mountain could be anywhere like in this range. That's kind of a broad thing. Same thing with like add a button to the home page. Like even just add a log out button to the home page. There are a lot of implicit
- 11:59
assumptions that go into that where you would assume that if you're adding a log out button to the home page, that would go in the top right hand corner of that page. The model needs to have the baked in understanding and discernment to figure that out from just a few words that you give it. The reason why Fable is so magical is because it is the best at intuiting kind of where this target is and it keeps itself on the right
- 12:21
path. Or sometimes even it will go to a target that you didn't intend for it to go to, but ends up actually being better. That's the magical feeling that it has that nothing else can quite replicate. 5 6 is incredible at quote unquote crossing the ocean into like call anything in here is a successful code change. Like it works, it runs, it functions, we're good, but it might not be exactly what you wanted. And if you
- 12:42
just imagine the straight line that we had from before, if you have a straight line here, this is the perfect change doing exactly what you intended for it to do, that starting angle right here matters a lot. And that starting angle is how well it interpreted what you wanted to actually do. Fable's really good at getting that right. 5 6 is really good at getting over here, but it's not always great at figuring out exactly what you want and can sometimes
- 13:04
start going down very much the wrong path, even to a subtle degree. The longer a task is, the more important it understanding where you're trying to go is because you just make a slight correction here, you just change that starting angle a little bit, you suddenly end up down here and you get something you really do not want. I've had this happen so many times with 5 6 where I will let it run for hours at a
- 13:24
time, spin off tons of sub agents and it will successfully output something that works. It's successful, quote unquote, but it's not at all what I wanted. It's too complicated. It adds too much random stuff. The code itself looks disgusting. And as a result, you can't trust the model as much. But the shorter the distance you go, the less that matters. Like, if it's a very specific change,
- 13:45
and you were just starting here, say, and you just need to get like from here to here, 5 6 is amazing for stuff like this. There's not much room for error, even if it's slightly misinterprets, like it's still going to hit it, and it's going to cross that distance perfectly every single time. And a lot of the models that I have down here, like V4 Flash, Grok 4.5, 5 6 Luna, all
- 14:05
of these models do really, really well on shorter time horizons. The places where they start to fall apart is just going this full distance. 5 6 and Fable are really the only two that can go that full distance, that can run for absurdly long periods of time. But it is changing and increasing. Like, I would say a couple months ago, even the best models probably couldn't go further than that. But we're already at the point where
- 14:26
even something like V4 Flash running on my Sparks in my closet can get pretty damn far. I've been watching it orchestrate sub agents, check its own work, reason through different problems, and figure out, "Oh, I shouldn't have been doing that," and correct itself down the right path, and get to a pretty good desired end state. It's improving a lot, but the only model that feels like it can effortlessly do this journey
- 14:47
right now is Fable. I think that's a combination of it being a really big model. It's a really good pre-train, and Anthropic is amazing at doing pre-train runs. I hope Open AI has an answer for this soon. Even things like Kimmy are getting closer to being an answer, cuz that's a really big model. That model does well at going this distance. But, yeah, I don't know. Like I said, this was a very weird, convoluted way of
- 15:07
explaining this, but hopefully that's kind of helpful for how I think about these things. Cuz this also heavily informs the way I think about the harnesses that I use day-to-day. And that's the other thing I wanted to talk about in this video. What am I using to interact with the model every single day. The two main services are Codex and Claude Code. The Codex CLI and desktop
- 15:28
app, and then the Claude Code CLI. I wouldn't don't use the Claude desktop app, it's not very good. The Codex desktop app is overall very good, and it kind of just feels natural and like home to me. The one that I probably should have added into this is actually Cursor. I've been using them more, especially like their background agent stuff is really nice. To me, Codex, Claude Code, Cursor, and Pi are all really, really
- 15:50
good ways to actually run the core model. But, in this diagram, I have it set up kind of weird here, where like I put these two down here deliberately to just make a separation, because I have Pi up here as an agent that controls other agents. Because when I did my custom Pi setup, a big thing I added to it was a sub-agent feature that can call off normal Pi sub-agents the way it would work in a Codex or Claude Code.
- 16:10
But, the other type of sub-agent it can spawn is a Codex sub-agent and a Claude Code sub-agent. So, it will use the app server for Codex and the SDK for Claude Code under the hood to do a generation there. So, I could have GPT 5.6 Sol running in Pi, that's what it's interacting with. But, 5.6 Sol running in Pi is using Fable under the hood to write plans, review its code, even do
- 16:31
some implementations, just orchestrating that thing. We're getting into this weird scenario where it actually makes a lot of sense for agents to control lots of other agents. This is the best way to use Opus 5 is I don't want to deal with it directly, but Fable is perfectly capable and patient enough to deal with Opus 5, and it doesn't burn your rate limits as hard. So, you can just kind of tell Fable, "Hey, use Opus 5 and keep a
- 16:51
close eye on what it's doing. Be very, very specific with what you tell it to do. It's kind of stupid." It'll do a good job with that. It can control those other agents. I really like Pi for this. Like I showed earlier, I do use a pretty large amount of Pi at this point. The main reason why it's not nearly as high as something like Codex is one, I really like having a real desktop app for a lot of this work. The Codex desktop app is
- 17:12
just kind of in a league of its own as far as on machine normal desktop apps go, to where I don't really feel the need to use all that much of anything else. The only other main interface I use for this stuff is T3 Code. Obviously, like I'm close with the Oh, yeah, yeah, yeah, whatever. But, the reason why I use it is because they genuinely do have the best remote experience. T3 Code is not a normal
- 17:34
harness like a Codex or Claude Code or Pi. The agent doesn't directly run in T3 Code. T3 Code is an interface on top of Codex and Claude Code. So, if you go into my T3 Code instance, you'll notice one, it's running on a URL here and that URL is a Tailscale URL that's going into the Linux box running in my home server so that I can send off prompts whenever
- 17:54
I want to from the website and then just shut my laptop and they'll run in the background. T3 Code is the best interface for this by far and basically all my Claude Code usage here and probably like half of my Codex usage here is running through T3 Code. I should mention briefly too the subscriptions I'm using for this stuff. Right now, I have two Claude Max plans entirely because of Fable. The limits on Fable are pretty brutal and I burn
- 18:15
through one sub way too fast. I haven't burned nearly as much this week cuz I was out on vacation with family last week so I wasn't using these things as much, but on heavy work weeks, I burn through two subs very, very quickly. I only have one Codex sub just because they're so generous with the resets to the point where I just don't really need another one and the model so efficient that I haven't run into any issues. In
- 18:36
the future, I might get to a point where I will need to have a second one possibly. I have all this stuff set up through Vi Proxy. It's an open source project that I have running on one of my Linux machines and since all my devices are always Tailscaled together, it exposes an endpoint and that endpoint allows me to hit like the Claude Code API and it will under the hood load balance between these two accounts. And
- 18:56
to be clear, I am only using this with Claude Code itself. I'm not using this to stick the Claude Code sub into some other harness. You can technically do that. It's totally possible by using it. I wouldn't recommend it cuz they're very trigger-happy on banning and TOS violations. They don't want you using Claude anywhere outside of Claude code itself. But I think this here, just using it to split between two different
- 19:17
subs, so I don't have to constantly log out and log back in, is generally fine. I also like having a web page I can go to to just quickly check on my limits, cuz these do burn down pretty quickly. One last thing I did want to mention before I close is that I have started a Substack. The main reason why I'm doing this is because one, I found that for all these videos, I think things through much better in writing. I just wanted a
- 19:38
place to dump more unfiltered thoughts and breakdowns of this stuff. Like I have a companion article to this video on like the harnesses and models I'm using. This is not a replacement for videos. You do not need to go to this or subscribe to this. It is totally just supplementary extra stuff if you're interested. And I do have it behind a paywall. I really was rolling this around a lot in my head of whether or not I wanted to do that. The main reason
- 19:59
I just wanted to have it be more brain dumpy and not have to worry about it as much. I want to be super clear, this is not replacing videos. Videos are not going anywhere. I'm not going to paywall those, ever. Videos and the info and things I'm doing, I do not want to be paywalled at all. I want that all to just be out there and available for anyone to get use out of, cuz people do seem to, which I do genuinely really appreciate. It is really cool that I'm
- 20:20
able to do this and just talk about the random stuff I spend way too much time thinking about and have people enjoy it. Really do appreciate that, and I'm not going to take that away. This is just an extra place for me to dump thoughts. I'm probably going to do this for a lot of my videos. I have one in drafts for my general thoughts on open weight models right now. That's another reason why I kind of wanted to have this like a little more gated off is because I'm going to get spicier in there. Open
- 20:42
weight models are kind of becoming a politically charged topic, which really sucks. I want to avoid that like the plague, but I don't think we have much of a choice. And I really really like open weight models and local models and stuff like that. I think it's really dope. So yeah, hopefully this is kind of helpful to see where my head is on these things. Again, 5-6B Fable 5, Luna, Flash, Grok. Those are the main models
- 21:02
I'm using. They're all fantastic. Just experiment a bunch and try pushing these things harder than you think is possible. We're really getting to a point where these models can do things that are far beyond anything I imagined would be possible for years. It's accelerating, it's scary and exciting and weird at the same exact time. I don't know. I I'm just enjoying the ride. It'll be what it'll be. If you
- 21:22
enjoyed this video, make sure you like and subscribe and I'll talk to you again soon.
Description
I wanted to try and explain how I think about all the different models and harnesses I use everyday. When should you use Fable? Or GPT-5.6? Or something else?
Thank you Encore for sponsoring! Check them out at: https://davis7.link/encore
My new substack is here: https://substack.com/@davis7
My X, Discord, and more: https://davis7.sh/
