I was wrong about GPT 5.5

Added: 2026-05-29

[00:00:00]Every other time I've talked about GBT 5.5 has been to tell people to use it on low reasoning. And I think this model is still really, really good on low reasoning. I do not take that back at all. The vast majority of the time I am still using this model on low reasoning, but I was wrong about X high reasoning.

[00:00:15]I have never been a huge fan of it on older GBT models like GBT 5.4. I pretty much entirely use that one on high. I never really changed it. This one I just recently over like the last week or so started really pushing it on extra high in the Codeex desktop app with all the crazy stuff they put in there. I It's something different. It is really really cool to the point where I felt like I kind of just had to make another video on this cuz I want to show you guys the kinds of workflows this unlocks and generally talk about where I think all of this stuff is maybe going. I've been pretty bad at predicting this stuff in the past and I'm sure I'll continue to be bad at predicting it, but I still want to talk about the weird way I've started Oh my god. monkey.

[00:00:55]He's really stupid. He's like really really stupid. Stripe, what are you doing? And I think what I was saying is that this unlocks a pretty insane number of workflows. And I think it's kind of emblematic of the direction things are going where we are just going to end up giving these agents more and more control and let them do more and more things. To start, I want to give you guys just a little example of how I generally speaking have been using these things. If I go into the Codex desktop app, and I'll just go here, use the computer use skill. What I want to do here is just have you go through and test the actual chat stream experience of this. Put in a message, have it stream in, make sure that all the formatting and content actually looks correct. And then also double check the functionality of like the retry button, the edit button, the copy button, all of that stuff. I already have the project open within Helium. You can just kind of go in there and do it. Keep going until you have all those satisfied. Give me a report at the end of the current functionality, as well as any proposed fixes to improve performance, UX, etc. So, this is one of the things I've been doing with it. It is really convenient for just general frontendy type tasks to have it go through and double check its work. I've mentioned this a bajillion times and I'll say it a bajillion times more. One of the best things you could do to get better outputs out of agents is to give them some feedback loop that is rooted in reality. Be it a schema check that makes sure its outputs are in the right shape or maybe for code stuff a check command that will actually make sure that the code compiles and is correct. And now with this, instead of just outputting a bunch of stuff that it can't really test, all it can do is make sure that the syntax is correct, but there could still be tons of logic errors hidden within this. It lets it actually go through and get feedback on whether or not this actually worked.

[00:02:19]Like one of the things I was dealing with earlier is I'm using clerk components for the org management on this project. It had these gross little boxes around all the icons, which was just a styling mistake the model made. I told it like, "Hey, this is what's happening. Use the computer use skill to make sure that's not happening anymore."

[00:02:32]And if you look closely in here, it's a little hard to see. I'll zoom in a bunch here, but uh it grabbed a screenshot of the actual current state of the site.

[00:02:40]So, it made its fixes and then went into the browser, ran the actual app, and was able to see, okay, now the boxes are gone from these two elements, which is super super nice. And as I'm doing this, you'll notice up here at the top, you look at my top bar here, I have the little Helium computer use icon. If I jump into Helium, there's a little cursor running around doing things in here. That cursor isn't me. I'm not touching the computer at all. It is just doing its thing to actually test the app where the first thing it did was a little copy test up here. So asked it for a TypeScript code snippet and it is currently trying to figure out how to send another one. Oh, there we go. It's a little slow and a little jank, but still the fact that it is just like doing this is really really cool. That's actually really funny. If you look at the output here, it was trying to do a test on the stop button and it wasn't able to click the stop button in time because it's too slow to actually take multiple actions and the model is generally pretty fast so it couldn't do it. It's currently figuring out a way to make a better one. I'm just going to let it keep doing its thing. That's not really the point. The point is that the model can now fully use the browser. And obviously, that's really useful for this kind of thing. But if you actually just think about it for a little bit, this unlocks an insane number of workflows that just weren't possible beforehand.

[00:03:46]And you know what else unlocks a bunch of stuff? Today's sponsor. Today's sponsor is a markdown file. It is the OMD file from work OS. This is a really, really cool new standard that they just introduced that makes it incredibly easy for your agents to authenticate basically anything. It's an open standard that works really well with work OS but can also be ported over to any other authentication platform because when they built this they took the standards really really seriously and it's just built on top of existing OAS specs. The way it works is seamless.

[00:04:12]Say I want to add firecrawl scraping to my app. I tell my agent to add firecall scraping to my app it'll analyze the repo and then when it goes out to try and add in firecrawl it'll find this off.md file on their servers. Get all the information it needs on how to sign up and register. Give me a yes or no on can I register on your behalf. I'll say allow. Now that that's allowed, it's going to provision everything I need.

[00:04:31]Make the account, make the API key, load it into myv, and that's it. Everything is set up and working. This has already been adopted by Cloudflare, Firecrawl, Resend, and many others. It's a dope standard, and it's just the kind of thing that works does. They deeply understand Oth everything from just adding a sign-in with Google button to your site all the way up to the crazy stuff you need to do to authenticate an enterprise customer. And there's a reason why everyone from OpenAI, Cursor, Perplexity, Verscell, FAL, so many others are using works. They are the right choice for scaling up your O and for letting agents off with your service. It's a great platform. If you're not already using them, you really should at d7.link/workos.

[00:05:06]One of the weird things that this actually helps a ton with is environment variables and shitty cloud console configs like a certain Google Cloud console. I have been getting more and more fed up with environment variables and stitching a bunch of random stuff together as I've been doing more with agents. One of the things you'll realize as you push these things further is that you want to have multiple of them running at the same time. Like as I was recording, I had that big run of codecs going in the background, but that's pretty much hogging up a my browser and b the current instance of the BTCA codebase. I want to have multiple instances running at once so that I can be doing some front-end stuff, but also be doing some backend stuff. But in order to do that, you need to have a way to parallelize these things with work trees. And work trees are good, but the problem with work trees isv files really, really suck. you can kind of copy them over a bunch, but then if you're using something like convex, you have to spin up a new convex instance for every single worksheet, otherwise they're going to conflict with each other. It's just a really hellish management problem. This doesn't solve that problem, but this does make dealing with environment variables and cloud services a little bit less painful in the short term. I'm hoping that in the long term we get to a place where like everything can actually be interacted with over API. Full disclosure, these guys have sponsored the podcast, but they are not sponsoring this video or the channel right now. Maybe they might in the future. I do really like them.

[00:06:16]But the thing I did want to talk about here is agent mail entirely because I think it's a really good example of how these things should work for agents where it gives your AI agents email inboxes. And what that means is that they can create new inboxes, update the inboxes, read threads, update threads, do whatever the hell they need to entirely over API and SDK. The most annoying thing ever is when I am working on a project, I have it working locally, it's really nice, and then I go to deploy and I need to stitch together a bunch of random environment variables and set up the preview deploys and set up the hosting and all of this random crap and stupid dashboards. Like, why can I not just run one command to actually do the thing? There are so many infra platforms that are getting close on this. Like CloudFlare's Wrangler is kind of going in this direction and it has a lot of the pieces that it needs, but still if you're going to deal with an external service, you still have to put that OpenAI API key in there somehow. And the whole thing just kind of sucks to a level that I just am tired of dealing with. The computer use thing does actually make that easier. Like one of the things I've been using it for is I've had to do a bunch of random stuff in the GCP dashboard. Not because I'm using Google Cloud, but because I need to use their APIs. I need to get into the YouTube API, the Gmail API, the Google Drive API, and a bunch of these things for internal tooling for my job because I'm building a bunch of custom agents with that. As I'm doing these things, I need to be able to properly set the scopes on the OOTH to authorize the right emails for all these things, it's just a lot of tedious clicking through stupid dashboards that I just don't want to deal with anymore.

[00:07:36]And I've honestly found that the easiest way to set up something like the Gogg CLI, which is a little project by Peter, the guy who made OpenClaw that effectively lets you grab your Gmail inbox or whatever for your agent to work with, is to just kind of let the agent go through and set up the project and credentials and all that stuff within the GCB dashboard. Copy paste that all into the places where it needs to be and then let you go through and do the actual login. Obviously, this is a security nightmare and this is another level beyond just dangerously run all permissions. I already have been doing that for a while. Like you'll notice if you go into my codecs, I've had full access turned on for the last 4 months and I have no intention of turning it off. I thought this was an insane idea 4 months ago. Like last December, I never in a million years could have imagined letting an agent have full route access to my computer at all times. But yet now I do. I just let it do its thing. I don't have any permission checks on it.

[00:08:23]And half the time I'm not watching what it's doing. Probably more than half the time. I'm just letting five or six agents run off in the background and do whatever they need to. I just kind of gave up. I trust the models enough at this point. While I'm not there yet on stuff like this where we're like dealing with very sensitive credentials in the GCB dashboard, I'm getting kind of close and I can definitely see a world where in a couple months I'm just not monitoring it anymore. I have my Mac Mini running in my server room over there. Maybe I just let that guy have full computer use all the time and if I need to do something like this, I'm like, "Hey, can you go get all this set up?" And I don't even know what it's doing. It just goes off, clicks a bunch of things in the dashboard with full root permissions and then hopefully it ends up working out. Obviously, there are a bajillion ways this can go very, very, very wrong. be it prompt injection, hallucination, or just a simple mistake where it ends up deleting something it shouldn't have. This creates a lot of potential problems, but also solves a ton of problems. And it's just it's a weird balance that we're kind of dealing with here. I don't really have the answer to yet. I'm also dealing with this with like the Hermes agent openclaw stuff. I've been going a lot harder with that. The only reason I haven't like fully gone all in on it yet is because again, the security and perms is a very big concern. The idea of giving it read rate access to the YouTube API or the Gmail API is still just kind of terrifying. There are safeguards in place for all of this.

[00:09:34]Like Hermes agent has a really good permission system right out of the box, but even still, this is what a lot of the command approval requests look like these days. This is a gigantic number of bash commands that it wants to run to execute whatever it's trying to do.

[00:09:46]Right now, I can read through this and see the RMRFs and like get an idea of what it's actually doing here and whether or not it makes sense. But it is time and effort to actually parse through and read this. And as you're just like lying in bed looking at this on your phone or like out doing something else, are you going to closely read this entire thing? As we get to the point where a lot of the ways these agents make edits is by just writing out a giant Python script and running it.

[00:10:07]It's not really parsible unless you really sit down and do it. And at that point, you might as well just like write it or have the agent write it with you at the same time. We're getting to a point where we kind of just have to let go more and let them do more. And that's uh terrifying. I'm I'm not there yet.

[00:10:20]I'm not saying I won't get there. I probably will. That's just the direction these things are going. Yeah. I I I don't have the answer. I just wanted to bring this up. Another thing worth bringing up here is the potential cost of this stuff. I have not been going 1 millionth as hard as someone like uh Peter is the creator of OpenClaw. He posted an update to Codexar which included his actual usage of OpenAI spend. He's currently working at OpenAI.

[00:10:41]So like he's getting these tokens for free. But in the last 30 days, he's done $1.3 million worth of API credit spend.

[00:10:48]He did mention in the replies that if you turn off fast mode, which is what I'm currently running it on, I'm doing 55 extra high on fast, it would go down to about 300 grand. But still, that's 300 grand a month of AI spend. He's going insane. And the whole point that he's trying to make here is he's trying to figure out what it would be like to run a product where tokens were free and unlimited. What could you actually do?

[00:11:07]He has a bajillion different sub aents running on this. Every single commit and change and issue gets run through an insane number of OpenClaw instances.

[00:11:14]Like whenever a PR lands, it will go through every single one of the issues and there are thousands of them and find every single little one that might have been related to the fixes shipped in that PR and automatically close them.

[00:11:24]Same thing with PRs. It'll automatically close them if they're duplicates or if they are currently being fixed or whatever is currently happening. It is letting the AIS autonomously manage this gigantic open source project for him.

[00:11:35]It's really cool what he's actually doing here. I think it's a sick northstar to have, but uh again, that's a lot of money. And my personal usage here, like you can see, I have 99% left, 98% left. I'm barely even touching my $200 a month sub. A lot of this is because I have the extra rate limits from the 55 party. If you didn't see that, OpenAI had like a 55 event. Anyone who applied or ended up going got 10x rate limits for a while and I'm currently still on those 10x rate limits. So like these aren't reflective of real usage. If you expand these out, my 5 hour I'd probably be down to like 90% and my 7day I'd be down to like 80%.

[00:12:09]But also like even still, I've been doing an insane amount of inference on this. Like I am definitely burning thousands of dollars worth of tokens to actually do this stuff. And at least for right now, the free money is there from OpenAI to actually do this stuff. Will it be there forever? I don't know. I really hope it will be. Enjoy it while it's here. I would go pretty hard with this. I think it is worth experimenting with because there are really cool workflows here. I also wanted to talk a little bit about the Codeex desktop app in general because I've been pushing it harder and trying to figure out the right ways to actually use this. I brought up work trees earlier and I don't have this set up on all my projects. I do have it on this one where whenever I create a new thread or whatever I can do new work tree and if I make a new work tree it'll make this work tree off of main and I'll just say we'll just do this on low. What does this project do? Just a dumb test prompt. The whole point is it's doing a bunch of stuff here where it is copying over thev.lo from my main branch. It is also setting up a new convex instance for this specific work tree that it can all just kind of run at once. And now that this is fully spun up, I have it running. It's doing its thinking. And if I went in here, I will stop that dev server. And then I will start the dev server here. So this is running within this workree. And I can open up a browser within the Codex desktop app and do a bunch of useful stuff in here. Like I'll move it so that my face isn't covering it. If I go over here, I can like take a screenshot and that'll automatically get put on my clipboard so I can paste it into the chat. Or I could be like, okay, I want to change this element, click it, and then add a comment in there that'll be passed into the next prompt. Like there's a lot of really useful things you can do within here. The Codeex desktop app has gotten really good and I'm pretty much using it for everything. I've even started using the automation system where I just have one right now. It's just keeping my developer directory clean because of what I do for work. I do a lot of experimenting and make a bunch of random stuff that ends up getting thrown away constantly. So my dev directory would just constantly get bloated with nonsense that I didn't want to have and it would be hard to parse through. So, I have an automation that every day at 5:00 p.m. it will go through and run a codec session that will clean up the directory, get rid of any stale stuff, keep notes of what I am and am not working on, save some of these memories.

[00:14:05]It's just a nice little babysitter for that directory. And that's the kind of thing you can kind of start making with this. I've also been using this a bunch on my Windows PC over there. I was hanging out with Sarah when he was in town at SF. He was showing me some of his workflows and helping me get local model stuff set up on my 5090 rig over there. And one of the things that I noticed as I was talking to him and working on this stuff is that I am not using these models enough. I need to be going harder with it. When he was getting like VLM studio, his little project for running local models. It's a really cool open source project he's been working on. The way he got it set up was just going into the Codeex desktop app and telling it to set that thing up. We didn't do any of the git clones. We didn't do any of the setup.

[00:14:40]We just had it do the thing and just went back and forth with it and asked it to make whatever changes were needed to get it running properly there. it wouldn't be enough to just blindly say these things into the void. Like the prompts he was giving it were very specific. He understood how the whole thing worked. And I think that's a common theme through all of this. If you know what you're doing and you're just telling it to take an action, it can take that action really well and do a lot of the config stuff for you. I like letting these things just kind of run computers at this point.

[00:15:05]It's one of the best uses I found for them. And that's a place where 55 on low really shines. I've even been really liking the mobile app experience. They added to ChatGpt a codec section where you can link your Codeex desktop app to your ChatGpt account so that when you're out doing something else and you're not at home on your computer, your computer can still be running and have the actual instance of Codeex running and just control the threads remotely. The experience is pretty seamless. It feels really good and I'm trying to get more and more of this stuff set up. Like my dream setup that I'm not done with yet.

[00:15:34]I'm currently working on because again the security stuff is a nightmare is having in my network closet my Mac Mini that is just kind of running forever. It has Hermes agent and other things like that on it, but I also wanted to have a codeex instance on it. So, if I want to be working on projects while I'm on the go, if I just have an idea for something, I can just send that off as a prompt and have it actually happen in the background. The other really good way of doing this, if you don't want to set up like a dedicated, longunning computer that's just always running codecs at all times. The cursor cloud agents are really, really good. I definitely have my issues with the current cursor desktop app experience.

[00:16:05]It's getting better, but it's still like glass is not where it needs to be. Their cloud agents are amazing. I really, really like them. The biggest problem with all of this though is again you have to set your projects up in a way that it is easy for the agents to work on them in parallel. That means as few dependencies as possible, as few environment variables as possible, and as simple of environments as possible.

[00:16:23]One of the things that Julius did for T3 code that I thought was really sick is he made it so that you can run the entire thing with nov files. I think you still have to have like an open router API key so that you can get the model inference to actually run that. But if you just want to boot it and look at it and sign into it, all of that can be simulated with no environment variables or external dependencies. And I think these days that's a really good thing to have not just for the development experience side of things, but also I'm really trying to work on all my projects getting the number of dependencies in the package.json as low as humanly possible. More security issues keep happening and they're going to keep happening. Supply chain attacks are getting very real and very bad very fast. I don't want to be exposed to that. I want my stuff to be a lot more secure. So, I'm going a lot harder on just instead of installing a package, letting the agent basically recreate the package within the project because they kind of just can't at this point. The way I'm using these things is just completely unrecognizable from how I was using them 3 months ago. This was kind of just an off-the- cuff rant of like what I'm currently feeling and doing with this stuff. If it was helpful, make sure to let me know. Like and subscribe.

[00:17:23]I will have more detailed breakdowns on this stuff in the future. But for now, all I would say is just push these things a little bit harder. Give 55 extra high a shot.