Hey Anthropic, WTF

Well, it seems like Anthropic has officially become the Apple of AI providers. They put out an update today

providers. They put out an update today which made it so that my Open Code instance, which is using Opus 45 through my Anthropic Cloud Max subscription, no longer works. Every time I try and

use it, I will get this credential is only authorized for use with Claude code and cannot be used for other API requests. And this really, really sucks.

requests. And this really, really sucks.

Not just because it breaks my Open Code setup. That sucks in and of itself. I

setup. That sucks in and of itself. I

hate that now if I want to get the good $200 a month price for Claude, which effectively gives you unlimited Opus, I have to use Claude Code, what I believe is an inferior harness to something like

Open Code. But it goes way beyond that.

Open Code. But it goes way beyond that.

I think that this is a long-term play to lock down the way agents are built going forward because of the way things are almost certainly going to change in 2026. I've been pretty convinced of this

2026. I've been pretty convinced of this for a while, and this has fully pushed me over the edge. We'll get to that in a second. I want to start with this,

second. I want to start with this, though. I put out this post when I first

though. I put out this post when I first saw it that first of all, yes, this never should have been possible in the first place. The $200 a month Claude

first place. The $200 a month Claude subscription, which effectively gives you unlimited sonnet or opus or whatever, and allows you to rack up what would be through the API, thousands of

dollars of usage for a flat $200 a month fee is kind of absurd. And if they didn't want people abusing that over API to build crazy stuff, they should never have allowed it in the first place. Cuz

the thing is they screwed up and they made it possible for me to authenticate open code with my cloud code subscription for months. I personally

been doing this since at least October and I think it was possible way before that. They never stopped it until now.

that. They never stopped it until now.

So now it feels like I'm getting rugged.

It feels like they sold me on this thing and I set up a really nice workflow for it and now that entire workflow is broken and I have to use claw code if I want to get the better price. Open code

still works with cloud. You can still do it, but you're going to be paying API usage, and API usage is a lot more expensive than this $20 a month plan.

Even the $20 a month plan is effectively subsidized. You can do $50 worth of

subsidized. You can do $50 worth of usage for $20. You are getting free money by using Claw Code instead of Open Code. This is a land grab to kill these

Code. This is a land grab to kill these other harnesses, but it goes a lot deeper than that. When I first posted this, someone responded with, "I don't really get why they're doing this because you're still paying for the Claw subscription if you're using Open Code, right? It doesn't they're still making

right? It doesn't they're still making money no matter what." But the thing is that's not the point. You're going to use Opus no matter what. It's the best option right now. But what happens in 6 months if it's not the best option? If

you build your whole system and workflow around a generic core like Open Code, that means that as soon as Gemini or OpenAI or maybe even Groc releases a better model than Opus, all I have to do

is type /models, go down here, and I can pick anything else I want. I can move over to Gemini. I can move over to OpenAI. It doesn't matter. I'm not

OpenAI. It doesn't matter. I'm not

locked in to using claude versus like imagine if I had a crazy setup in cloud code. I had a bunch of custom tools,

code. I had a bunch of custom tools, commands, hooks, all this stuff dialed in to give me this amazing app layer experience for building apps with AI.

Even if GBT 6 comes out and it benches like 10% better and it feels a little bit better like OpenAI finally gets their [ __ ] together. They give us a model with updated training data. They

make it so it doesn't reason for 500 years. It's actually fast. It actually

years. It's actually fast. It actually

calls tools well. Yeah, maybe I'd want to go use that. But if it's only a little bit better, I'm not going to go through the pain of switching out my harness to Codeex. I'm just going to stick with Opus because it's good enough. And what they'll have done there

enough. And what they'll have done there is they will have completely locked me into the anthropic ecosystem. And in

doing so, I will have become an Anthropic customer forever. Just like

this thing. I am an Apple customer forever. I've always had an iPhone. I am

forever. I've always had an iPhone. I am

on a Mac. I have an iPad. I have an Apple Watch. All of my tech is Apple.

Apple Watch. All of my tech is Apple.

And it will continue to be Apple forever because they have completely locked me into their walled garden. And that is what I fundamentally believe this is.

This is the blue text bubbles moment of the AI race. They're trying to lock you into the anthropic ecosystem so that you never ever leave, even if there's something better. You know, the the

something better. You know, the the Pixel Fold looks really really cool. I

would like to have that phone. I'd like

to use it. But then I think about, oh god, what would it be like to use Android? I'd get a worse app store. I

Android? I'd get a worse app store. I

wouldn't get the Blue Bubbles anymore, which is dumb, but it does kind of matter. People care about that stuff.

matter. People care about that stuff.

Like there are so many little pieces of friction to where I'm like, eh, even if the phone's like 10% better, I'm just going to stick with my iPhone. Who gives

a [ __ ] And Apple won. That's what they want. It is this line in the quick start

want. It is this line in the quick start guide for the Claude agents SDK package.

If you've already authenticated Claude code by running Claude in your terminal, the SDK uses that authentication automatically. And a lot of you are

automatically. And a lot of you are almost certainly wondering what is the agents SDK? Because this is very new,

agents SDK? Because this is very new, very bleeding edge. This is something that I've been betting will be huge in like 2 to 3 months from now. It's

already starting to pick up traction.

But that is using these coding agents, the open codes, the cloud codes, whatever, as the cores for new agent experiences. A great example of this is

experiences. A great example of this is something I've been working on for a little bit here, uh, BTCA, the better context app, which is effectively a wrapper around open code, where I do a bunch of stuff under the hood to load

some GitHub repos onto your machine.

Then it uses the Open Code SDK, not the Open Code CLI, but they expose an actual SDK that powers Open Code itself. They

did a really really good job architecting the whole thing where they have a client server model where the core of Open Code is actually an HTTP server that fires up on your machine.

Like when I run Open Code to start it, it starts an HTTP server. and the TUI.

And the TUI communicates with the server over HTTP, which allows them to do really cool things like make a desktop app super easily, make a TUI super easily, make a mobile app super easily,

and also build on top of it super easily with their TypeScript SDK that's fully open sourced. So, I can go and build

open sourced. So, I can go and build something around that. And when I hit enter here, it no longer works. I get

this API error where this credential is only authorized to be for use with cloud code. Cannot be blah blah blah. You've

code. Cannot be blah blah blah. You've

heard it before. this really cool new pattern of using the Open Code SDK to build crazy agents on top of it, which I think is super underexplored. There have

been some really cool attempts, one of which I'll talk about in a little bit, but it's not even remotely explored yet.

I think this is going to be huge and they have instantly cut it off at the knees. You can't do it anymore. Yeah,

knees. You can't do it anymore. Yeah,

you still can over API, but it's going to be way more expensive. You know, the reality is if you have one option that is $200 for $1,500 or $1,500 for $1,500,

which one are you going to pick? Because

guess what? Anthropic has. They have

their own alternative. This is a code snippet using Anthropic's SDK. I'm not

going to make any commentary on their SDK itself. I haven't built anything

SDK itself. I haven't built anything serious with it, so I don't have a strong opinion about it, but you can see like you have this stream. You can call a query, which would be like typing something into cloud code itself. Uh you

send a prompt, send some options, read out your messages, got type safety on that. All all like normal good stuff.

that. All all like normal good stuff.

But you'll notice in this package, I don't have any authentication set up.

There's no API key here, anything like that. And when I run it, it works. It's

that. And when I run it, it works. It's

using my cloud code subscription under the hood because that's already authenticated on this device to do this new agent. So you can imagine I could

new agent. So you can imagine I could build some really cool agent on top of it like better context and the end users can basically use it for free by just installing it on their machine already paying the cloud sub and it's done. It

works. Which takes me back to the walled garden point. If you're having people

garden point. If you're having people build these really cool agents on top of coding agents, they want to lock that down so that those super dynamic agents have to be built with the claw SDK to

get a good price because of this same exact reason. Yes, Opus 45 is the best

exact reason. Yes, Opus 45 is the best model out right now by a very wide margin. But if you've paid any attention

margin. But if you've paid any attention over the last 12 months, that typically doesn't last. Sonnet 4 was really,

doesn't last. Sonnet 4 was really, really good. And then Gemini 3 Pro came

really good. And then Gemini 3 Pro came out and everyone was hyped on that. Then

GBT 5.2 came out and everyone was hyped on that. Then Opus 45 came out. What's

on that. Then Opus 45 came out. What's

going to happen with GPT6? What's going

to happen with Gemini 4? They're making

so even if other labs catch up on the model layer, they're still going to get killed at the app layer. And that's what they're trying to do. They're trying to win application share with this change.

And it sucks. I hate it. I really,

really, really hate this change. It's a

great business decision. Whichever

executive came up with this, good job.

You're going to make a bunch of money off of it. But it really, really sucks as someone who wants to build with this.

And to give you another really good example of this, this is a project I worked on a while ago. It's what better context will turn into where I can spin up a Daytona sandbox, which they're a sandboxing company where you can just

spin up basically a compute instance in the cloud with whatever you want on it.

So I spin that up. I run this command.

I'm going to attach to that remote sandbox instance. And suddenly I have an

sandbox instance. And suddenly I have an Open Code instance which is running the Open Code SDK in the cloud connecting to it locally. And I can search through the

it locally. And I can search through the Spelt docs with this. I have those already loaded into the sandbox. And

again, you can do way crazier, cooler things than just like searching a codebase. These are early days. I should

codebase. These are early days. I should

also mention for the sake of transparency, Daytona is not sponsoring this video at all, but I am working with them and they will show up in ad spots in future videos. I just love the company. I think they're great. Their

company. I think they're great. Their

founder is awesome. They have no idea that I am mentioning them or talking about any of this. I just wanted to be super clear on that. It's something that I think we're going to hear a lot more about this year because again, these coding agents can be used for a lot more

than just writing code. It's like the Verscell AI SDK on steroids. And Daytona

is the compute primitive that I would bet a lot of people are going to be using to run these really cool new agents. It unlocks a whole new way of

agents. It unlocks a whole new way of thinking and doing things. It just

[ __ ] sucks that they are cutting us off at the knees before this can even happen. And to show you where this can

happen. And to show you where this can and almost certainly will go, uh, this is a really, really dope project that Theo sent me this morning. I hadn't even heard of, but it's sick. It's called

Claudebot, which is a personal AI assistant that can hook into any of your random things, your iMessage, your local files, your projects, whatever you want, and do a bunch of stuff with them, and

you can just text with it. It's really,

really sick. And you'll notice on their getting started page, they have subscriptions on here where you can ooth to Anthropic or you can ow to OpenAI. So

that when you're running Claudebot instead of having to pay API pricing which would get disgustingly expensive on something like this, you can just use your existing Cloud Max subscription or Codeex Pro subscription or whatever

OpenAI is calling it. It just works with this. Peter, the guy who's making it

this. Peter, the guy who's making it already responded and said that they've already found a fix and you can still use Cloudbot with the subscription right now. Uh if you're not following Peter

now. Uh if you're not following Peter and haven't seen his stuff, I'll have his Twitter link down below. He's very

very very good. You should very much be following what he's doing. But honestly,

I don't think that this really matters.

I yeah, we're going to have it fixed for now. And yeah, there are going to be

now. And yeah, there are going to be workarounds, but the reality is Anthropic has put a target on this.

They're going to make sure that anytime a workaround is discovered, it's going to get quashed. They're just going to keep patching the API over and over and over again until the game of whack-a-ole is won, and you just you can't do it

anymore. And there's going to be no way

anymore. And there's going to be no way to use this without the claw SDK. And

that [ __ ] sucks.

I I don't know, man. This is a this is a very unfortunate scenario with a lot of bizarre factors. Like if you even just

bizarre factors. Like if you even just go back to the root of this, the $200 a month subscription, which you can get $1,500 out of already, the economics on that are stupid. We're we're dealing

with weird stuff here. The pricing on AI is going to become a very real problem as this stuff eventually gets less subsidized. I'm super hyped to use Opus

subsidized. I'm super hyped to use Opus now. Everyone's super hyped to use Opus

now. Everyone's super hyped to use Opus now through the clawed subscription. But

what happens when I'm suddenly paying the actual price for it? It's still

going to be worth it to someone like me and to a lot of other people, but it's going to become a lot less accessible.

And a lot of these like really cool fun projects like Claudebot are either going to become incredibly expensive or they're going to get locked down to just working in Claude land, which is already talked about means that you lose out on

the benefits of any other lab making something cool. I know these labs are in

something cool. I know these labs are in a very very cutthroat race right now.

The reality is I don't think that there's a world where all of these companies survive. in the end, it's

companies survive. in the end, it's probably going to be like one or two of them actually make it out. And it seems like Anthropic is taking the strategy that they want to be the apple of this wave. And that's kind of where they're

wave. And that's kind of where they're positioned now. They're branding

positioned now. They're branding themselves as kind of like the artisan one. They're doing the walled garden

one. They're doing the walled garden thing. They do have the better models

thing. They do have the better models and for all Apple's problems, they do have the better products. The best

laptops in the world are the MacBook Pros right now. The M series laptops are incredible. I love them to death. I

incredible. I love them to death. I

would never use anything else. All I can really say at this point is I hope another lab can catch up and make models that can match anthropics so we actually get choice and we aren't forced at

gunpoint into building into their ecosystem which is just worse. I'm sorry

but claude code is a worse experience than open code. Use them both side by side and you will know why one is worse than the other. I'm very skeptical of Google models. I really haven't liked

Google models. I really haven't liked any of them for anything other than like document processing. Soon as you try and

document processing. Soon as you try and have Gemini do tool calls, it gets bad fast. it will randomly stop for no

fast. it will randomly stop for no reason. Open AAI has kind of [ __ ] the

reason. Open AAI has kind of [ __ ] the bed over the last year. I hope I hope hope that they are able to figure out a way to come back, but we'll see. We'll

see what happens. And until next time, if you like this video, make sure you like and subscribe. And yeah.