WEBVTT

00:00:00.000 --> 00:00:06.400
So Anthropic recently announced more changes to their billing, more specifically around Claude minus p and the agent SDK.

00:00:06.560 --> 00:02:04.590
For those of you who have been building from the hype channels out there, this is gonna be a massive problem because it's gonna change the entire way that your AI operating system works. In this video, I'm gonna show you exactly why it's a problem, then I'm gonna show you how to fix this problem so that you don't run into it ever again. Let's get into it. So at the time of me filming this, everything that you're currently doing at Claude is inside your subscription. So your Claude minus p runs and anything you're doing with the agent SDK all neatly sits inside here. But like I said, from June 15, that's all gonna be different. For those of you who don't know, Claude minus p is essentially just running Claude headlessly for you. So if you've set things up with a cron job or running something silently in the background, that is using Claude minus p. The other side of this is the agent SDK where you would import this into any app that you're building or scripts so that you can use various aspects of the Claude SDK. Both of these will now be moving to a new billing model. In terms of headless Claude, I imagine this will affect pretty much anyone who's been using Claude in an IDE or for some form of software development as well as those, like I said, who have been building crazy apps around this trying to integrate Hermes into Claude and building all of those task based apps that you see on YouTube at the moment. But then more importantly, if you have built an AI operating system based on any of the hype channels out there, you're gonna be into a much larger problem because they build those elaborate front end systems where they have named agents that run away and build things for you, and everything is clickable on a front end prototype. In another video that I made previously, I said everyone is building their AIS wrong specifically for that reason. If Anthropic ever made a change, that entire system would break, and here we are two weeks later. So for those of you who are just running a few cron jobs and a few other things, that's much simpler to solve. But But for those of you who unfortunately went and built your entire business around this, you're gonna have to do some much heavy lifting to get things going. So in a nutshell, the changes are not gonna change anything you do in your interactive parts of using Claude. So if you're using Claude in the IDE or in the desktop app, none of that stuff is gonna be changing. This is specifically for Claude minus p and for the agent SDK, and you're gonna be getting a new credit system over here depending on the plan that you're on. If you're on the pro plan, you're gonna get $20, and that obviously scales accordingly to your max plan over here ending on $200.

00:02:04.725 --> 00:02:13.045
And this is built entirely on the API rate. So this credit just goes towards that API. Once you've hit that, obviously, it will stop unless you've got extra usage credits enabled,

00:02:13.205 --> 00:02:18.485
but I wouldn't really wanna do that because this costs way more than it actually needs to compared to running it inside your subscription.

00:02:19.020 --> 00:02:57.860
So there's a different path that we need to take to fix all of this. So the first thing that I want you to do is not to panic. You just need to figure out what it is that you've currently got running, how it's running, and then measure that over a period of seven days to see how many tokens you're using, how often the skill needs to run, what sort of data it's bringing back, and what systems it's actually going into. You should already have a map of this considering you went and built the thing. But if you did install one of those plug and play things from a more hype channel, you might have to take a more ground up approach, which you can do by watching my videos at the bottom of the screen over here. It'll show you exactly how to set this thing up from scratch the proper way. In terms of measuring how many tokens you're using and figuring out a whole bunch of other things inside your environment, I definitely recommend setting up observability.

00:02:57.860 --> 00:03:28.310
So I have an entire video, again, that I'll link below. It will set all of this up for you, including this pretty dashboard so that you can see what skills are running in your environment, how long they run, and a whole bunch of other stuff related to context and security. All I'm doing here is reading the logs that are stored locally on my computer, so I'm not running any of that Claude minus p stuff within this unless we get down to this task board over here. But even when I filmed this video, I said people shouldn't be using this mission control task board because it's just unnecessarily wasting tokens. But back to the video, the whole point here is to measure exactly what is running. So if we had to go over to our activity

00:03:28.390 --> 00:03:49.690
or our skills tab, we would be able to see the cost per runs that we've had over the last thirty days based on the models that we've been using. So you can see my most hungry skill over here is the Sketchnote Illustrator, which is the one I used to draw these pretty slides for the videos that I put out there. The idea here though is that you want to track your token usage for your skills, because like I said, we still get some money from Anthropic. They're gonna give us between 20 and $200.

00:03:49.690 --> 00:04:40.185
So the important part here for measuring is understanding what runs, how often it runs, and how much it costs. Because based off of that, that depends on the actions that we're gonna take. So if you already have one of these elaborate systems and it turns out that you're actually within those credit limits, probably don't need to change anything. I still recommend that you would build it differently because, again, that is entirely elaborate and not necessary at all. But if you just had some cron jobs as well, you might fall inside this whole token limit thing. But without doing any of this measuring upfront, we cannot possibly make a decision that is actually directed from data. It would just be on FOMO and hype and worry, and those are the worst types of things to make decisions on. The other thing that we can measure under our skills and MCP tab over here is we can see how many times a skill was invoked, and that's really important as well. Like I said, it's gonna tie into how we're actually gonna be running this thing, because now that Cron is gonna start charging us and various other things are gonna start charging us, we obviously need to and should have already been using Anthropic's

00:04:40.185 --> 00:04:41.385
native functionality

00:04:41.625 --> 00:06:13.450
inside Claude Cowork. That's why they built it, and it works pretty damn well. So we've now measured everything, and we built somewhat of an audit of our environment. We understand exactly what skills are running, when they're running, and how much they cost. Now we need to decide if we actually need to move these things. If their cost and the amount of times that they run are within that billing credit that Anthropic is gonna give us, we probably don't need to do anything. If they're not, then we need to sort them into various categories that we understand the different methods that we can use to have them scheduled or to have our work running. So your first option is to use co work schedule tasks. We can just come on over here. You can create it with Claude if you need a copilot to help you do it, or you can set it up manually. And then this is pretty much just a cron job, but because it's running inside Anthropic's ecosystem, you're not gonna be charged as if you're outside running Claude minus p on a headless mode as you would if you were running it inside your IDE, which is where I run most of my stuff. Again, because I've got the $200 and the things that were scheduled inside here, they were so minor, I don't actually need to move them. But for some of the bigger things like lead gen, which only runs once a week, it does consume a lot of tokens, that will add up despite the fact that I have $200. So the simplest thing for me to do is to literally just chuck it in here in exactly the same way that I had it running before. There's not much more to change beyond that. Another option that you will have is routines. I'm not gonna go into those in-depth. I have an entire video that breaks this down step by step. I'll link that as well below. But this is the exact same thing. It just runs in Anthropix Cloud, so you have the benefit of their infrastructure taking care of this for you. You don't need to have your device always on as you would if you were running a co work scheduled task because this thing is currently running locally. They don't have any cloud functionality for this. So if you don't have a device that's always on, I would then recommend moving to routines.

00:06:13.690 --> 00:07:50.875
Something to note on routines though, depending on the size of your plan, again, you're gonna have rate limits on that. I think it's between five up to 15. That will probably change as the product matures as well. Again, that that is why it is so important to start with these first two steps over here because when you have the actual data, you know where things can go. Without that, you're just gonna be guessing, and you might be using a routine when you absolutely do not need one. So I would base my decisions on whether I've had this always on device and whether I need that cloud functionality. And then from that, I would just make my decision on where to put those automated skills. But for those of you who have a much larger problem, let's say you have one of these task board things with an AI as an employee that's doing all of the stuff in the background that probably really wasn't working in the first place, you're gonna have to rearchitect this entire thing because every single run that this agent does, it's gonna be running off of Claude minus p. This is all headless. It's all outside of Anthropic's infrastructure. It's not the end of the world because your skills and all of that stuff, it's obviously universal. You can just pop them straight into co work, and that's exactly what I recommend that you're doing. So if you are in that situation, watch the videos that I put down below about how to build an AI operating system, and it will walk you through setting things up inside here step by step with probably the exact same functionality that you've been using, and there's elaborate front ends from the other channels, but it will be way more robust, and you will never run into this problem again, specifically because Anthropic is moving towards having AI as an employee. We've seen they're bringing out Kairos, which is gonna be their always on task management agent. We just don't know when it's coming in. It was in the leaked data from a few months ago. So the signals have always been there, and that's why I've stepped away from making these really pretty front ends and things like that on this channel. Because I knew as soon as Anthropic said they weren't gonna allow third party people, OpenAI then followed and didn't even give anyone credits to run things in headless mode. You just have to pay for the API automatically,

00:07:51.035 --> 00:08:10.460
it was very clear that they're trying to lock everyone into the ecosystem. So if that doesn't sit well with you, then obviously you can go and use OpenClaw and you run an open source model or something like that. But really, Cowork is a great product, even Codex is starting to turn out to be a great product, So you're not gonna end up on the wrong path as long as you stick to the best practices that you should be doing, because in business, you want reliability, determinism,

00:08:10.460 --> 00:08:55.955
and you wanna make sure that it's just doing the same thing that you're doing every day. For instance, if you have a business that is making sales, you don't need an AI employee to adapt on the fly and stuff like that. You just need to keep doing whatever you're doing and refine the skill over time alongside the agent because there is never gonna be an AI grader that can learn something without you, the human in the loop, deciding whether what it's actually outputting is good or not. The TLDR here is that you already know what should make your business successful. So if you've taken your standard operating procedures on how you get leads, on how you deliver success to clients, all you're doing is turning them into skills. So get really good at making skills and then just schedule them. Set them up inside co work or code, and have them stay in a robust infrastructure and build it that way. It's boring, but you will never run into problems like this as you would if you follow a hype based custom approach

00:08:55.955 --> 00:09:11.070
that makes no business sense. I hope this short video was helpful. If you have any more comments, leave them down below. I will help you get through this if you need that. Otherwise, check out the videos on the screen now. They'll definitely help you on your journey, or you can check out my community where we are helping business leaders achieve success with AI every single day. Thanks very much for watch