Anthropic's "dedicated monthly credit" is actually a huge cut

An important announcement I have today for anyone who's interested in using Claude programmatically. Anyone who's interested in using agents for AFK workflows for running in the background while you go and do other work. A couple of hours ago, Claude devs posted this update. Starting June 15th, paid Claude plans can claim a dedicated monthly credit for programmatic usage. This credit covers usage of Claude code agent SDK, Claude P, Claude Code, GitHub actions, and third party apps built on the agent SDK. We have heard your questions about SDK and claw P usage sharing your subscription rate limits with claw code and chat. This I think is quite a confusing announcement and it's gotten quite a baffled response. And so what I would like to do is break down exactly what it is, what they're actually saying and this thing that sounds like a bonus, why it might not actually be a bonus. We've got to start here by talking about the three things that they're actually mentioning in that tweet. The first is the Claude agent SDK, which is a TypeScript and Python SDK for running Claude, or rather it's the thing that I'm pretty sure drives Clawude code in the back end, or at least that's how they sell it. And it means that you can access Claw code by just running a script. This means that in theory, you could run that script and just get Claw Code to build something for you. The Clawed Agent SDK is used in lots of third party applications. And of course, you can use this and use any of this with API credits. So you can pay via Claude's API to use these tools. But until recently, these have also been able to be used with your subscription, too. If we move on to Claude P, Claude P is essentially the same as the agent SDK. You just call it via a CLI. For instance, I can run Claude Pello here. And after a little bit of waiting, after Claude code sort of runs in the back end here, it comes back with a response saying, "Hi, what would you like to work on?" Now, this is a very simple demo, but I can do much much smarter things with this. This is actually running Clawed Code in the back end here, or like just invisibly running an entire Clawed Code instance. It can read files, write files, run bash scripts, do all the things that claude code can do. It's just you can run it totally programmatically without you being there. This sounds like a toy, but I have built an entire software factory on this called San Castle, which is a TypeScript library for orchestrating AI coding agents in isolated sandbox. And under the hood, this uses Claude P. And finally, Claude GitHub actions, which are fairly self-explanatory. You can run Claude inside a GitHub action so that when you tag issues with at Claude, it will actually go and do work on those GitHub actions. Again, another version of programmatic usage. This whole bucket I'm going to refer to as AFK use cases, away from the keyboard use cases. So you're able to run clawed code and just step away from the keyboard and let it do its thing. I have found this is the most productive way of working with agents. And if you're interested in learning how to do this, by the way, then I do have a ton of skills for this on aihero.dev. You should check out the link below. So it's important then to separate between the AFK use cases that Claude has and the human in the loop use cases. So the ones where the human actually needs to sit there and converse with the application. So Claude.ai of course the classic one, then Claude code for working on code, then Claude co-work for sort of acting like Claude code but in a friendlier to normies way. And what Claude is doing with this announcement or Anthropic is doing is they're saying that they are prioritizing the human in the loop use cases way above the AFK ones. Before this announcement, you could use a single subscription. So Claude Pro or 5X or 20X to either fund your human in the loop use cases or your AFK use cases. In other words, you could use your subscription to run Claude hyphen P. There wasn't necessarily like it wasn't a firstass solution from Anthropic. You had to do this slight rigma roll, but they did allow you to do it. But now what they've said in their announcement is that they will keep the like same usage budget you've always had for the human loop use case here and you will get a little extra subsidy that is for AFK. Now for a lot of use cases and mine included this is an enormous cut in terms of usage limits for AFK work specifically. Now technically technically this is a bonus right because they are actually giving us something extra which our subscription didn't have before right you know that we've got this whole extra thing that has been given to us so we're technically allowed to use more anthropic more claude than we did before but for folks like me who do most of their work through AFK workloads this is a massive massive cut how much of a cut is quite hard to work out because Anthropic is a little bit shy with just how much like how much of a good deal Pro 5X and 20X Max is. Most of the estimates are from third parties and so I find it hard to trust them, but I think it's fairly reasonable, especially if you're on 5X like me or 20X, that it's often 10 times cheaper to do an 20x subscription than it is to run it through uh API credits. What I'm particularly nervous about is going and running my AFK workflows. And I will have to report back on quite how quickly they burn out. But let's just do a close reading of everything inside the um announcement just to see what's there. So the way this works is you claim a monthly credit and programmatic usage will draw from it automatically. That's just for the AFK stuff. If usage credits are turned off, usage pauses until the credit reset. Okay? So, he's basically saying, you know, you can't like take credit from one month and stick it in another month. And you get a bit of credit per plan. So, pro, you get $20, max five, you get $100, max 20, you get $200. Uh, team standard, team premium, you get the idea. And this is sort of feels like it makes sense because it's the same as you pay every month. So, it almost feels like it's a double bonus, right? You get uh a double subsidy. How interesting. But that's not quite right. So, because you get such a multiplier, such a subsidy from Anthropic versus the API credits, this feels more like a 5x or a 10x cut. I've seen sources saying that the 20x u max is one of the best deals in terms of API credits, in terms of raw uh tokens that you get. And many folks have been quoting that it's like as much as $5,000 worth of API credits a month, which is crazy. So, this does mean that these rules are clearer, right? Because it means that third party tools built on the agent SDK like Conductor and OpenClaw work with your clawed plan, but will draw from your credit in the same way your own scripts do. So, OpenClaw is considered an AFK use case, which makes sense. You'll get an email to claim your credits on June 8th, and this goes into effect on June 15th. Okay, let's check this for more information. Whoa, light mode. What the credit covers, we can see the human the loop AFK splits here. The claude agent SDK usage for your own projects, Python or TypeScript, Claude P command, the claw code GitHub actions integration, and it doesn't apply to human in the loop use cases. Claude code in the terminal or IDE, claw code conversations on the web, desktop or mobile apps, or claude co-work. So effectively because this is such a drastic cut, this means that I'm need to consider moving off claude code for a lot of my AFK work. I still can see myself using it for a lot of human in the loop work such as planning which is basically half my work. But then I have been looking at greener pastures in terms of just raw subsidies. For instance, codeex is starting to look kind of nice to me. Incredibly, I've not really tried codeex in anger yet. I've definitely had people saying it's pretty good. Uh, and it's also quite tempting to try PI now, knowing that you can just take your codec subscription and use it there because Codeex doesn't have this AFK versus human in the loop split. They just say, "Yep, you've got a subscription. You can use it for anything." We'll see how long that lasts. I don't know. I'm not a pundit, but Codeex is where the greener pastures are right now. So, I'm kind of tempted to go there and try. Now, the question that you're probably wondering most of all is, am I happy? What's my take on this? We know it's a big cut to AFK use cases, but on the whole, I'm actually weirdly happy about this. I made this tweet a few, what is it, a few weeks ago now, a month ago, saying that okay, yeah, it's I don't know what the fuss is about. Anthropics rules on subscriptions are very simple. Claw code fine. Claude's online platform, fine. Claude agent SDK running in personal software, okayish. CL agent SDK running commercial software, not okay, of course. Claw code running in CI. Um, no idea. Agent SDK running in CI. Claude P running in CI. Claw P running a personal software. That's okay. Claw P running on open source software but running on a personal computer. Don't know. Lots of weird rules were here and just so many edge cases and with a stroke. This announcement cuts through the Gordian knot there. It was a very difficult thing that the Claude devs had to figure out, I think. and they have found a way of doing it that is clear. It is clear that they don't want us or don't really care about us using Claude for AFK purposes. It is clear that they want to kind of you know do a little bit of vendor locky and pull us into the human in the loop use cases more and they want to control the amount of compute that people are spending on AFK. Now guess what? I am a teacher. I am not a pundit. I it's not my job to say whether Anthropic has done a good or a bad thing here. What I am pleased about is that they've increased the clarity, but it's also clear that I need to move off clawed code for my AFK use cases if I want to get the best bang for my buck. So, on this channel, I'm going to be experimenting more with codecs. I'm going to be trying out um I'm going to be purchasing an OpenAI subscription, which is not something I've actually had before. Uh I've been an anthropic boy for a fair while now. And I'm going to be certainly keeping Claude code because I do love it for planning and things like that. I love the human in the loop use cases and that's obviously what they want you to use it for which is fine and but I will be using codeex a bit more for the AFK stuff. So I'm excited to see how that goes. If you want to follow along too then I have a newsletter where I'm going to be posting about all of this stuff or of course you can just follow the channel, click subscribe, do all the good stuff. Thank you so much for watching and I will see you in the next