I’m done.

Anthropic finally responded and I couldn't be more pumped. They finally came out and support T3 code directly. That's why I got my Claude hat on and I'm sharing this update. Starting June 15th, paid Claude plans can claim a dedicated monthly credit for programmatic use. Wait, uh, credit? This credit covers the usage of the agent SDK, Claude-P, the Claude Code GitHub actions in thirdparty apps built on the agent SDK. Huh. For those who haven't kept up with this drama, unlike pretty much every other provider, Enthropic is really strict about what you can use your Cloud Code sub for. They have historically supported things built on top of the agent SDK. So, if you wanted to build, I don't know, a better guey for coding with cloud code and codecs. Mind you, an open- source guey that just calls the official Claude code bindings directly. I'm not stealing their author anything here. I'm just calling their stuff programmatically. Seems like they are not going to let us do any of that anymore. And for those saying this doesn't target me directly because I I know how you guys are. I see your comments far too often. Here's the PM for Claude Code literally directly citing tools like T3 Code and OpenClaw that you can use with these credits. There is actually a good idea tucked deeply within the mess that they have unfurled here, but I want to emphasize just how much a mess this is. Anthropic kind of backed themselves into a corner here after promising us more clarity over the last five months or so. Poor Matt PCO who just put out a course about Cloud Code. Literally sold a course to teach you how to use Cloud Code more deeply. Wanted to know how much he could package his workflows and share them with people and couldn't get an answer. Kept bugging them for an ETA. And now two months after this last ETA request, we have an answer. The answer kind of feels like the finger. I'm upset again. There are pieces here that are good and I will talk about them. But first, I have to crash out. But before I can do that, we should take a quick break for something that I don't hate so much. Today's sponsor. I have a lot of very cool sponsors that have made it possible for me to do awesome stuff. But there aren't very many that have prevented me from doing bad stuff. In fact, Code Rabbit's probably the only one that has stopped me from shipping hundreds of bugs. I am not exaggerating. The sheer volume of things Code Rabbit has prevented me and my team from doing wrong is genuinely kind of hilarious. I can spend a bunch of time telling you about how great the code reviews are, how you can just copy the feedback and paste it in your agent and get it fixed immediately, or even about how good the CLI and IDE plugins are. The CLI one's great cuz your agents can get review during their own loops. But I'm going to go a different route with this ad because I just like working with Code Rabbit and not just in the sponsor sense, in the this company cares sense. One of my followers hit me up asking me to review the sign up process of Code Rabbit because the onboarding wasn't great for him. This confused me because my onboarding was literally three clicks. So, I immediately took this tweet and sent it to the team. Usually, when I do this, a brand will respond with a bunch of excuses in Slack and then maybe they'll fix two or three bugs at best. That's not what happened here. Within literally 45 seconds of me linking the tweet in Slack, Arvin had replied to the original post asking for more information. also tagging in two other high-up product people at Code Rabbit. They care a lot, more than I would expect a company of their size and growth to do, and they're still moving fast in shipping, too. You should try out Code Rabbit for the highquality code reviews that'll keep you from shipping bugs, and you're going to stay there because it's a team that cares and makes great product. Ship better software in less time at soyb.link/codeb. Chad is telling me it's probably time to remove the hat, and I agree. Sad. So, let's go through what they shared here. First, paid Claude plans can claim a dedicated monthly credit for programmatic usage. There's a lot to dig into here. First off, claim. We'll get to that word in a bit, but the dedicated monthly credit part and the programmatic usage parts are the ones that we have to dig into. It's crazy how much layering and [ __ ] they fit into one sentence like this in order to try and make it sound good, but there's a lot of layers to dig into here. We've heard your questions about SDK and Claude-P usage, sharing your subscription rate limits with Claude Code and chat. Starting June 15th, programmatic usage gets its own dedicated budget instead. Your subscription limits don't change. They're now reserved for interactive use. How does this work? Claim the monthly credit once and programmatic usage will draw from it automatically. You have to manually claim it, which is just disgusting, but sure. When it runs out, you can keep going with usage credits, which are build at API rates when you turn them on and off. And if usage credits are turned off, usage pauses until the credit resets. If you're on the $20 tier, you get $20 a credit a month. If you're on the $100 tier, you get 100 bucks a month. And if you're on the $200 tier, you get 200 bucks a month. After you claim the credit, it resets each billing cycle. Credits do not roll over. This means that third-party tools built on the agent SDK like Conductor and OpenClaw will work with your Clawed plan, but will draw from your credit the same way your own scripts do. There is nothing you need to do today. Users will get an email on June 8th to claim their credits, and this change goes into effect on June 15th. I think this is incredibly misleading and does not really solve the problem, which sucks because there are really good ideas within this. In order to understand any of this, I have to break down the different categories of usage. If you're paying for Claude, the most simple use case is just using the claw.ai site for chat. So, if you're just going to claw.ai chat to send messages and get responses. Obviously, that's included in the subscriptions. That's what people expect to use the subs for. Makes all the sense in the world. And then on the other side, there is the opposite extreme, which is API calls. This is for something like T3 Chat where obviously they wouldn't want me to subscribe to Claude's chat and then use that token to resell to other people to use T3 chat because on T3 chat you're paying me or my company a monthly fee and then use the chat with other providers and services that we have API keys for and then we are build from them per token and then hopefully you use less than you're paying us so we can make some money. Sometimes you do, sometimes you don't. It is what it is. Obviously, API calls when you're hitting their endpoints to do weird specific programmatic stuff can and probably should be built differently from a subscription model. But there is a wide range of things between these two and that's where things start to get messy. First, we have Claude Code. Claude Code is a thing you install on your computer that you also customize, add things to, change, and use to do stuff on your machine. Generally speaking, any two Claude AI users have a very similar system prompt, very similar context beyond what they paste in, similar sets of skills and tools and plugins that the model can use. The usage here is very similar between any two users. In cloud code, that usage can vary quite a bit more. Some users might use cloud code in a way where they just have it go through local files and almost treat it like they would treat the chat site. Others use this to do really complex programming on their applications. Some people use it to automate things to like generate reports based on the poll request being merged all those types of things. People use cloud code in all sorts of different ways and the result is that the amount of money one is spending in cloud code in terms of the compute it takes to use it can vary a lot depending on the user and their use case. The last remaining piece between cloud code and API calls in my opinion is stuff like open code. Different harnesses, different tools that have their own entirely different ways of behaving. Maybe they don't cache right. Maybe they burn tokens. Maybe they just are broken and causing maybe they're just broken and cause problems for the users. One way to frame this is the amount of control anthropic has of what you're doing. When you're over here in the chat on the site, they have a lot of control. When you move to cloud code, they have less, but they still have quite a bit. When you move to open code, they have meaningfully less control. And once you're just calling the API and doing whatever you want with it, they have almost no control at all. The other issue here is how billing works across things where in the cloud chat you pay 20 bucks a month and get effectively unlimited use for chat use cases. Obviously, you can run out if you use it aggressively enough. But generally speaking, if you're paying the $20 a month and just using cloud chat and not using Opus with PDF parsing for everything, this can be relatively unlimited feeling. Where on the API call side, you're paying per request and you're paying for the tokens and it's much more expensive. The main issue with all of this is how big the gap is between what you get when you pay for the sub versus what you get when you pay for tokens directly. If you're paying for the $200 a month plan for Claude Code, for example, the amount of inference you can get for that is estimated to be as much as $5,000 per month. So you're getting a 25x cost decrease effectively if you use Cloud Code and actually max out your subscription. They also just bumped weekly limits to be 50% higher after doubling the 5-h hour limits. And that combo means that that $5,000 is probably $7,500 a month. Now, that puts you at nearly a 40x subsidization. Obviously, assuming you use all of your subscription, this means that Anthropic has to do things to keep you from using all of that 40x, especially during hours where they only have enough compute available for a certain portion of their customers. If you're burning through their GPUs during the night when nobody's really using them and they're not selling them to enterprises as aggressively, they don't care as much. But if you're doing this during working hours, they care more. So figuring out how to balance the GPUs they have available for all of these things is a meaningfully difficult challenge. They also are kind of treating this $200 a month as a marketing spend. You're paying the 200 bucks. You're getting between $5 and $5,000 of usage out of it. Hopefully, they're using GPUs they already had that weren't too busy at the time for that usage. And then you're more likely to talk about Claude, use it more aggressively, bump your inference in the amount that you're using it a ton. And then when you get used to it, you want to go use it in the workplace. And when you want to use it in the workplace, the companies don't get the same subsidization. These $200 a month plans are not the plans that enterprises get. Enterprises have to pay API costs. they might get a discount of 20 to 30%, but 20 to 30% is a lot less a discount than the 40x that you can get through the subs. So, if I'm personally subbed to the $200 a month plan, I do $2,000 a month of inference on it and then I go to work and I want to use it there and I'm used to using it in this very token hungry way, I end up spending a ton of money at my company, not even knowing it because I'm just using cloud code blindly and then Anthropic makes a bunch of money off that. So, the thing a lot of people don't realize is the subsidization isn't being subsidized by the users who barely use cloud code. It's being subsidized by the enterprises that are hiring the people who are addicted to cla code and then they go to the company and use it the same way they used it personally and go through shitloads of tokens. I know engineers at companies that are doing like 10k of inference a month and they're doing that through the API, not through these subsidized plans. Anthropic makes a killing off of that. That's why their growth has been so nuts. That's why as of recent reporting, Enthropic's revenue is actually higher than OpenAIs right now because people are just burning tokens on the enterprise side. So, what does that have to do with what we're talking about today? Well, there's a hard line here between Cloud Code and Open Code. A problem happened before, and you've probably heard about this if you watch my channel much, where Open Code previously had a way that you could grab your token from Cloud Code once you pay for the $200 a month plan and use it in Open Code instead. Why would you want to do it? Well, clad code's the worst harness you can use anthropic models in. All of the others perform meaningfully better. Cursor, in particular, performs much better with Opus models than Cloud Code does. Open code's also a much better user experience. It's much less buggy. It's just nicer overall. It has support for other models and providers. Open code also is open- source unlike cloud code. So, what's the problem? Anthropic wasn't getting what they wanted when people took their subs and brought them to open code. They weren't getting the lock in. They weren't getting people addicted to the loop and using cloud code in such a way that they burned through lots of tokens. They also probably saw that open code users both used their subscriptions more and probably cost more money on average and importantly they probably weren't caching as well. So the average request likely cost more in open code than it did in cloud code. A lot of cloud code subs who were on open code started complaining about how quickly they were hitting their rate limits. And this is, I think, the genesis moment where Enthropic started locking out external players. By the way, Anthropic's caching style is also [ __ ] DAX did some tests in open code with tool call pruning. What this means is when a tool call happens, cleaning up some of the noise afterwards so that you don't have all of this bloat in your context window anymore. The issue is if I remove some responses from a tool call from four messages ago, that breaks the cache cuz the cache is up to a certain point in chat. So if you edit things from earlier, the cash breaks, which should make everything slower and more expensive, right? Well, it turns out cleaning up the tool call history actually ends up being cheaper overall because you have less bloat in your context. You also end up hitting caches more often because you don't have this massive cache that has to be managed and compressed constantly because you're cleaning it up as you go. So even though DAX's implementation in open code breaks cash more, it ends up being cheaper and faster. So, it's [ __ ] But here is where I have to reveal something. I kind of lied when I broke things up like this because there's another thing between Claude Code and Open Code that is right on this line. The agent SDK. When Anthropic built Claude Code, they realized how powerful it was and that people were going to want to use it to do things in their code directly, not just to use it in the terminal to write code. They might want to programmatically use it. Maybe you have an API endpoint that gets hit when a pull request is filed and it will do a review of it. Maybe you have a bot that you set up in Slack that will go use cloud code to update something when you ping it about some [ __ ] It's a very common desire for programmers to programmatically do a thing they're already doing. And the point of the agent SDK was to take a lot of the power and capabilities of cloud code and give you the ability to use that programmatically. You can assign the different tools it has access to. You can write prompts. You get back responses. You can create threads. You do all the things you would normally do in Cloud Code, but with code instead of prompts. Historically, the Claude Code team has said it's okay to use your Claude Code subscription for the agent SDK. Sadly, we never got a good answer as to how much we could use that. They even had a section here in the docs where they describe how to authenticate the Claude agent SDK, suggesting that you can use API keys for all these different services. But then this little call out here, unless previously approved, Anthropic does not allow third-party devs to offer claude.AI login or rate limits for their product, including agents built on the agent SDK. Please use the API key authentication methods described in the document instead. But this has a lot of that vague wording that makes anthropic so annoying to deal with. third party devs. Sure. But what if like I make a package that people are installing themselves into their stuff that they're building around the agent SDK? Because the agent SDK allowed you to use your cloud code sub for personal use. They were very explicit about this before, at least on Twitter. Thor even said as much as of February this year. We want to encourage local development and experimentation within the agent SDK and cloud-p. If you're not familiar with cloud-p, it's how you programmatically use the cloud CLI instead of having the fancy interface that takes over your whole screen where normally when you open cloud, it does this and it's the fancy UI. Cloud-p lets you programmatically write a prompt and then it will respond in line like a normal CLI app does. You would think this isn't that different, but you wouldn't be thinking like anthropic if you thought so. But as I mentioned before, they're very explicit that they want to encourage local development and experimentation with the agent SDK and cloud-p. If you're building a business on top of agent SDK, you should use an API key instead. This distinction was annoying, especially for things like open source apps you run locally on your own machine. And this is what we've been trying to get clarification on for a while. Everyone from Matt PCO to Jose Valim, the creator of Elixir, to myself, has been begging for an answer to this question for over half a year at this point. And this was almost an answer. We got a slightly better almost answer from Boris when someone asked directly. Subs can be used for personal local tools that use or wrap cloud harness with like cloud code and cloud code headless and agent SDK. Right. Yep. Working on improving clarity here to make it more explicit in April, a bit over a month ago. I even was nice here. I know I give you guys a lot of [ __ ] but I hope you know how much we appreciate the clarity here. Thank you, man. Just replied now. I take it back. I know I cite this Matt PCO tweet a lot, but it's so good. I don't know what the fuss is about. Anthropic's rules and using subscriptions are very simple. Claude code, okay. Claude's online platform, okay. Agent SDK running in personal software, okay. Is agent SDK running in commercial software? Not okay. Cloud code running in CI, we don't know. Maybe it's not so simple. What about the agent SDK in CI? What about cla P in CI? What about claw-p running a personal software? Okay, as far as we know, but quad-P running an open- source software that is run on my personal computer, no idea. Claw-P running on distributed sandbox is kicked off by me. yada yada, you get the idea. And Matt clarified at the end here, let me be clear. I have never before experienced from any developer tool such a frustrating lack of clarity over the basic terms of usage. Remember, Matt PCO released and sold a Claude Code course. He dedicated months of his life to supporting Anthropic and making it easier to use Claude Code heavily, the exact things they want. Matt has probably personally brought millions of dollars of revenue to Anthropic. And they stopped acknowledging him because of how hard he was trying to just get an answer to this basic question. And instead of an answer, we got Anthropic glazing themselves, pretending they're being so generous as to give us this additional credit. And here's the problem. This spectrum is so much wider than I've even made it out to here because if you're using cla code interactively where you just type claude in your terminal and use it that way, that is one thing, but claude-p is a different thing. You're still using cla code, so you're still getting all the benefits of their caching their everything else. As SDK is more directly programmatic, open code is a different tool entirely that they don't like or support. And then API calls you can use for whatever you want. I would also say that open claw fits in here too. There's a rough idea of the spectrum of different uses for a tool like claude and your claude subscription. And the further you go from the left to the right here, the less they want you using your subscription for that use case. There are so many other things I want to crash out about here. For example, the fact that the agent SDK is so [ __ ] terrible and doesn't support all of the standards everything else does and is closed source for no good reason. Obnoxious. There are like a dozen better AI SDKs you can use in TypeScript that I can think of off the top of my head. Seriously, just go use the PI SDK or maybe use the Versell AI SDK. It's pretty good, too. There's always new ones coming out that are fine. I know the Tanstack one's pretty solid. If you want to do client side stuff for the like UI and React, you have a hundred better options than the Asian SDK. I'd go as far as to say the only reason to use the agent SDK was because you could use your cloud code sub with it. We have so much code in T3 code built around this because it was the only way that they told us we would be allowed to use your subscription in local software that is literally just calling Claude code. I'm going to give them the benefit of the doubt here and try to steel man this a little bit. Realistically speaking, you do have to draw a line somewhere in this spectrum. If I need to sell API calls for an amount of money, but I also want to subsidize them for chat usage for various reasons, you have to choose where the subsidization stops. I would argue the right place to draw the line is here. Once you're doing programmatic API calls in enterprise use cases, you should be paying the API rates. They are much too high, but I know why they're doing it. They have to make revenue targets happen in order to succeed as a business. I get it. But you have to draw this line somewhere. The line is where does your subscription stop benefiting you? Where does your subscription stop being a solution? And since their intent with the subscription is to get you into cloud code and claude chat, it makes sense that they would want to cut you off past the things that they built. Especially with their previous justification, which was that external tools like Open Code and Open Claw caused absurd amounts of usage and didn't cache properly. If one prompt in claude costs $5 of inference and the same prompt in open code costs $15, that makes sense. And considering that OpenClaw runs a cron job that is outside of the cache window, it just burns money. On my personal Open Claw that I'm barely even using, it does over $5 a day of inference. Just hitting Opus 47 asking, "Do I need to do anything?" No. Cool. Just running constantly throughout the day. So, I get why they would want to cut that off because if half of the people on the $200 plan are using OpenClaw, they are just burning money for Anthropic and they don't want to do that. I understand. I think the solution would be lowering the rate limits, but I get why they don't want to do that either. Whatever. Realistically, Anthropic is really okay with you burning a shitload of their compute and money as long as you're doing it in cloud code and the cloud site. They get less okay as you go down the path because they banned explicit usage of your O externally as in I use open code. I hit the same endpoint that cla code hits, but I'm using the token from cloud code in open code. I get why they would ban that. It's not great, but they did it because they did that. And their reasoning for it is that we weren't using the building blocks they had made for cloud code. I get why they would ban that. So, a lot of us ended up building on top of cloud-p agent SDK as a result. This includes things like zed, things like gene, things like T3 code, things like all the cool CI integrations people have done. There was even an attempt to make this work with open code. It's also a really good integration for open claw as well as for Hermes agent. All of these are either using cloud-p or agent SDK in order to use the official anthropic provided harness with your official anthropic subscription. And this is obnoxious. Every other provider that we support in tools like T3 code has a much happier standardized integration path, but we had to build everything ourselves for claude because there was no other way to let you use your subscription. That was one of the goals that we had with T3 code. This should solve all their problems, right? Like this solves their problem with caching. This solves their problem with us not using their tools. We are building heavily around their closed source [ __ ] In fact, the the only closed source dependency that's a meaningful portion of the T3 code codebase is the [ __ ] agent SDK. So, this all should be good and dandy. We've just been waiting for clarification for them to give us the formal thumbs up that this is okay. They certainly put a finger up to us and it's not their thumb sadly because the change they announced is [ __ ] What they have effectively done is draw another line here. The sub is okay. If you're using the cloud site or you're using cloud code in the standard like interface like this or in their desktop app, your subscription works. If you're inside of this section, the sub used to work, but they would do hacky things to prevent it from working if you were using tools they didn't like. For example, if you were using OpenClaw or Hermes agent with your sub by calling Claude-P, which is how these were implemented, OpenClaw would work by using the Claude CLI as a workaround because you couldn't call the API directly. So they started looking for specific words inside of the prompt and inside of the system prompt that would suggest that you're using OpenClaw or Hermes agent and then would block the traffic or charge you for it instead. The implementation of this was so cringe that I had a viral video about it because it would literally trigger if you had commit messages that mentioned OpenClaw or Hermes in certain ways. Genuinely absurd that certain file names being mentioned in commits was enough for this to break and then bill you money instead of giving you your free inference over your subscription. So what they've done instead which and this is why I am annoyed previously there was kind of a gap between these where tools like zed gene t3 code open code and CI integrations all of these were honoring the original intent of cloud code we were using it to code we were just providing different interfaces for it but then there were things like openclaw and hermes agent which are meaningfully different from what you're supposed to do with claude code that also automatically burn tokens these things are token farms that just burn tokens constantly ly these things are UI wrappers on top of claw code that are just different ways to use it. So you would imagine that Anthropic's goal here would be to draw a cleaner line between these to separate the things that just waste tokens from the things that are just trying to provide better UIs. Like I built T3 code because I wanted to use Cloudmore, but I hated using Claude code because it was falling the [ __ ] apart and their desktop app was even worse. So their solution is they now give you a limited credit for everything here. So uh so we got [ __ ] They drew a really [ __ ] distinction here that I am genuinely mad at them for. They refer to the difference here as programmatic and interactive use. Interactive is being inside of the terminal, talking to it, asking questions, and being there when it does things. Programmatic use is writing code or automating the process of calling things. I don't know about you guys, but I kind of don't see T3 Code as a programmatic usage. T3 Code is just a different interface to do the same thing. So, if you use Claude Code through the Claude Code terminal on the $200 plan, you get between $5,000 and $7500 a month of usage, but if you switch over to T3 Code, you get $200. At the very least, I can now move off of their [ __ ] agent SDK and Claude-P stuff, right? Because they're putting this bucket here. Nope. because everything past this line is still API billing directly. So they now have three tiers of usage. If the UI on your screen is from Anthropic, you can use your subscription. If the UI on your screen is powered by Anthropic, but you wrote your own custom layer, even if you're just writing the prompt without using their UI, you just dash P instead, you get a limited usage bucket that is this new $200 thing that you can pull from instead. And if you're doing any of the other things, like if you want to move off agent SDK and use something competent, maybe you want to move over to open code or something instead, nope. You can't even use those credits anymore. You're paying API prices. So people building in the middle here, which reminder are people who love cloud code and want to support it. They want to make these models and these subscriptions and these things accessible to more people because believe it or not, this interface is not the most accessible thing. If you're blind or have specific physical disabilities, maybe you need higher contrast or a UI where you can actually paste an image and have it [ __ ] work. Maybe you just want copy paste that functions. Maybe you have any of a billion reasons why you don't want to use a broken shitty terminal app and you would like to use a web app that functions to do the same thing. Well, I hope you have 40 times more money to spend because that's your only choice. I said I was going to try and steal man. So, let me do that. Let's imagine a world where the only things you have are cla chat, claude code, not even the interactive and API calls. Imagine these are the only things you have. You buy the sub so you can use cla chat. You start playing with cloud code and enjoying it. And now you want to do other things. Maybe you want to set up a chatbot or want to set up an automation that will trigger when PRs get filed on your project. Anthropic doesn't want you to do API calls using your subscription because that's a different product, different thing. Makes sense. Imagine you run a restaurant that serves food, but you also own the distribution for food that you sell to other restaurants. You have an all you can eat buffet that reasonably speaking, people come in and eat as much food as they can. Some eat a little more than they spent the money on, but most don't. But if somebody who runs a restaurant comes in and collects as much food as they possibly can on that one flat fee they paid for the all you can eat and then they go resell that food, not good. Makes sense economically why they can't do that. The API calls are what they sell to other restaurants. Cloud code is the all you can eat inside of the restaurant. The interesting thing that it feels like they're trying to do here is as part of your subscription give you some credit to do the API call thing with. So again, with the restaurant example, if you go to this all you can eat buffet, maybe they give you a voucher where you can get $200 of free food from their distribution that you can then get for whatever use case you want. This is actually kind of cool. If you start in Claude chat, you realize Claude Code could potentially be used for building stuff. So you use it and you build some stuff and then you realize, oh, maybe I can use Claude over the API for some things, too. And then you go do that. That's actually kind of cool. If I start using Claude code and then I want to expose something for others or programmatically build something that manages my PRs on GitHub or whatever, having a $200 a month bucket that I can use for [ __ ] around with things that aren't clawed code itself, but as a gift almost as part of my sub to go experiment with this other thing. That's kind of cool. No one else does this, by the way. Nobody has a plan where I pay for the monthly fee to use the platform, but then I can go hit the API to go experiment and do other stuff. I think you kind of can do this with open code, but it's definitely not a blessed path. I like this idea. The idea that I can use my cloud code sub to build a thing that uses claude and have a little bit of usage that I'm getting as part of my subscription to test it, try it, see if it's useful, and then eventually get upsold to be a traditional API customer paying API rates. That all seems fine. And if this is the goal, I think it's kind of cool. I even said as much to Lydia here. What she she framed this as the new bucket means you'll be able to build on claude without needing an API key. Everything running on the agent SDK will instead draw from a new monthly credit separate from your subscription limits. And as I said here, I do actually like the framing. Thinking of this as an accessible way to build new things to compliment cloud code and experiment is cool. Totally valid for new ideas like CI integrations, programmatic audits, etc. I don't love it for cloud code rappers and UIs though. So that's the first complication. This new bucket that they're making for people to experiment and build cool [ __ ] with is not being used that way. It's being used to limit realistic things that are just cloud code rappers. That's already pretty dumb. But where it gets even dumber is you can't [ __ ] use the free API calls. You can only use the credit for other things Anthropic made. You cannot use the $200 a month in code you wrote. You have to use it on top of code they wrote programmatically and you have to go manually claim this in the dashboard too which just is the icing on the [ __ ] [ __ ] cake. The way they could have done this that would have been less cringe is if they moved the limited credits for everything over to be for stuff like open code, open claw and API calls and they had a little bucket I could play with for those things. And then when I use the Asian SDK and -p, it just goes towards the credit here. And maybe if they notice OpenClaw and Hermes are using too much usage, they will start to throttle you when you're using those things or be more aggressive with the limits for that. Don't block them or bill them, but be more aggressive with limiting these users if you're concerned about it. But that's not what they did here. They drew their line far too early. And instead of having two clear tiers, they have three vague ones. Somebody asked in chat, "Isn't cla-p just the API with extra steps?" No. The API is a thing you hit with fetch and define everything around. Claude-P is using all of Cloud code, all of its system prompt, all of its skills, all of its harness, all of its everything that it has, its permission system, and all their [ __ ] half of which doesn't even [ __ ] work. But you're calling it in the CLI with a prompt passed to it directly instead of in their CLI UI [ __ ] Is the cloud ACP also affected? There's no such [ __ ] thing as the Claude ACP. You can use Claude over an ACP binding that is built on top of the agent SDK. So again, if you're not using Claude code inside of Claude's UI if the thing you're looking at wasn't made by Anthropic, you are being build and you are being scammed. That is just a [ __ ] fact. There is no in between. There is no workaround. There will be all sorts of temporary hacks. Somebody made a fake Claude-P that will behind the scenes run something like Zmox or T-Mox in order to abuse normal Claude to pass it a prompt and give you a response in the end. But this is going to get patched. All these things will be patched. And if you try to work around the APIs or something, good luck. I'm pretty sure the reason that they are delaying this change till June 15th is they're going to start offuscating the endpoint even more. So if you try to hit whatever API endpoint they're hitting in cloud code, you're not going to be able to interpret the results because it's going to be some encrypted obuscated [ __ ] and there's going to be a long ass cat-and- mouse game around that. But by giving us a month lead time here, they are padding people getting on to the latest version so that if you're on an old one, you'll just get an error saying make sure you update to the latest so that it's harder to reverse engineer and like build something else on top of. They are clearly explicitly banning all use cases that don't involve you using a cloud code UI that you can take a screenshot of and post on Twitter. They just don't want to support anything else anymore. My honest guess on what happened here is they had this idea of a bucket you got alongside your sub to use for things like API requests to play with using Claude in the things you built. And that's an honestly novel good, noble effort. I think that would have been really cool if they just gave you an API key you could use for whatever you want with a small but reasonable limit on it as part of your subscription. That'd be [ __ ] awesome. That'd be really cool. And more people would build more apps with Claude. It would get them all the things they want and it would be really cool. But what happens at a company that is addicted to inshitification is a good idea gets proposed and by the time it actually happens, it is just being used as yet another [ __ ] bludgeoning stick. This could have been really good, but instead it feels like a middle finger. And I feel like we've been punished for trusting anthropic that if we build around the agent SDK, we can continue supporting cloud code users. But we can't. The reality is that anyone using cloud code in things like zed or t3 code or gene or any of these other awesome harness wrapper UI tools went from getting thousands of dollars a month in inference to $200 a month. Your ability to use your sub and T3 code has been cut by 25 to 40x as a result of this change after we already put all the work in building a special happy path in an open- source app that you run on your computer. We did everything they asked and they responded with the finger. And I can't actually believe they did this. I did not think this is how they would go. Don't worry though, I'm working on a solution. If Anthropic won't let us make a better UI for cloud models, we will use their shitty UI instead. What we're going to have to do is add an option in T3 code where if you want the good UI, you get the reduced limits. And if you want the higher limits, you get their shitty [ __ ] terminal. If this is what you want, Anthropic, this is what you get. We will replace our carefully crafted, slaved over UI built on top of your [ __ ] SDK that no one should have to use. And we'll replace it with your shitty [ __ ] terminal app that you spent way too much money gimping because you hate developers who like your [ __ ] We're going to have to ship this soon. It'll certainly be out by the time this change goes live in June. But I hate that we have to provide a worse experience to our users because Anthropic is such a shitty evil [ __ ] company. They took what could have been a good idea, framed it like it was a kind gift of them when they're actually giving the finger to everyone who cares the most about building great experiences with Claude. I can't help but feel like they're punishing me for being angry that the Claude desktop app is bad because the Claude desktop app is so bad. Apparently, some people don't understand the fact that this change is a hard ban for things like ACP and integrations and other tools. Agent client protocol is really cool and is supported in most things. Anthropic has never supported it. All of the use cases for ACP are tools that we in the community built. Zed built the Claude ACP adapter and they built it on top of, you guessed it, the Claude agent SDK, which is now not going to use your subscription limits. So tools like Zed are explicitly screwed by the same thing, which is why tools like Zed are copying what I'm doing here with the terminal and doing it themselves in their editor as well. We're all equally screwed here. If you don't see the stupid Claude logo when you open it up if you're not in a UI that you don't see the source code of, which is another funny way of putting this, you can no longer use Cloud Code subs on anything open source at all. You just cannot do it. It's not allowed. If you're using your Cloud Code sub, you are using their closed source software explicitly. As soon as one layer is between cloud code and what you're doing, you can't use your sub anymore. And they're patting this out, pretending it's a nice thing that they're doing. They're being so generous. They hurt us with our questions. And that's why they are giving us this new generous credit. It's so [ __ ] stupid. I know how you guys are. You make all sorts of theories about how we can solve this problem. Just make a virtual terminal or hack this in other ways. They're going to patch all that. They're going to be aggressive. The framing I'm going to put into this and I I will stand behind this. In its current state, all of this is an attack on open source. If you're building open- source things in, around, and on top of cloud code, you have to pay 40 times more. Closed source use cases on top of their closed source [ __ ] subsidized. As soon as you build anything on top, you're paying 40x more. Lydia drew out this diagram to make it easier for people to understand. They split it up as interactive and programmatic. Interactive meaning you're staring at cloud codes closed source [ __ ] These are using your unchanged subscription limits. The programmatic use. So anything on top of agent SDK or just using -p with claude. How crazy is it? Calling cloud with -p instead of calling it directly now cuts your usage by 40x. Chad's confused. Wait, so just using -p in my own terminal, I'll be affected. Yep. Yep. Great call out from Calvin here. If you're using cloud code interactively, you're not working with AI the way we should be in the future. Absolutely. They want to own the interface. Good call out from Ma here. Keep in mind the claiming system is so ass that if you temporarily don't have a plan, you just can't. I have been a pro subscriber since the pro subscription had Claude code. Once there was extra credits I could claim due to an issue, but because they couldn't charge his card once due to a random bank issue, his subscription stopped. And if the subscription fails to pay once, they just wipe your credit and they won't give you a single day of grace. Yep. Yep. I'm going to go over some fun things that are now banned as a result of this change. Familiar with Project Glass Wing? A big part of how it works is the agent will start on different files in your codebase programmatically and say, "Look for security issues. Start here and explore." Since that's done programmatically, probably with cloud-prite where Jared is programmatically going through every file and every directory inside of all of bun, rewriting it in Rust from Zigg. Done. Wait, can you hear it programmatically? You can't do this with your sub anymore. He can because he works there and he gets unlimited inference anyways, but you can't. Do you want to set up quad code in your CI on GitLab or GitHub so that it automatically runs when things happen? Guess what? You use the word automatically. That's banned now, too. They haven't split AFK and human in the loop because something like T3 code is human in the loop. This is not automatic. And there attempts to break up between automatic and in the loop, they [ __ ] everybody who's building in the loop tools on top. That's a really good question. If you ask cla code to run claude-P, will you get build? Yep. If you ask other agents, too, because something I've actually done a lot is I'll use something like codeex to make changes and then it'll tell it, hey, run cla, tell it to review your code and address the feedback it gives. That cost you money. Now, this is a good point from knocking. They did this because they couldn't find out a good reliable way to detect the difference between something like OpenClaw and Hermes versus something like T3 Code. They couldn't. So instead, they're just going to ban all of us and charge us all for daring to make a better UI. You notice that I'm more pissed than usual. It's cuz I put a lot of effort into this. It was my understanding from all the conversations I've had and all the coms they've put out publicly that building on top of the agent SDK and an open source piece of software that was just an alternative UI for cloud code was totally fine. and supported and I got [ __ ] beat down from them for it. We put so much effort in. It was obnoxious getting Asian SDK to fit into a competent system because it's incompetent software. And I'm not going to mince my words anymore. The engineers building these things at Anthropic either suck at their jobs or have their hands tied behind their back because every piece of software they produce is [ __ ] trash. All of them. 100% of the time. And after putting way too much effort in with my team to build something good around their [ __ ] so that you guys, the users who are paying them, can have a better experience than they give you. We built this for free and open source. There is no way to pay us for T3 Code. This is a free piece of software. T3 Code has made more money for Anthropic and OpenAI than it is made for us. It's made way more for Open Code as well. We did this because we wanted a better experience for people who used Claude and used Claude models. Our reward for that is having to build a shitty terminal app that breaks half the time because their [ __ ] CLI is so bad. Like this is what we have to deal with now because Anthropic, to put it frankly, just kind of sucks. If you're using cloud code the way it exists today, you're not using it the way it's meant to be used in the future. What's the correct way? Interfaces that aren't so hands-on. Things like a bot that you can text to make changes. Things like a Slack bot that monitors when you're having a discussion and prompts you saying, "Hey, maybe this is a useful thing we should do." UIs that aren't so involving of you and the work. If someone uses the cloud primarily through the CLI or desktop app with less than $200 worth of cloud-p usage, they now get a free extra $200 a month. Yeah, if they're hitting the max on their other stuff. Yeah, this is an additional bucket. That's how they framed it. And they focused so much on that that I didn't want to give them that. You can still try code with cloud code right now. It'll still work totally fine till the 15th and then you have to use it through a shitty terminal instead. I just posted this because I want it to be more widely broadcasted. I can't help but feel personally burned by the cloud code changes that were announced today. We put so much work into wrapping the atrocious claent SDK in T3 code. It was the only path they supported. So, we made it work. It was hell. To be clear, they told us this path was supported. And now, as a result, our users are getting their rate limits cut by 40x despite us doing everything right. I listened to the Cloud Code team. I had my issues with their direction, but I trusted them and took them at their word. I will never make that mistake again. Until we see significant change, it is safe to assume any statement from an enthropic employee is a lie on a timer. The rug will be pulled no matter how many promises are made beforehand. Two last things before we end. First, I would like to remind you that there's lots of other options that are nowhere near as evil and rug pulley. Within T3 code, we support claude and codecs and cursor and open code with integrations for things like Gemini and GitHub coming in the with integrations for things like Gemini and GitHub copilot coming in the future as well as hopefully the full ACP registry. So every ACP compliant thing will be supported. So you don't have to have a CL or so there's still plenty of uses. Honestly, most of our users and most of my usage is with codecs anyways. 5.5 on low fast is so good by the way. Definitely worth trying. All of these are totally fine and support us and are nice to work with and actually reliable. So, you should consider using other things. And I'll end with Matt Poc's statement. This is the clarity that we've been crying out for, but it's a poisoned chalice. It's a 10x, no, it's a 40x cut to Claude-P disguised as a monthly bonus. Anthropic is discouraging any kind of programmatic usage, and that's fine. No subsidy lasts forever. But it's time to finally try Codeex. I think it's finally time. I held strong through all the other [ __ ] but I'm done. My plan has been cancelled. They had the opportunity to do this right. They chose not to. I don't know how better to end this than with a Codeex hoodie and a GPU poor hat. So, that's what I'm going to do. Sorry for the crash out. I haven't had a company screw me like this in a very long time, and I'm far from the only one being screwed. Sorry to anybody who trusted Anthropic to support you building on top of them. That was a mistake we all fell for. That the company whose whole brand is being the nice company that loves to make things for everyone that are safe and reliable is making things that are untrustworthy and terrible. I wish it was better, but I don't think it can be. We have now seen the true I don't know what else to say. I'm done. Peace nerds, I guess. Cool.