It is official. Anthropic has rolled out and announced new rate limits, basically how much you can use the AI platform because of a specific tool and because of a specific type of person potentially abusing their whole system.
Now, today on the podcast, I want to talk about the Anthropic rate limits, what this means for other AI companies, because I don't think Anthropic's going to be the only one that has to do this.
I'm going to talk about why they have these rate limits, what the rate limits are, where I think this is going in the future, and Are rate limits even going to be a thing in the future?
Or are we eventually just going to get to unlimited free compute?
All right, we're going to get into all of that on the podcast today.
Before we do, I wanted to mention, if you've ever wanted to use all of the top AI models that I mentioned on the podcast, I'd love for you to try out my platform, which is called AIbox.ai.
Essentially, it allows you to try the top four 40 different AI models all on one platform, Anthropic, Google, Gemini, Rock, OpenAI, everything for just $20 a month.
So you don't have to have subscriptions to every single platform.
There's also the ability for you to compare results side by side from different AI models and the ability for you to have a media storage that shows you everything you've ever generated in the past, the prompt you used to generate it and help you get back to the prompt and the chat that you had to begin with.
Okay. So that's cool, but let's get into what is going on with Anthropic starting with a tweet that they put out, um, that says, quote, we're rolling out new fresh or new weekly rate limits for Claude pro and max in the late of August.
So it's not right now, but coming soon, we estimate you'll they'll apply to less than 5% of subscribers based on current usage.
Basically the big thing that I think is important is it's like, okay, look, This is only going to apply to like 5% of our subscribers based off of what they're seeing.
The other thing that I do think is important is as more and more people are trying this, more and more people are getting familiar with these tools.
I think like give it a month or two months and that number would have been 10% or 15% or 20%, right?
Because people are using it more and more.
And so I think that maybe it's 5% of subscribers today, but like by the time this rolls out at the end of the month, it could be impacting a little bit more.
Anyways, all this to say the 5% I don't think is.
As small as you'd think, it'll get bigger.
But anyways, this is what they said specifically.
Claude Code. So Claude Code, for those who are unaware, is basically Anthropic's coding platform.
It ties into your code base. I use it at AI Box.
It ties into our back end and our front end.
We tell it how, you know, we would like it to code something.
It looks at all of our code precedent. So it designs things that look similar to our current design styles.
We basically, if I'm being 100% honest, have replaced our... designer over AI Box with this because it's able to help a lot with the code.
Now you do need, in my opinion, a talented coder to help give it directions and to understand what's going on.
You have to understand the code base you're prompting against the back end and the front end and everything else.
But basically, if you understand what you're doing, you can get a 10x on what you're able to do with Cloud Code. as a developer.
So they said that cloud code has seen unprecedented demand.
They said, especially as part of our max plans, basically they're, so this is the other thing with it.
I was spending over at AI box. We were spending like a hundred dollars every couple of days because we're just using, getting so much usage out of cloud code.
And then they unveiled this thing called the max plan, which was basically unlimited usage for $200 a month.
Who could say no? We're already spending so much money every few days that we're like, oh yeah, this is a no-brainer.
So we sign up for that. It's kind of tricky.
You actually have to message, or I think they're like support, or you got to do something, jump through a couple hoops.
But basically, once you get it, it saved us a ton of money.
Now, evidently, all the money we saved, they're not happy about.
And they said, we'll continue to support this growth while we work on making Cloud Code even better.
But for now, we need to make some changes.
Some of the biggest Cloud Code fans are running it continuously in the back end 24-7.
All right, I have an example of one person that is doing this.
They said... These are remarkable and we want to enable them, but few outlying cases are very costly to support.
For example, one user consumed tens of thousands of model usage on a $200 plan, right?
So you're spending 30 grand, getting 30 grand worth of compute for $20.
I thought, You know, we're probably saving a couple thousand dollars a month going from 2000 to 200.
Now I'm thrilled. Well, it sounds like there's people going from 20,000 or 40,000 to 200.
So evidently this isn't profitable for them.
And I know we can all complain about that in particular.
It's profits profit again, bleed 40 grand a month.
All right. They said in other cases, a small number of users are violating our uses policies by sharing and reselling accounts.
This impacts capability. for all cloud users, and we're taking appropriate actions to stop it.
Okay, this one makes perfect sense. Obviously, if you're getting an account for 200 bucks and selling it to a ton of people so tons of people can use it all at once, that should be against the rules.
You gotta use an API if you wanna resell their stuff, and this is for individual users, so I'm not really...
I'm not mad about that and it makes perfect business sense.
They said from August 28th, we'll introduce new weekly limits.
So this is the new thing that they're having.
They basically had some limits in the past.
And while some people are complaining about this, I'm honestly just happy that they've clarified what the rate limits are.
Because up until now, it's been very opaque and hard to even understand.
So they're now going to have these new weekly limits before they just kind of had...
I believe like hourly limits or total limits or monthly limits.
Now it's weekly. They say that will mitigate these problems while impacting as few customers as possible.
We'll also support max plan users by buying additional usage at standard API rates if they choose to.
Okay, now this is the part that gets a little bit crazy and is reminding a lot of people of basically the drama that happened with cursor earlier in the month where They had a similar kind of like unlimited plan.
Cursor was, in my opinion, and they apologized, but I think it was shady.
They switched so that basically if you had a credit card on file with Cursor, you had their unlimited plan.
They like switched without really telling a lot of people so that they didn't have an unlimited plan anymore, right?
It was rate limited. But if you went over the rate limits, they automatically billed you for the extra usage, which I'm like... come on, you got to give people a warning, you have to opt in.
So people were just like waking up at the end of the month, and they had to have like a $2,000 cursor bill when they're on the $20 a month plan.
And they're like, what? So basically, I'm hoping Anthropic handles it better, because it says if they choose to, so it seems like this is an opt in thing, right?
Basically, it's gonna say, you've reached your weekly limit, do you want to just use like the API and continue, but you're gonna just have to like pay a la carte for your usage?
I think a lot of people are gonna say yes, because Basically, as of last month, I was spending thousands of dollars, a couple hundred bucks every month.
Every few days. I mean, it does make sense.
And hopefully it's not going to be expensive.
And hopefully it's like you don't hit that rate limit maybe every day or all the time.
It's just at times when you have like a big project or a big deadline or a big push and you're really cranking on usage, then you could go over a little bit and make it happen without because what's the alternative?
The alternative is to like make multiple cloud code accounts.
So you don't really want to have to like max out one account, switch the next account, max out that account, switch the next.
It seems a little ridiculous if you could, if you could be, you know, make it all on the same account.
All right. Then they said, and this seems like the final kind of like backpedal where they're like, look, guys, I mean, we're just figuring stuff out.
They said, we're still exploring the best ways to ensure as broad access as possible to cloud code.
If you're a power user who has feedback on how we can support your specific use case of cloud code, we'd love to hear from you.
I mean, they're kind of leaving the door open, but basically if you're using it and using an insane amount of money, I don't think they're going to just still continue to give it to you for 200 bucks a month.
All right, what are the responses that people are giving?
First of all, I think you can imagine, and I'm gonna break down the exact actual rate limits, but first I wanna go on a couple funny posts on Twitter.
If you're on Spotify or YouTube, you can see me sharing my screen on all of this, but otherwise I'll explain.
Top comment on this, someone says, "'Towards intelligence, too cheap to meter.'"
I mean, and this is SmokeAway, who they got good posts if you're not following already on Twitter.
But I mean, basically, the concept here is like everyone keeps saying, you know, eventually the AI is going to be so cheap.
We're going to have so many data centers.
It's going to be so smart. It's such unlimited intelligence.
And it's so cheap. It's basically free. And this is kind of the promise that we've heard from a lot of people, Sam Altman, like scaling up so we can make everything cheaper, allegedly.
But we are getting to this point where things are just getting more and more expensive.
And you're also getting more and more rate limits.
And so people kind of complain. But in any case, someone has a picture of a cat who is pondering and it says, we had a good run.
It is true. I mean, for those that were getting $40,000 worth of compute for $200 a month and they're complaining now, you got a good run and I hope you enjoyed it.
Now, this is an example of someone that actually was doing the 24-7 cloud code run.
So it's McKay Wrigley. He said... He said about that and he basically reshared one of his projects that he had posted earlier this month.
He said, so I gave Claude Code a Mac mini and it's called Claude Pewter.
It runs 24 seven and it's allowed to do whatever it wants.
It's in complete control of the computer.
He has a two minute demo where it's just like running it, doing iterating on tons of things.
It's trying tons of different projects. Basically, he's one of those use cases.
Potentially, I'm not sure like how much code he's actually spitting out or how hard he can push it.
But basically, it's running 24 seven. So. he is one of those 24 7 examples and i think he he gets a place in the in the books for legendary someone said and of course everyone is suddenly going to be in that five percent right so that's kind of what i was getting at at the beginning was uh they say it's only in effect five percent of users and i'm like and somehow everybody is in the top five percent congratulations you guys are all incredible and cracked um someone said rug pulls are a weekly occurrence now i thought that was hilarious also i love i love complaining about this because like people are getting you're getting so much bang for your buck so i love to call it a rug pull as i'm literally like getting thousands of dollars of usage for 200 bucks a month um i think it's kind of funny also especially because it's not like a permanent rate limit you can pay more you can pay the regular api price if you want more.
Anyways, someone had a post that says, but the sign said all you can eat and it's a picture of Homer Simpson getting dragged out of I think a buffet.
I think that's hilarious. Yeah, all you can eat.
Well, turns out some people eat too much.
So we are putting some limits on it. All right.
I think that's pretty funny for the reactions.
I wanted to get into some of the specific rate limits that are actually happening and what that actually entails.
So based off of this whole announcement, Anthropic has a status page that shows... Okay, so another reason why I believe this is happening.
There's a status page over that Anthropic has, and it's basically showing the outages.
And you can also see that based off of the last, I think the last... seven times that Claude Code had an outage.
The last seven times were literally in the last month.
So they've had seven outages in the last month.
Obviously, this thing is getting run up to quite a high degree.
So it's, you know, if they felt like they had to do something about this.
All right. What are the actual limits? Okay.
Anthropic commented on the story, basically, and they said that most pro users can expect to get about 20 to 40 hours of Sonnet 4.
Sonnet 4 isn't necessarily their best, so 40 to 80 hours, that's their weekly limit, which sounds like a lot.
Um, and that is just pro users. And of course they're using like a lower model because it's a lower tier, the pro users.
Okay. But what are the, uh, really what I'm interested in is the max, because that's the plan that I'm on, but I guess everyone has different plans.
So I'll, I'll just shout them all out. Um, if you're on their hundred dollar a month max plan, you can get a 40, 140 to 280 hours of sonnet.
But that's not what people care about. People really care about the Opus 4, I think, a lot because that's technically their best model.
Some people say it doesn't code as good.
Whatever. I won't get into the argument there.
But basically, their top model, you get 15 to 35 hours a week of Opus Pro.
Now, if you're working 40 hours a week, you could almost use it the whole time, almost, although 15 to 35 is kind of a big range.
And so if it's the 15 end of the 40 hours, you're kind of doomed.
Um, also we know that coders do not work 40 hours a week.
They work 80 because they're crazy. Um, but also if you are on there, but that's a hundred dollar a month plan.
Okay. If you're on the $200 a month plan, max plan, which is what our company AI box has for our developers, you get about 240 to 480 hours of Sonnet four.
But more importantly, you get 24 to 40 hours of Opus four.
So basically you get 40 hours technically.
Um, And I suspect we'll be buying the API additional credits for a bunch of our developers that want to do more than 40 hours or 24 hours.
I mean, really, it depends how much code you're spitting out.
But in any case, the company has noted that usage is also going to vary based off of your code base size and other factors, right?
So if you have a small code base, much faster, big code base, big context windows that it's taking in and out data.
So it's going to cost you a lot more. I have a bad feeling that AI Box's massive code base is going to doom us into the big code base bucket, but whatever.
It's also not super clear how Anthropic is measuring their usage, right?
So they claim that $200 a month plan has about 20x more usage than the pro plan.
They're also saying in hours, but they're giving really wide ranges of hours.
This isn't very clear. It's not like an X amount of tokens.
They're like, look, you get 40 hours. It's like, bro like what 40 hours like some lax developer that gets a runs a couple prompts every once every here and there to you know just get a little supplemental help is that 40 hours or is it the dude that's like rewrite this entire feature in this entire page no make it like this add this thing like huge things and right and this is kind of what i think what's exciting about cloud code is that you could do those big changes so i think the hourly thing is still quite opaque Um, but based off of the updated figures now, because so they said that, you know, Anthropic basically claimed that the $200 a month plan had 20 times more usage than the pro plan, right?
Just their basic paid plan for people. But based off of the updated numbers, they've just shared.
So subscribers now get only about six times as many cloud code hours as pro user.
So they basically scaled it back, but they haven't updated the fact that they said you get 20 X. Now you only get six X, which is still a lot more.
But in any case, um, I think they're intentionally opaque.
They're measuring things in hours instead of in tokens because they're trying to figure out where they're profitable and how much people abuse them or complain as they rate limit them.
In any case, I think this is very interesting.
They're going to be, you know, not the first person that is making these big updates we've had.
Like I mentioned, Cursor did this and they had a bunch of drama about it.
They apologized. We also had a replica make some similar pricing changes.
So it seems like people are all, it seems like this is kind of an industry trend.
And I think we're going to see a lot more companies as it becomes less and less profitable.
Companies are going to try to thread the needle of not making too many of their users upset, becoming profitable.
You also could use like, you know, basically a user that has a $200 a month plan, but maybe only uses $100 worth of credits.
They're they're, you know, they're offsetting the loss of someone that might be using like $250 worth of credits a month.
And so I think Anthropic wants to have as many users as possible.
They don't want to like curb stomp all of the hopes and dreams of their power users either.
Because if those people go other places, you know, it just makes their whole ecosystem get used less.
So they like they want to keep as many users, but also they have to make money.
So it's gonna be interesting to see where they, to see where this all shakes out.
Thank you so much for tuning into the podcast today.
I hope you learned something new and I hope this was insightful and gave you some good feedback from the community and feedback from exactly what Anthropic is actually doing.
Thanks so much for tuning in. If you wouldn't mind leaving a rating and review on wherever you get your podcasts or if you're on YouTube, to like the video and subscribe.
It helps the channel out a ton. Thanks so much for tuning in.
Make sure you go check out AIbox.ai. There's a link in the description to try out all of the different AI models to see which one is the best at coding or writing or any sort of project.
You can compare different AI models side by side.
We're currently in beta. So thanks so much for tuning in and I will catch you in the next episode.
