Theo - t3․gg · 16:46 · https://www.youtube.com/watch?v=y5TEATUDDfc
0:00 I'm not going to lie, it kind of feels like Anthropic just can't stop winning. Recently, they've been on quite a tear with all of the recent model releases blowing us away, price changes that make a lot of sense, and generally just not being as annoying to work with. Today is another step in that direction because Anthropic has just changed their plans. They've just introduced the ultra tier to their subscriptions. And don't worry, they're not here to charge a whole bunch more money.
0:21 It's the exact same plan as the 20X was before, that is the $200 a month plan, but they made some subtle changes to it that are really worth noting. In particular, they removed the 5-hour limit. This is just a great change all around, right? Like there isn't much to cover here. I should be able to just wrap this video up fast. Oh, for verified ultra accounts.
0:43 >> [sighs] >> You never make this easy for me, do you, Anthropic? It wouldn't be an Anthropic release without at least a little bit of weirdness, right? That verified callout is part of it. The way that they enforce the 5-hour limits is another part of it because remember, they're only removing this from the $200 plan, even though those 5-hour limits tend to hurt a lot more on the cheaper plans. Ah, there's a lot to dive into here, and I'm going to do my best to cover it all and also talk about how this affects us with T3 Code, which sadly, this does directly affect us with T3 Code.
1:11 Don't worry too much, though. We're going to be good. Either way, we got bills to pay, though, so let's take a quick break for today's sponsor. If you love bugs and buggy software, you should skip this ad. But for everybody else, you should really consider using Grappl. Grappl's AI code review platform is a lot more than just code review. If I'm being honest, it feels like they're the only ones who really understand what we want as developers.
1:30 If you've ever had to pull down a pull request just to make sure the changes worked, you'll understand why T-Rex is so cool. It allows their review agents to run your code in a sandbox to make sure the changes are actually valid and working. If you've ever had to keep your docs and code base in sync, you'll love their new knowledge base product that allows you to do that automatically. As your code changes, the knowledge base does, too, and it's all fully editable if you go in and want to make changes.
1:51 It's also exposed over MCP, so your code agents can do better at making changes because they have all of the knowledge available to them through the MCP exposed by Grappl. And this one is small, but as I'm watching my cost balloon, I really appreciate it. They have three different tiers of review in Graptile. So, if you want to just quickly get some feedback, you could use their base tier, which is relatively cheap through their credit system.
2:11 Or, if you want super thorough reviews for those big scary 20-plus file PRs, that's where the Apex tier comes in. You get to pick what level of review you want when you're working with Graptile, which allows you to right-size the effort based on what changes you're making. Everybody shipping faster nowadays, do it with fewer bugs and more practically at swadev.link/graptile. It is worth noting that part of this name change is to clean up some misleading things that existed before.
2:37 Specifically, that the 5x and the 20x aren't actually reflective of your overall limits. These are just for the 5-hour limit. Because, believe it or not, back in the day when these plans were introduced, the 5-hour limit is all they really cared about. And that was what people were hitting all the time and why they wanted upgrades. The weekly limits were only introduced as our usage patterns changed fundamentally, and they wanted to keep up the super generous usage that we got when we were used to using these with like Sonic 4 in the OG Claude code CLI.
3:05 The $20 plan was meant to be the like all you can eat buffet style, like super, super generous plan. It still is very generous, like absurdly so. But, that generosity is because they set a certain bar for how much work you could do, and people wanted to still have that without feeling restricted under the same limits. Mike, who's the head of product for Claude subs, made a really good follow-up post here about this. He also went on about how when they introduced the $20 plan in April last year, April 2025, it only had the 5-hour limit.
3:34 And the point of this plan was that it gave you 20 times the amount in those 5 hours that the Pro plan did. So, they called it Max 20x. And when they added a weekly limit to Max 20x, they set them to be 10 times the limit of Pro. Many people assumed that the 20x applied to both limits. They should have done better, and he personally apologized. Actually really nice to see. This it it just feels much more human than I'm used to from Anthropic.
3:55 Good work, guys. This is an improvement. The new naming is much clearer. They are very specific about these limits in all of their pages and everything now. Like it just directly says it's 10 times more than Pro, which is what it is. Awesome. Cool. So, what's next? There are four subscription plans for Claude: free, Pro, Max, and Ultra.
4:12 Free is of course free, Pro is 20 bucks a month, Max is 100, and Ultra is 200. They all offer different limits for people who use Claude different amounts. And none of them are enough for super heavy power users like myself balancing seven different Claude accounts. But for most people, the $200 account will be more than enough for day-to-day realistic dev work. Hell, even some heavier stuff would be totally fine in Ultra. When I turned off all my hill climbs, I struggled to even do half an account with Ultra right now, if I'm being real.
4:38 Max is also quite generous at the $100 offering there, but it does still have the 5-hour limits. We should talk about those really briefly. Claude subs have two limits by default. They have the weekly limit, which is how much Claude you can do in a given week before you have to start paying API prices. And separately, they've a 5-hour limit that keeps you from using the whole thing up in one burst. Why do they have that 5-hour limit that's separate from the 7-day limit?
5:01 The reason is because of compute availability. The level of usage that the subscriptions give you is absurd when you consider how expensive the models can be over API. The $200 plan gets you as much as 12,000 dollars a month in inference, which means if you burn that all at a specific time, it can cost Anthropic a lot of money. Not because it's just like melting money in their bank accounts. It's not like every time you send a prompt, it costs them a certain amount of money.
5:25 Their compute already exists, and it's either being used or isn't. So, as long as our usage in the subscriptions does not take away from availability that they could sell to enterprises, all is good and dandy. The problem is the way that usage occurs over the day. I just ran the numbers against our analytics for T3 Code to give you an idea of how much this varies throughout the day. On a given weekday, there's a lot of variance, but it really starts to spike around say like 4:00 a.m.
5:50 Pacific time during the weekdays. Absolute peak of prompts being sent is between 7:00 and 9:00 a.m. Pacific time. And then, as we approach the evening from 6:00 p.m. up until like midnight or so, you have a huge lull. There is a 2x gap in how much prompting is being done from the lowest lows and the highest highs just in normal variance of a given weekday. I bring this up because uh what happens if you spend whole of your usage for your week right here in the peak?
6:21 Eventually, you're going to hit a point where the amount of available computing Prophetic has is just not enough for them to serve these customers and then everybody sees these not enough compute available errors and starts getting upset. So, the 5-hour limits were introduced as a way to prevent a subscription user from taking more than a certain percentage of their weekly sub in any one of these work windows. If you broke up this chart into 5-hour chunks, the most a user could do would be I don't know, like 10 to 15% of their weekly limit in any one of those chunks, which massively impacts the amount of available compute that Anthropic has when they are trying to sell it during the workdays.
7:01 Back in the day, Anthropic even went as far as giving you worse limits during these windows and more generous limits outside of these windows because of how bad the compute scarcity was affecting them. And the other harsh reality is that during these late-night hours, there's probably a lot of GPUs they're paying for that are sitting not doing much. They don't mind when you burn during these windows. In fact, if you take the time to go through all of the resets that we've been given for Codex subs, you'll notice they tend to be either late at night or on weekends because that is when they have the most compute available, where the compute isn't being used heavily by the companies that they sell to.
7:37 Because it's also worth noting that these $200 subs don't make the labs much, if any, money. The point's to get us addicted to using 12 grand of tokens a month so that when we go to our employer, we're expecting to spend a similar amount there and they'll have to pay those API prices. There is a small benefit to the 5-hour limit to us as users. Makes it slightly harder to accidentally overrun your whole subscription. I know I and others have had this silly mistake that like, let's be real, any us can make,
8:03 where we send a prompt we think will be resolved quickly, leave, come back, and it's just burned an insane amount of tokens that we never thought it would have. The 5-hour limit helps a little bit with preventing that, which is nice, but overall, I'm happy to see it gone, at least from the verified Ultra accounts. Uh
8:20 yeah, we'll talk about the verification thing in just a sec, but first we have to time travel a little bit all the way back to May. Feels like years ago that this happened. It's crazy it was only a few months. This was when Anthropic announced their partnership with SpaceX for compute. This was a massive multi-billion dollar deal that allowed for Anthropic to borrow SpaceX's GPUs from their old data center that they weren't using as heavily, and I don't think they're using it at all.
8:44 I'm pretty sure Anthropic just runs that whole data center now. And when they did this deal, suddenly Anthropic had access to more GPUs and more compute, so they could run more Claude. And with that, they made a huge change. They doubled the Claude code 5-hour rate limits, and they removed the peak hour limit reductions for Claude code. The reason they could do this is, again, compute was available finally. They had to have these limits to restrict how much we could abuse Claude in a given hour.
9:09 To be clear, this wasn't restricting how many dollars we could spend in terms of API prices. This was limiting how much GPU we could use during those work hours where they needed those GPUs to sell enterprise inference to these bigger companies. They don't care about how much money our usage costs. They care about how much GPU they have available for enterprises. So, when they had more GPU available, they got way more generous with their limits. Note the wording here though, cuz we're going to come back to this.
9:34 Doubling Claude code's 5-hour limits. We'll be back. Because now we have to talk about the verified aspect here. Why are they limiting this to verified Ultra accounts, and what even is a verified account? They specify a little bit here, saying that this starts today with most existing subscribers. Newer accounts keep a temporary limit until they're verified, which helps us prevent fraud and abuse. You'll get an email when it's removed. Yeah, I don't love this. I specifically don't love the lack of transparency here.
10:06 They have been truly like legendarily vague about what verification even is. They say they remove the limit automatically when your account is verified, so there's nothing you need to do. Verifying your phone number at checkout is a separate step and does not remove this limit. To check, go to settings usage. If you don't see a current session bar, your account is verified. Yeah, not great. It would be really nice for them to give a little more detail on how this works.
10:33 But when you understand why they did it, I think you'll understand why they can't. One of the biggest issues Anthropic has to deal with right now is distillation. In order to do distillation, you need a lot of real-world requests and responses from real-world Claude usage. One of the best ways to generate this is to get an API key. And if you can't do those because, I don't know, maybe your country is sanctioned and you're not allowed to do that, or you don't have the money and you're trying to save because again, $200 in, $12,000 out.
10:59 It's a lot cheaper to use subs. So, as you can probably guess, there's a lot of interest from these Chinese labs and other companies that want to do the distillation to get these subs, burn them as quick as possible, generate a ton of responses, and then use that for distillation. The removal of the 5-hour limit does make it slightly easier to do that, which is why they want to enforce things more strictly this way, but it's also just weird and hits users that are doing things the honest way a little bit too hard.
11:29 From my calculations, you can do roughly 30% of your weekly limit through one of those 5-hour windows, so you get three and a half or so 5-hour limits in a given week. But here's where we get to the final catch, that hint I gave earlier in the video. Remember when I pointed out they specified that they were doubling Claude codes' 5-hour rate limits? They meant what they said. And what they meant was only Claude code.
11:52 This has been incredibly annoying for us to figure out because as people have started using T3 Code more heavily, they have identified the fact that their 5-hour limits burn much faster in apps like T3 Code. This user pointed out that T3 Code uses around 1.8 times more of the 5-hour window per token than the interactive CLI does. Even {dash} P does the same thing, by the way. Incredibly annoying. This is an issue that's opened on the Anthropic repo that I became aware of because of how many people are reporting it to us in T3 Code.
12:24 I did my own calculations to confirm this, and it is indeed true. I've been working with Anthropic on this because I was hoping it would get fixed. I patiently waited, and they did this change, which to be fair, for a significant portion of our users, does resolve the problem because anybody on the $200 plans no longer has 5-hour limits. So, you're good. We're done. We're This is solved now, right? Well, the $20 and $100 tier people are extra screwed because their 5-hour limits are worse, which already sucks.
12:54 And when you combine that with the 2x decrease for agent SDK users, it's just garbage. I am not fond of this. I've been trying to get Anthropic to do something about it for a minute, and uh not had the best luck, if I'm being real. I was hopeful. I've lost hope, which is why I'm going to tell you guys what I do. I already see the comment happening, so I'm going to jump on it now, but Theo Open Code has the Claude sub.
13:15 How are they doing it without agent SDK? I was really excited about this release. I thought this was Anthropic finally being chill and letting Open Code integrate Claude directly. I was wrong. They're using agent SDK under the hood, too. If you are using Claude subs in your apps, you either have to have your app just call the Claude Code CLI directly with like the terminal visible, or you have to use the agent SDK, which means you're going to be hit with this 50% decrease in the 5-hour limits. Obnoxious.
13:43 I really hope they decide to fix this instead of causing more problems. I will say before I go any further, I cannot be held responsible for any account bans. Okay, now that said, there's some very easy solutions here. Solution one is there is a field called Claude Code entry point that we currently have set to SDK-TS for TypeScript for the official TypeScript agent SDK. If you change this to literally anything, the problem goes away. There are lots of different ways to set this.
14:08 Hypothetically speaking, you can make this problem go away pretty easy in this certain JSON file that will be loaded properly into T3 code. I think we should be reporting this as T3-code anyways because, let's be real, that is the entry point. The entry point isn't SDK. TS The entry point is T3 code. So, I think this PR is an accuracy win, but I'm scared if I merge it, might cause problems when Anthropic sees a huge spike in this new entry point.
14:32 So, do as you please here. Well, good news. Things changed literally while I was filming. You might see this 2 minutes ago. Yeah, I got semi-solid permission to merge this which should, hypothetically, fingers crossed, get around the 5-hour limit changes that affect T3 code so that you will get the same usage for the 5-hour limit as well as for the longer-term weekly limit in T3 code as you would in Claude code directly. So, yeah.
15:01 Thank you to my friends at Anthropic for being reasonable with all of this. I know it is not trivial to convince that company to let you do things, but yeah. I wanted these limits to not affect our users, and as of this moment, literally as I'm recording, once this release goes out, they no longer will. Huge win. That all said, I didn't notice these problems because I've been routing through CLI proxy anyways, which already applied things like this to prevent this issue.
15:27 So, again, if you wanted to have multiple accounts or manage all of that, I don't recommend doing it in T3 code directly. I will not be held accountable for any of these things getting your account banned. I'm not telling you you should go do this and that it's totally safe and friendly and fine and Anthropic loves it. I'm just saying what my experience has been. And from my experience, if you hunt through the FAQ, you might just find a guest somewhere that shows my exact setup that your agents should be able to take and understand and work against.
15:56 And this is the exact setup that I've been using for all of my inference for at least 4 months now, probably more, and I haven't had any issues personally. So, again, take it as you will. Few things I have to say about the 5-hour limits though. That's all I have to talk about here. A nice, simple, short video all about the new Claude sub plan that isn't really all that new. I'm sure there won't be anything else I have to cover similar like, I don't know, a new Codex plan Thanks, Tibo.
16:18 Coming up very soon will be my video on the new Codex limits and how they compare to what Claude offers right now. Because before this change, the Codex plans were just not a good value in comparison to the Claude plans. Now the math is very different and I will go through all of that math with you guys in a video coming very soon. So subscribe if you haven't. Let me know what you think. Hopefully this was a useful breakdown, a bit different from what I normally do, but just wanted to get this news out as quickly as I could.
16:42 Let me know how y'all feel about it and until next time, peace nerds.