Does the $200 Codex plan suck now?
TTTheo - t3.gg@t3dotgg · 572K subscribers
- Views
- 166K
- 165,985
- Outlier
- ×1.0
- Median 167K
- Length
- 33:33
- Published
- Oct 7
- Wednesday
- Comments
- 317
- 2.3K likes
Script
How it's written
- Words
- 6,614
- Words a minute
- 197
- Reading time
- 28 min
- Segments
- 24
The hook
0:00–0:25Someday, far into the future, my children will be using their Samma coins to try and buy dinner, and I'll have to explain to them that when I was a kid, we used to use US dollars in order to get our Samma coins, and we used to pay $200 for $1,200 of Samma coins. I know I'm memeing pretty hard here, but seriously, just a few weeks ago, a $200 plan on your Codex sub would get you up to $12,000
Structure
Someday, far into the future, my children will be using their Samma coins to try and buy dinner, and I'll have to explain to them that when I was a kid, we used to use US dollars in order to get our Samma coins, and we used to pay $200 for $1,200 of Samma coins. I know I'm memeing pretty hard here, but seriously, just a few weeks ago, a $200 plan on your Codex sub would get you up to $12,000 a month in inference. And on the Claude plan, it wasn't quite as much, but it was still 200 bucks in and $8,000 out. We got here in a way that was honestly kind of understandable. The $200 plans were introduced to feel unlimited back in the day where we would send maybe 10 or 15 prompts on our heaviest days, and we would get a response back in under 5 minutes, usually just touching a few files here and there. And when we were all using Sonnet to explore our code bases, $200 plans were practically unlimited. But as models got more and more capable and more and more expensive, neither OpenAI nor Anthropic wanted to take the optics hit of somebody being happy with how much usage they get for 200 bucks and just to have it stop feeling unlimited when the new model comes out. That has resulted in this crazy arms race where Anthropic and OpenAI kept subsidizing these plans more and more until we got to these crazy numbers. Right before Dev Day, Tibo made the biggest announcement that OpenAI has made arguably this year for us vibe coders. He announced that they would be changing how they calculate usage for
Watch from 0:00the $200 plan, in effect giving you half as many dollars as you would have had before. That's pretty bad by itself, but I ran the numbers. I only got $570 out of a week of usage, which adds up to around two grand a month. So, is that the end? Did we really just go from 12K a month of usage to 2K and just smile and wave? Is there any reason to use the Codex sub and not the Claude sub, or is this just some secret plot to force us all onto the new $500 tier instead? And I can't wait to tell you those answers right after a quick break for today's sponsor. Agents need access to the internet to do tasks well. It's just a fact. It doesn't matter how smart the model is if it doesn't know how to get to the internet and reference the things it needs, it's not going to do a very good job. That's why your agents need a great browser. Ideally, today's sponsor, Browserbase. These guys built the perfect browser for your agents so that they have access to the entire web. And I don't mean just the browser. I mean all the different APIs, search capabilities, and more that your agents need to do all of this. None of that is my favorite part of Browserbase though. I've said for a while that Stagehand is the coolest thing they're building. If you're not familiar, Stagehand is an open-source SDK that allows your agents to do things on web pages with plain English instead of having to define every single element that they're interacting with. I know showing code isn't the cool thing to do nowadays, but look at what I mean here. You can grab fields from the page by awaiting stagehand.observe description, like find the email input or find the password input. And once you've identified these fields, you can fill them with different variables and whatever else you want to do. This is so useful, especially when
Watch from 1:30you combine it with their Zod implementation that lets you get validated data out of a page. This is great for you and me writing the code, but let's be real. Neither you nor me are writing the code. That's why we built the demo with agents showing just how powerful this is. This demo is going to fetch the latest releases from these three different repos on GitHub. We click scan, we can see it opening and it's navigating in a real browser because this is actually controlling a browser. It's not just editing HTML for you. And now using those plain English instructions, it's able to find the right page and to get the data we want. What's even cooler with this is since it's not fixed on specific elements, if the page gets updated in the future, you don't have to rewrite all your code to deal with one element moving slightly on the page. And now we have all the results. A nice little brief that shows all of the releases for the repos we put here. Make your agents more resilient and get them better data at so dev.link/browserbase. So, what's going on here? If OpenAI wants so badly to be seen as a better value prop than Anthropic, why are they giving us so much less than they used to? This is just over, right? Like there's no reason to have the $200 Codex sub anymore, right? There's layers to this one. The biggest issue that we need to jump it on top of now is the gap in these numbers. The number on the right seems just insane compared to the number on the left. They're basically giving us a crazy cheap number here. It's almost like made up. And when you compare that to what these models actually cost over API, it's just it's nonsensical. And when you look at my numbers for the last 7 days, it seems even crazier. I'm doing 20 grand of API spend in 7 days across
Watch from 3:02my machines. It's nuts. Obviously, I can't pay that. Like I'm doing okay, but I I don't have it in me to pay 20 grand a month to use these things over API. The reality is that not only is this $200 number made up, so is the API price on the right. Both of these numbers are fake, and neither of them represent the actual costs to the companies when they're running these things. The number on the right is the number that they have chosen to charge enterprises hitting these things over APIs to see what they can get away with. And the margins on these numbers have historically been insane, upwards of like 95% against the hardware and the energy costs. Obviously, they had to put a bunch of money into training the model, hiring all the employees, doing everything else. But the actual compute and energy costs for burning $100 in tokens is closer to like two to five dollars than it is to that $100 number. Since this number is so outlandishly inflated by these crazy margins, this number feels insane, and the gap between them feels even more insane. But in an ideal world with enough competition, Anthropic and OpenAI would both have enough pressure to lower the number on the right that the gap between the subscription subsidization and just paying for the model for what it costs them, that gap would get collapsed relatively quickly. And it seems like we are at the start of that now.
Watch from 4:33We will get to what I mean there in a second, but I also want to talk about the weird things that made the Claude sub feel worse before. The first thing, the one that pissed me off the most by far, was that you could only use it in Claude code. The Claude sub isn't a sub to do $8,000 a month of Claude. The Claude sub is a sub to do a large amount of inference in Claude code itself. You can't use your off here other places. You can use it in T3 code because in T3 code we are just calling Claude code directly through the agent SDK. But if you wanted to throw it into something like Pi or Open Claude or Hermes, you can't because those things aren't Claude code. And if you think you can be clever and add the instructions to work like a tool like Pi in your Claude code system prompt, good luck. The request will fail. They are that petty at Anthropic sadly. I am hoping that they're going to lift their restrictions here in the near future simply because a lot more people are maxing out their subs. So it matters less if they're doing it inside or outside of Claude code. Regardless, they've been awful about that. OpenAI has been the opposite. You can use your Codex sub for basically whatever you want, basically wherever you want, as long as you're not using it to serve user traffic. So if I put my Codex sub in the cloud and then charge you $5 to access my subscription, that's not allowed. But if I use my Codex sub to run on CI to review everyone's PRs when they come in, that is mostly allowed.
Watch from 5:55Just want to draw the distinction here because what you're allowed to do with your Codex plan goes a lot further than what you're allowed to do with your Claude plan. Claude also has a restriction on the usage of the biggest model. Fable usage cannot use your whole sub because Fable's price is not what it probably was meant to be. I have two sets of models on the screen here. Set one, Opus 4.5 to Opus 5.5 and GPT 5.6 Soul plus GPT 6 Soul. The second group is Fable 5 and 5.1 and GPT-6 Astra. What is the difference between these two sets of models? It seems obvious, right? The bottom ones are the big expensive ones and the top ones are the small cheap ones. You're also noticing something weird. I said GPT-6 Soul, not 6.1. Which side do you think 6.1 goes on for the comparison I'm trying to make here? GPT-6 1 Soul is on the other side of this line. And you might be thinking, "Oh, is that cuz it's a big model and the others are smaller? Are you saying 6 1 Soul is bigger and more capable than Opus or some other dumb thing?" I know how the comments tend to go. I'm putting 6 1 Soul here for a different reason. Remember Project Glasswing? AKA the announcement for Claude Mythos when Anthropic confirmed they had this crazy model internally that they weren't going to sell? They actually were going to sell. They were doing this big commitment early, but they planned to continue offering Claude Mythos to participating companies at $25 per million tokens in and My assumption here is that these numbers were calculated using the previous margins.
Watch from 7:14If Anthropic was used to having a 95% margin on their compute, so they would spend $5 and have 100 come in for their power and compute costs, they planned to maintain those margins here. Let's just play with that 95 number. That would put the actual cost for a million tokens out, not at the 125 they claimed, but at 625. So, let's just like write out these numbers quick. So, say Mythos was going to cost 125 per mil out. And that would be $6.25 in cost. That keeps them at their 95% margin. But, you might be confused about those prices because that's not the price Fable came out at. And Fable and Mythos are the same model. I know some people seem confused about that. I'm not going to entertain that right now. They are the same model because when Fable and Mythos 5 came out, it was a different price. Fable 5 and Mythos 5 are being offered at $10 per million in and 50 per mil out, less than half the price of the Mythos preview. It's not like they made the model smaller and dumber before releasing it. They were very clear about that. They actually just made it bigger and more capable. So, again, we'll copy this and say Fable 5 {slash} Mythos 5, price is now $50 per mil out, but it's the same model size, same hosting. Which means that their costs probably didn't go down. Which means their margins have dropped pretty massively from 95% to 87.5%.
Watch from 8:48If you were to look at this as the profit that they're getting at a given time, it's even more brutal. Where in this instance, they would have profited $118.75, here they're only profiting $43.75. The margin drop is big, but the return drop way bigger. They used to spend 625 in compute and make $118. Now, they spend the same in compute and they make $43. The line I drew here is the start of the end. What is it the start of the end for? The subsidization. Anthropic chose with Fable 5 and Mythos 5 to price it way cheaper than they originally planned in order to just really make Open AI's life harder. And they succeeded with that. That was the start of the panic at Open AI, for sure. And even if the change in margins doesn't seem that much, the change in the revenue is world's difference. And if you bump that up to what was possibly a 98% margin before, these numbers get even crazier. The point I'm trying to make here is that the amount of revenue Anthropic makes per amount of energy and compute used has gone down meaningfully with Fable. But they chose to do that because the margins are still insane. Like they're still just printing money with this. It's just that the amount they're printing has gone down significantly. And this also forced OpenAI to do the same. I don't know what their planned price was for Astra, but they released it at the exact same price as Fable. So, we have no idea like what
Watch from 10:15they would have charged before Fable because they made the model with a certain price target in mind probably. But, they did a few other things on top of that in particular with the subs that made it really brutal. They took this hit on the cost, so they might have intended to charge 100 bucks per million output. Instead, they're only charging 50, so their margins are worse. Meanwhile, Anthropic doesn't even give you your full usage of your sub with Fable. You only get half your allocation for Fable. The other half can can be used for whatever other models, but you're not using 100% of your usage with Fable. It's limited massively. OpenAI chose to not do that, and they let you use Astra for your whole weekly limit, not just a portion of it. And if you're curious where that 50% cut came from with Anthropic, it came from them not doing it for the 3 days before the ban, reading all the data, and also here. It's not exact, but $43 is half to a third of $118. They're profiting half as much, so they give you half as much. Makes all the sense in the world, right? Let's keep going down this rabbit hole a bit because there's one other model here that does not seem to fit that well. 6.1 Soul. Considering the way things are currently priced in the industry, my honest assumption would be that 6 Soul was kind of meant to replace Terra because nobody used Terra. And eventually, they would do something more like Opus between Soul and Astra. They didn't do that. In fact, they spit out 6.1 Soul suspiciously close to 6 Soul,
Watch from 11:46which suggests that this was not necessarily what the model was meant to be. As such, I'm going to pretend that they were looking at the existing numbers for the competition cuz we've now established, hopefully, that OpenAI's pricing is based less on their margins and more on their competitors because they set Astra's price to be literally identical to Fable. So, they were probably planning when they were training to make 6.1 Soul a little closer to Opus. Not Opus 5.5 cuz it didn't exist yet, but for Opus 5, which was $5 per million and $25 per million out. We will assume these are the numbers that they were operating with. The new model they were planning was going to be $5 per million, 25 per million out, nothing special with cash discounts. We'll assume the same 95% margin. That means for $25 out, it would cost them around $1.25, which gives them $23.75 profit. Not bad. You spend $1.25, you get $23.75 out. Pretty good deal. But that is not the price that this model came out at. 6.1 Soul came out differently. It came out. It's $2 per million, $10 per million out, and massive cash cost decrease.
Watch from 13:16The cash reads are half as expensive as they were before. They used to be 20 cents cuz it would be a 90% discount. So, it would go from $2 per million to 20 cents per million. It's now 10 cents per million. Doesn't sound like a big deal. 20 cents to 10 cents, 90% to 95%, but when you look at the percentage of your cost and see that for OpenAI models, half your cost is cash reads, now it's 25% of your cost because you just took half your cost and you cut it in half. So, it would have been 50 cents per million read from cash is now 10 cents per million read from cash. That's a 5x decrease from what they had originally planned here. Crazy. So, the results here is that your actual effective costs using 6.1 Soul aren't just 2.5x cheaper than they expected. It's probably closer to 5x cheaper than they expected. What I'm trying to say is their margins for this model are garbage. This 95% no, no. I would legitimately guess their margins for this model are like in the 50% range. Just seeing how much slower it is than 6 soul, it was definitely planned to be 525. So, 210 with the 10 cent cash reads is just insane. So, they destroyed their profit margins. Like they're still making money. I'm not going to say like 50% profit margin is bad, but it's pretty obvious that they aren't just printing money on the inference the way they were before. Remember earlier I was talking about that 50% limit for fable?
Watch from 14:33How they cut your limit in half when you were using fable models? Open AI had said they didn't want to put an arbitrary limit on Astra. That's what happened here. But it would have been the other way. Since the model they cut the subsidization for the most was 61 soul, they didn't want to put a custom soul limit. They could have done that. Open AI absolutely could have put in a custom limit that was your soul only counts for half your usage. But instead they're taking the opportunity to try and collapse this gap in the subsidization a bit so that your account sub doesn't feel quite as insanely subsidized compared to paying API prices. I don't think they should have announced this the time they did. I love the attempts at transparency from Open AI. They were pretty transparent here, but this did not come off great, especially because so many people are still using Astra. But this is where we have to dig into something a little more painful, which is the effective costs. I ran a bunch of real numbers for my day-to-day usage. Previously, I was seeing similar to those 12K numbers before. Especially when you combined the fact that we got a reset every 3 days on average many months, you were getting 24 grand if you were truly maxing out every single window that you had. Now it's closer to $2,500 a month instead of $24,000 a month. Roughly a 10X drop. Because I only saw $573.86 of usage, and I was able to kill a whole account in 29 hours in a loop. Not great. But let's scroll down a little cuz I have things I want to talk about in this.
Watch from 15:52My usage. My GPT-6 Astra usage was for around 1,800 responses. In in these 1,800 responses, it did 766,000 output tokens and it cost $534. It was API prices. 61 Soul did almost as many responses. It's a little less, but it was like 1,500 versus 1,800. It's not that big a gap. More input tokens, fewer cash input tokens. I think this might have been a bug in one of my setups or something. I haven't figured out, don't care to. Again, very similar number of output tokens. $38.27. Do you understand? Roughly the same amount of usage, greater than a 10x decrease. The reason they made these changes, all of these changes, is they wanted to make 61 Soul a great value. They wanted to make this $200 sub feel unlimited again, the way it used to, all the way back in the 54 and 55 days. And they wanted to make sure they didn't go bankrupt because of the insane amount of compute that you're getting for that amount of money. The reason they had to do that is cuz of people like me and a lot of y'all who are token maxing right now. Since I use these subs and I push them to their absolute limit, if they didn't make these changes, I would be burning two to four times more compute from OpenAI every month and that's just insane cuz you can absolutely, and I do genuinely believe for any even vaguely real-world use case, you're going to get practically unlimited use of a $200 plan as long as you only use 61 Soul and you never touch Max cuz Max makes no sense at all. Do I think this is a good change though? Meh.
Watch from 17:22I'm going to be so real with you guys. There is more to talk about here, especially the $500 plan. I'm going to be incredibly real and honest with y'all. If your goal here is to use your subscription to write code that you plan to ship, the insane value you get out of Opus 5.5 on the $200 Claude plan is impossible to beat. I thought I would cancel subs with Claude because it was so much more efficient, but I ended up just pushing my own limits to what I could do, and I ended up with a six Claude account. And if I refresh now, it's probably closer to drained. Yeah, like I have killed three of my six. Two of them were dead this morning and just reset. I'm pushing these things. I am very much getting more value than I would ever have expected out of my Claude plans. It's crazy. And the craziest part here is I have not touched any of my fable. I have 600% of my fable. I can't use it on the dead accounts because that gets combined with your weekly. So, I can't use it in those accounts. But, it's hilarious that I've just straight-up not touching fable anymore. So, I still think the Claude plan is the better value by far, but I also think 61 Soul is the best value a model has ever been by far. If I was forced to pay API prices, I'd be picking Soul without any question. But, I'm not. You get less subsidization from the Codex plan, but the model's also so much cheaper that it balances out relatively well. And I'll show you this with a new project that I actually built with Astra Ultra fast, funny enough, Slopolitics. I wanted it to be easier to visualize the score against the dollar spent for these benchmarks. And you can see very clearly here, 61 Soul is an absurd value on the
Watch from 18:52Artificial Analysis Intelligence Index, which is what this data is from. If I switch to the linear view, it's even funnier looking. Truly hilarious. I also ran this on Terminal Bench myself because I was not happy with how others ran Terminal Bench. And not only did 61 Soul get the best score on it when you gave it access to its native harness, I ran it with Codex and Claude code, so I got better numbers than a lot of others did. Not only was the 61 Soul max run the best score ever, the 61 Soul X high run pretty much exactly tied Opus 5.5, but Opus cost $13.11 per task. 61 Soul cost 96 cents per task. A 13th the price. Do you understand? Your subscription might not be as much dollar value, but the number of tasks you are completing with that money is way higher. So, if you're paying API prices and you're paying Opus, you better be damn certain that Opus is the only model that does what you need because if you're comparing the cash value you get in the subs by the pay to tokens, none of this ends up making sense. And this is the problem is we have two levels of like subsidization happening here. We have the amount of dollars you're getting for your $200 sub and we have the cost that those dollars represent to the lab and we have how many tokens it take to solve a problem because if I switch this to show per tokens, which I have in Slop lytics, so do I have it No, I have tokens for task here.
Watch from 20:23You'll see that medium on Opus 55, sorry, high on Opus 55 uses roughly as many tokens as Max does on 61 Soul. And the tokens are cheaper. So, cost per task is what we should be looking at and real-world cost per task on something like Terminal Bench 4 is a 13x decrease. So, the dollar value of the sub being half as much ends up still being five times more tasks depending on the work you're doing. So, again, I don't think 61 Soul is the better model. I would take my Claude sub over my Codex sub any day, but this isn't some crazy rug pull like everyone's making it out to be unless you insist on continuing to use Astra with the much higher cost and much higher spiky nonsense rates. But, here's where I have to stop being nice to OpenAI. Cuz I'm about to crash out really, really hard. And everybody who's accused me of being paid by OpenAI to say nice things, this is going to prove that I'm not, which means I do need to be paid, so we're going to do one more quick sponsor break first. If I told you that you could 5x your potential users with one simple trick, you'd probably say that I'm insane and I understand because you probably didn't know about general translation. These guys are incredible. They made it so easy to translate real-world applications across various different surfaces. There are so many little edges that are hard to get right when you build translations into your apps. Localization is just not an easy problem to solve. I had to do it myself at Twitch and it was miserably difficult. I'm very thankful for my team of much smarter engineers for helping us all through that. Going to do a thing I
Watch from 21:49don't get to much anymore and show you some code because it really shows how magical what they built is. In React applications, you use their helpers in your JSX to actually indicate what section to be translated. They also have helpers for handling annoying things like number formatting and pluralization, which is not trivial in lots of different languages. Where things get much more powerful is their dictionaries that allow you to have consistent translations across various surfaces. So, if you have a feature with a specific name, you don't have to worry about that feature being named incorrectly in three different places. Now, you can finally know for a fact that the same translation and the same verb will be used across all these different surfaces. We tried this out for the T3 code marketing site and when I say us, I really mean our agents because agents are really, really good at using general translation. And after two prompts, we have everything we need to showcase our application in all sorts of different languages. And the best part is we can share these definitions across all of our services, whether it's our mobile app, our desktop app, our website, and more. Reach the 85% of the world that doesn't speak English at swyd.link/gt. Time for the crash out. I'm going to do more of a detailed version of this in a future video, but I'm going to give you the CliffNotes of it now. The $500 plan was a mistake. Not the existence of a $500 plan. Doing it right now. Absolute optical show. To tell people that the 20x plan is now 10x, but don't worry, you can get the 25x plan if you pay 2.5 times more, pisses everyone off, myself included. This was just a mistake. If they were going to do this, which obviously they had to eventually,
Watch from 23:20they should not have done these things at the same time. It is such an unnecessary self-inflicted L that if I defend anything they do, you guys are going to call me a pay shill and I understand because this is such an absurd blunder optically that it makes them look insidious. The reality is that these subscription plans are not how they make their money. They make their money off the consumer plans like the $8 and $20 plan and selling API access to the labs, not from these $200 subs. This is not real profit for them. These are all about marketing. So, if this $500 plan did not help them with marketing, it failed. And I can comfortably say it failed. There was a bigger mistake though. And this is the one that's going to get a dedicated video in the near future. Ultrafast. All of the things I have said up until this point made sense for one reason. 16k is so cheap that cutting your cash value in half does not meaningfully decrease the amount of value you're getting from your subscription as long as you move to soul cuz the model is so much cheaper that 50% cut doesn't feel like a cut. Ultrafast is the opposite. Astra priced at a dollar per mil in cash, $10 per mil in normally, $50 per mil out. God forbid you have a bigger context window. Now you're paying double. Then you turn on ultrafast.
Watch from 24:50You're not hallucinating. That just went from $10 in to $60 in. The cash input just went to $6 in. Cash rights are now $75 and the output tokens are $300 and that's before the long context. God forbid you don't have a context window limit on which thankfully we all do in Codex. Input's now $120 in. Cash input's $12 in. Cash rights are 150 bucks and the output tokens are $450 per mil out. I had Astra on ultrafast review two PRs. It cost $600. $600. We're now in the category where it would have been roughly the same cost to pay an engineer. It would have been slower, but I could have paid him. So, Open AI, the same day, put out the Pareto frontier cheapest model for the value you get ever by far. Like, let me go back to Slopilitics and show you the Pareto frontier really quick. They own it. Luna and Soul are so deep in the Pareto line that you can't even see them when I highlight it cuz they're just the Pareto line. Opus has a section that isn't that you see highlighted there when I hover because low and medium are not worth using compared to just using 61 Soul.
Watch from 26:06So, they put out the cheapest model for that level of intelligence by far. Didn't put it on Cerebrus cuz that's how ultrafast runs. It's on different infra that is built for crazy inference. And it is incredibly expensive for them to run it. They just don't have the computer available. And it's like special chips from Cerebrus. So, they're trying their hardest to make enough to support this. It's crazy costs. I get it. But they put out the cheapest thing ever and the most expensive thing ever the same day. And the day before announced a cut a massive cut in how much cash value you get with your sub. This is all just pathetic. And the reason it happened or Astro ultrafast is Nvidia. I thought that they were used so they straight up just haven't gotten Cerebrus working. Have they just given up? Apparently, that's why they did this tweet because Cerebrus didn't work out for Astro ultrafast. Damn, I didn't even know that. That the cost makes even more sense now cuz that means that they are like massively over provisioning in order to make it that fast. God, yeah, no. The margins on the ultrafast stuff are Like, way, way lower. Apparently, Jane Street bought all the Cerebrus chips and that's why they aren't able to use them. Fun. Good to know. Regardless, this sucks and makes the value of that $500 plan feel awful because they're branding it as the plan with ultrafast. Ultrafast is an ultrafast way to burn through your limits. I did a bunch of math with my $500 plan using ultrafast. You get about $1,000 of ultrafast API value per week on that $500 plan, which means from our estimates 2.1 hours of generation. Yeah, my weekly limit can be killed in 2.1
Watch from 27:26hours. This was a mistake. They probably shouldn't have put it in the subscription at all. And combining all of these things at once, the release of 61 Soul with way lower margins that people don't understand, combine that with the limits being cut in the accounts so drastically before people understand this change, combine that with a new $500 plan that obviously people are going to feel pressured to upgrade to, combine that with ultra fast, which is the fastest way to burn the through that new plan, and any goodwill that could have been won by 61 Soul's efficiency has effectively been burned even faster than my limits. That's the problem. They did not release these things in an order that makes sense so that we could learn these changes. And I don't blame anyone for not understanding that 61 Soul has way lower margins. And I'm also not surprised that there are people who upgraded to this new plan excited for the ultra fast and then were blown away when their account was locked after like an hour or two of prompting. All of these things suck. All of these things hurt the trust that OpenAI has. And I'll be real, almost all of these things are OpenAI's fault. Doesn't mean I don't think they can come back, just means that Anthropic has put them in a tough place right now. The biggest benefit OpenAI used to have is that they were better at post-training in RL, so they could make smaller models that were really, really capable. Anthropic figured this out with the recent Sonnet and Opus releases, so that gap isn't really there anymore. And here we are.
Watch from 28:59I guess Sam was right. I'm an Anthropic fanboy in the end. But yeah, the value you get on your account is a lot lower. My ultra fast usage on the $500 plan gets me around $4,000 a month, whereas just a few weeks ago, my $200 plan used to get me $12,000 a month. So my price got two and 1/2 times higher and my usage got three times lower. That just feels bad as long as you're thinking of it that way. It feels bad. I went from months ago being able to use 5.4 or even 5.5 literally all day and if worst case I'd hit like 10 to 15% of my usage in that window to not even surviving half a day with my $200 or even $500 plans, it does feel bad. And Open AI needs to fix it. They need to fix it by making things that are actually better than Astra. So, there's no reason to reach for that model. They need to make things that are as capable as Opus 5.5 that are as pleasant to use, that are way faster than they have currently for reasonable prices that can feel unlimited in our subscriptions. And they need to never ever ever brand the $500 plan as the ultra-fast plan. That was a mistake from the front. I warned them ahead of time. I still think that was a bad decision. So, to answer the original question, does the $200 sub suck now? If you are thinking about it as the Astra plan, the way I used to think of Claude as the Fable plan, yes. If you're buying the $200 plan just to use Astra or you would buy the $200 Claude plan to use Opus and Fable, yeah, go get the Claude plan. But if you can do both, I personally get a
Watch from 30:22ton of value from having Claude spin up sole sub-agents to review its work because Open AI models I've still found are a bit more thorough with their analysis and will find things deeper than Anthropic models tend to. So, I do a wield and the vast majority of my Codex usage is either me using it to like do my email and computer you or it's coming from Claude calling Codex via T3 code in order to review Claude's work. And for those tasks, I have found it to be incredible and I get plenty of value out of my $200 subs. But the Claude plan is absolutely mandatory in my opinion if you want to ship serious engineering work regularly. It's just such a good value and you get so much. But I do suspect in the near future as Open AI and Anthropic get more and more competitive with enterprises and start driving those API prices down more and more, that the amount of subsidization we see here will go down and people will complain. There'll be lots of crazy graphs people are showing on Twitter how much worse the value it is, but I promise you you will still get more work done with the new models on the same plan in a year than you're getting from it today, or something went really, really, really wrong. Hopefully you now better understand the change to that $200 plan. Do I think everyone should rush to go sign up for it right now? No.
Watch from 31:52Do I think the Claude plan is a better deal? Honestly, kind of yeah. But I also think 6-1 is all is one of the best values in models ever and you shouldn't discount that just because your discount rate is a bit lower than it was before. Let me know how y'all feel and how you think I'm still an OpenAI hardcore fanboy after a video where I tear them apart like this. I I know how this one will go. But until next time, peace nerds.
Watch from 33:10Full transcript
- 0:00
Someday, far into the future, my children will be using their Samma coins to try and buy dinner, and I'll have to explain to them that when I was a kid, we used to use US dollars in order to get our Samma coins, and we used to pay $200 for $1,200 of Samma coins. I know I'm memeing pretty hard here, but seriously, just a few weeks ago, a $200
- 0:21
plan on your Codex sub would get you up to $12,000 a month in inference. And on the Claude plan, it wasn't quite as much, but it was still 200 bucks in and $8,000 out. We got here in a way that was honestly kind of understandable. The $200 plans were introduced to feel unlimited back in the day where we would send maybe 10
- 0:43
or 15 prompts on our heaviest days, and we would get a response back in under 5 minutes, usually just touching a few files here and there. And when we were all using Sonnet to explore our code bases, $200 plans were practically unlimited. But as models got more and more capable and more and more expensive, neither OpenAI nor Anthropic
- 1:03
wanted to take the optics hit of somebody being happy with how much usage they get for 200 bucks and just to have it stop feeling unlimited when the new model comes out. That has resulted in this crazy arms race where Anthropic and OpenAI kept subsidizing these plans more and more until we got to these crazy numbers. Right before Dev Day, Tibo made the biggest announcement that OpenAI has
- 1:25
made arguably this year for us vibe coders. He announced that they would be changing how they calculate usage for the $200 plan, in effect giving you half as many dollars as you would have had before. That's pretty bad by itself, but I ran the numbers. I only got $570 out of a week of usage, which adds up to
- 1:45
around two grand a month. So, is that the end? Did we really just go from 12K a month of usage to 2K and just smile and wave? Is there any reason to use the Codex sub and not the Claude sub, or is this just some secret plot to force us all onto the new $500 tier instead? And I can't wait to tell you those answers right after a quick break for today's
- 2:06
sponsor. Agents need access to the internet to do tasks well. It's just a fact. It doesn't matter how smart the model is if it doesn't know how to get to the internet and reference the things it needs, it's not going to do a very good job. That's why your agents need a great browser. Ideally, today's sponsor, Browserbase. These guys built the perfect browser for your agents so that they have access to the entire web. And I don't mean just the browser. I mean all the different APIs, search
- 2:26
capabilities, and more that your agents need to do all of this. None of that is my favorite part of Browserbase though. I've said for a while that Stagehand is the coolest thing they're building. If you're not familiar, Stagehand is an open-source SDK that allows your agents to do things on web pages with plain English instead of having to define every single element that they're interacting with. I know showing code isn't the cool thing to do nowadays, but
- 2:46
look at what I mean here. You can grab fields from the page by awaiting stagehand.observe description, like find the email input or find the password input. And once you've identified these fields, you can fill them with different variables and whatever else you want to do. This is so useful, especially when you combine it with their Zod implementation that lets you get validated data out of a page. This is
- 3:07
great for you and me writing the code, but let's be real. Neither you nor me are writing the code. That's why we built the demo with agents showing just how powerful this is. This demo is going to fetch the latest releases from these three different repos on GitHub. We click scan, we can see it opening and it's navigating in a real browser because this is actually controlling a browser. It's not just editing HTML for you. And now using those plain English
- 3:28
instructions, it's able to find the right page and to get the data we want. What's even cooler with this is since it's not fixed on specific elements, if the page gets updated in the future, you don't have to rewrite all your code to deal with one element moving slightly on the page. And now we have all the results. A nice little brief that shows all of the releases for the repos we put here. Make your agents more resilient and get them better data at so
- 3:49
dev.link/browserbase. So, what's going on here? If OpenAI wants so badly to be seen as a better value prop than Anthropic, why are they giving us so much less than they used to? This is just over, right? Like there's no reason to have the $200 Codex sub anymore, right? There's layers to this one. The biggest issue that we need to jump it on top of now is the gap in these
- 4:10
numbers. The number on the right seems just insane compared to the number on the left. They're basically giving us a crazy cheap number here. It's almost like made up. And when you compare that to what these models actually cost over API, it's just it's nonsensical. And when you look at my numbers for the last 7 days, it seems even crazier. I'm doing 20 grand of API spend in 7 days across
- 4:33
my machines. It's nuts. Obviously, I can't pay that. Like I'm doing okay, but I I don't have it in me to pay 20 grand a month to use these things over API. The reality is that not only is this $200 number made up, so is the API price on the right. Both of these numbers are fake, and neither of them represent the actual costs to the companies when
- 4:55
they're running these things. The number on the right is the number that they have chosen to charge enterprises hitting these things over APIs to see what they can get away with. And the margins on these numbers have historically been insane, upwards of like 95% against the hardware and the energy costs. Obviously, they had to put a bunch of money into training the model, hiring all the employees, doing
- 5:15
everything else. But the actual compute and energy costs for burning $100 in tokens is closer to like two to five dollars than it is to that $100 number. Since this number is so outlandishly inflated by these crazy margins, this number feels insane, and the gap between them feels even more insane. But in an ideal world with enough competition,
- 5:38
Anthropic and OpenAI would both have enough pressure to lower the number on the right that the gap between the subscription subsidization and just paying for the model for what it costs them, that gap would get collapsed relatively quickly. And it seems like we are at the start of that now. We will get to what I mean there in a second, but I also want to talk about
- 5:59
the weird things that made the Claude sub feel worse before. The first thing, the one that pissed me off the most by far, was that you could only use it in Claude code. The Claude sub isn't a sub to do $8,000 a month of Claude. The Claude sub is a sub to do a large amount of inference in Claude code itself. You can't use your off here other places.
- 6:20
You can use it in T3 code because in T3 code we are just calling Claude code directly through the agent SDK. But if you wanted to throw it into something like Pi or Open Claude or Hermes, you can't because those things aren't Claude code. And if you think you can be clever and add the instructions to work like a tool like Pi in your Claude code system prompt, good luck. The request will
- 6:41
fail. They are that petty at Anthropic sadly. I am hoping that they're going to lift their restrictions here in the near future simply because a lot more people are maxing out their subs. So it matters less if they're doing it inside or outside of Claude code. Regardless, they've been awful about that. OpenAI has been the opposite. You can use your Codex sub for basically whatever you want, basically wherever you want, as long as you're not using it to serve
- 7:01
user traffic. So if I put my Codex sub in the cloud and then charge you $5 to access my subscription, that's not allowed. But if I use my Codex sub to run on CI to review everyone's PRs when they come in, that is mostly allowed. Just want to draw the distinction here because what you're allowed to do with your Codex plan goes a lot further than what you're allowed to do with your Claude plan. Claude also has a
- 7:21
restriction on the usage of the biggest model. Fable usage cannot use your whole sub because Fable's price is not what it probably was meant to be. I have two sets of models on the screen here. Set one, Opus 4.5 to Opus 5.5 and GPT 5.6 Soul plus GPT 6 Soul. The second group is Fable 5 and 5.1 and GPT-6 Astra. What
- 7:44
is the difference between these two sets of models? It seems obvious, right? The bottom ones are the big expensive ones and the top ones are the small cheap ones. You're also noticing something weird. I said GPT-6 Soul, not 6.1. Which side do you think 6.1 goes on for the comparison I'm trying to make here? GPT-6 1 Soul is on the other side of this line.
- 8:04
And you might be thinking, "Oh, is that cuz it's a big model and the others are smaller? Are you saying 6 1 Soul is bigger and more capable than Opus or some other dumb thing?" I know how the comments tend to go. I'm putting 6 1 Soul here for a different reason. Remember Project Glasswing? AKA the announcement for Claude Mythos when Anthropic confirmed they had this crazy
- 8:25
model internally that they weren't going to sell? They actually were going to sell. They were doing this big commitment early, but they planned to continue offering Claude Mythos to participating companies at $25 per million tokens in and My assumption here is that these numbers were calculated using the previous margins.
- 8:48
If Anthropic was used to having a 95% margin on their compute, so they would spend $5 and have 100 come in for their power and compute costs, they planned to maintain those margins here. Let's just play with that 95 number. That would put the actual cost for a million tokens out, not at the 125 they claimed, but at
- 9:09
625. So, let's just like write out these numbers quick. So, say Mythos was going to cost 125 per mil out. And that would be $6.25 in cost. That keeps them at their 95% margin. But, you might be confused about those prices because that's not the price
- 9:30
Fable came out at. And Fable and Mythos are the same model. I know some people seem confused about that. I'm not going to entertain that right now. They are the same model because when Fable and Mythos 5 came out, it was a different price. Fable 5 and Mythos 5 are being offered at $10 per million in and 50 per mil out, less than half the price of the Mythos preview. It's not like they made the model smaller and dumber before
- 9:51
releasing it. They were very clear about that. They actually just made it bigger and more capable. So, again, we'll copy this and say Fable 5 {slash} Mythos 5, price is now $50 per mil out, but it's the same model size, same hosting. Which means that their costs probably didn't go down. Which means their margins have dropped pretty massively from 95% to
- 10:13
87.5%. If you were to look at this as the profit that they're getting at a given time, it's even more brutal. Where in this instance, they would have profited $118.75, here they're only profiting $43.75. The margin drop is big, but the return drop way bigger. They used to spend 625 in
- 10:37
compute and make $118. Now, they spend the same in compute and they make $43. The line I drew here is the start of the end. What is it the start of the end for? The subsidization. Anthropic chose with Fable 5 and Mythos 5 to price it way cheaper than they
- 10:57
originally planned in order to just really make Open AI's life harder. And they succeeded with that. That was the start of the panic at Open AI, for sure. And even if the change in margins doesn't seem that much, the change in the revenue is world's difference. And if you bump that up to what was possibly a 98% margin before, these numbers get
- 11:18
even crazier. The point I'm trying to make here is that the amount of revenue Anthropic makes per amount of energy and compute used has gone down meaningfully with Fable. But they chose to do that because the margins are still insane. Like they're still just printing money with this. It's just that the amount they're printing has gone down significantly. And this also forced OpenAI to do the same. I don't know what
- 11:38
their planned price was for Astra, but they released it at the exact same price as Fable. So, we have no idea like what they would have charged before Fable because they made the model with a certain price target in mind probably. But, they did a few other things on top of that in particular with the subs that made it really brutal. They took this hit on the cost, so they might have intended to charge 100 bucks per million
- 11:59
output. Instead, they're only charging 50, so their margins are worse. Meanwhile, Anthropic doesn't even give you your full usage of your sub with Fable. You only get half your allocation for Fable. The other half can can be used for whatever other models, but you're not using 100% of your usage with Fable. It's limited massively. OpenAI chose to not do that, and they let you
- 12:20
use Astra for your whole weekly limit, not just a portion of it. And if you're curious where that 50% cut came from with Anthropic, it came from them not doing it for the 3 days before the ban, reading all the data, and also here. It's not exact, but $43 is half to a third of $118. They're profiting half
- 12:41
as much, so they give you half as much. Makes all the sense in the world, right? Let's keep going down this rabbit hole a bit because there's one other model here that does not seem to fit that well. 6.1 Soul. Considering the way things are currently priced in the industry, my honest assumption would be that 6 Soul
- 13:02
was kind of meant to replace Terra because nobody used Terra. And eventually, they would do something more like Opus between Soul and Astra. They didn't do that. In fact, they spit out 6.1 Soul suspiciously close to 6 Soul, which suggests that this was not necessarily what the model was meant to be. As such, I'm going to pretend that
- 13:23
they were looking at the existing numbers for the competition cuz we've now established, hopefully, that OpenAI's pricing is based less on their margins and more on their competitors because they set Astra's price to be literally identical to Fable. So, they were probably planning when they were training to make 6.1 Soul a little closer to Opus. Not Opus 5.5 cuz it didn't exist yet, but for Opus 5, which
- 13:45
was $5 per million and $25 per million out. We will assume these are the numbers that they were operating with. The new model they were planning was going to be $5 per million, 25 per million out, nothing special with cash discounts. We'll assume the same 95% margin. That means for $25 out, it would
- 14:05
cost them around $1.25, which gives them $23.75 profit. Not bad. You spend $1.25, you get $23.75 out. Pretty good deal. But that is not the price that this model came out at. 6.1 Soul came out differently. It came out. It's
- 14:25
$2 per million, $10 per million out, and massive cash cost decrease. The cash reads are half as expensive as they were before. They used to be 20 cents cuz it would be a 90% discount. So, it would go from $2 per million to 20 cents per million. It's now 10 cents per million. Doesn't sound like a big deal. 20 cents to 10 cents, 90% to 95%,
- 14:46
but when you look at the percentage of your cost and see that for OpenAI models, half your cost is cash reads, now it's 25% of your cost because you just took half your cost and you cut it in half. So, it would have been 50 cents per million read from cash is now 10 cents per million read from cash. That's a 5x decrease from what they had originally planned
- 15:07
here. Crazy. So, the results here is that your actual effective costs using 6.1 Soul aren't just 2.5x cheaper than they expected. It's probably closer to 5x cheaper than they expected. What I'm trying to say is their margins for this model are garbage. This 95% no, no. I would legitimately guess their margins for this model are like in the 50%
- 15:27
range. Just seeing how much slower it is than 6 soul, it was definitely planned to be 525. So, 210 with the 10 cent cash reads is just insane. So, they destroyed their profit margins. Like they're still making money. I'm not going to say like 50% profit margin is bad, but it's pretty obvious that they aren't just printing money on the inference the way they were before. Remember earlier I was
- 15:48
talking about that 50% limit for fable? How they cut your limit in half when you were using fable models? Open AI had said they didn't want to put an arbitrary limit on Astra. That's what happened here. But it would have been the other way. Since the model they cut the subsidization for the most was 61 soul, they didn't want to put a custom soul limit. They could have done that. Open AI absolutely could have put in a
- 16:10
custom limit that was your soul only counts for half your usage. But instead they're taking the opportunity to try and collapse this gap in the subsidization a bit so that your account sub doesn't feel quite as insanely subsidized compared to paying API prices. I don't think they should have announced this the time they did. I love the attempts at transparency from
- 16:31
Open AI. They were pretty transparent here, but this did not come off great, especially because so many people are still using Astra. But this is where we have to dig into something a little more painful, which is the effective costs. I ran a bunch of real numbers for my day-to-day usage. Previously, I was seeing similar to those 12K numbers before. Especially when you combined the
- 16:52
fact that we got a reset every 3 days on average many months, you were getting 24 grand if you were truly maxing out every single window that you had. Now it's closer to $2,500 a month instead of $24,000 a month. Roughly a 10X drop. Because I only saw $573.86 of usage, and I was able to kill a whole
- 17:14
account in 29 hours in a loop. Not great. But let's scroll down a little cuz I have things I want to talk about in this. My usage. My GPT-6 Astra usage was for around 1,800 responses. In in these 1,800 responses, it did 766,000 output tokens and it cost $534. It was API prices. 61 Soul did
- 17:37
almost as many responses. It's a little less, but it was like 1,500 versus 1,800. It's not that big a gap. More input tokens, fewer cash input tokens. I think this might have been a bug in one of my setups or something. I haven't figured out, don't care to. Again, very similar number of output tokens. $38.27. Do you understand? Roughly the same amount of usage, greater than a 10x decrease. The reason they made these
- 17:59
changes, all of these changes, is they wanted to make 61 Soul a great value. They wanted to make this $200 sub feel unlimited again, the way it used to, all the way back in the 54 and 55 days. And they wanted to make sure they didn't go bankrupt because of the insane amount of compute that you're getting for that amount of money. The reason they had to
- 18:20
do that is cuz of people like me and a lot of y'all who are token maxing right now. Since I use these subs and I push them to their absolute limit, if they didn't make these changes, I would be burning two to four times more compute from OpenAI every month and that's just insane cuz you can absolutely, and I do genuinely believe for any even vaguely real-world use case, you're going to get
- 18:41
practically unlimited use of a $200 plan as long as you only use 61 Soul and you never touch Max cuz Max makes no sense at all. Do I think this is a good change though? Meh. I'm going to be so real with you guys. There is more to talk about here, especially the $500 plan. I'm going to be incredibly real and honest with y'all. If your goal here is to use your subscription to write code that you plan
- 19:03
to ship, the insane value you get out of Opus 5.5 on the $200 Claude plan is impossible to beat. I thought I would cancel subs with Claude because it was so much more efficient, but I ended up just pushing my own limits to what I could do, and I ended up with a six Claude account. And if I refresh now, it's probably closer to drained. Yeah, like I have killed three of my six. Two
- 19:23
of them were dead this morning and just reset. I'm pushing these things. I am very much getting more value than I would ever have expected out of my Claude plans. It's crazy. And the craziest part here is I have not touched any of my fable. I have 600% of my fable. I can't use it on the dead accounts because that gets combined with your weekly. So, I can't use it in those
- 19:44
accounts. But, it's hilarious that I've just straight-up not touching fable anymore. So, I still think the Claude plan is the better value by far, but I also think 61 Soul is the best value a model has ever been by far. If I was forced to pay API prices, I'd be picking Soul without any question. But, I'm not. You get less subsidization from the Codex plan, but the model's also so much
- 20:05
cheaper that it balances out relatively well. And I'll show you this with a new project that I actually built with Astra Ultra fast, funny enough, Slopolitics. I wanted it to be easier to visualize the score against the dollar spent for these benchmarks. And you can see very clearly here, 61 Soul is an absurd value on the Artificial Analysis Intelligence Index,
- 20:25
which is what this data is from. If I switch to the linear view, it's even funnier looking. Truly hilarious. I also ran this on Terminal Bench myself because I was not happy with how others ran Terminal Bench. And not only did 61 Soul get the best score on it when you gave it access to its native harness, I ran it with Codex and Claude code, so I got better numbers than a lot of others did. Not only was the 61 Soul max run
- 20:46
the best score ever, the 61 Soul X high run pretty much exactly tied Opus 5.5, but Opus cost $13.11 per task. 61 Soul cost 96 cents per task. A 13th the price. Do you understand? Your subscription might not be as much dollar
- 21:07
value, but the number of tasks you are completing with that money is way higher. So, if you're paying API prices and you're paying Opus, you better be damn certain that Opus is the only model that does what you need because if you're comparing the cash value you get in the subs by the pay to tokens, none of this ends up making sense. And this
- 21:28
is the problem is we have two levels of like subsidization happening here. We have the amount of dollars you're getting for your $200 sub and we have the cost that those dollars represent to the lab and we have how many tokens it take to solve a problem because if I switch this to show per tokens, which I have in Slop lytics, so do I have it No, I have tokens for task here.
- 21:49
You'll see that medium on Opus 55, sorry, high on Opus 55 uses roughly as many tokens as Max does on 61 Soul. And the tokens are cheaper. So, cost per task is what we should be looking at and real-world cost per task on something like Terminal Bench 4 is a 13x decrease. So, the dollar value of the sub being
- 22:11
half as much ends up still being five times more tasks depending on the work you're doing. So, again, I don't think 61 Soul is the better model. I would take my Claude sub over my Codex sub any day, but this isn't some crazy rug pull like everyone's making it out to be unless you insist on continuing to use Astra with the much higher cost and much higher spiky nonsense rates. But, here's
- 22:34
where I have to stop being nice to OpenAI. Cuz I'm about to crash out really, really hard. And everybody who's accused me of being paid by OpenAI to say nice things, this is going to prove that I'm not, which means I do need to be paid, so we're going to do one more quick sponsor break first. If I told you that you could 5x your potential users with one simple trick, you'd probably say
- 22:54
that I'm insane and I understand because you probably didn't know about general translation. These guys are incredible. They made it so easy to translate real-world applications across various different surfaces. There are so many little edges that are hard to get right when you build translations into your apps. Localization is just not an easy problem to solve. I had to do it myself at Twitch and it was miserably
- 23:16
difficult. I'm very thankful for my team of much smarter engineers for helping us all through that. Going to do a thing I don't get to much anymore and show you some code because it really shows how magical what they built is. In React applications, you use their helpers in your JSX to actually indicate what section to be translated. They also have helpers for handling annoying things like number formatting and pluralization, which is not trivial in lots of different languages. Where
- 23:37
things get much more powerful is their dictionaries that allow you to have consistent translations across various surfaces. So, if you have a feature with a specific name, you don't have to worry about that feature being named incorrectly in three different places. Now, you can finally know for a fact that the same translation and the same verb will be used across all these different surfaces. We tried this out for the T3 code marketing site and when I say us, I really mean our agents
- 23:57
because agents are really, really good at using general translation. And after two prompts, we have everything we need to showcase our application in all sorts of different languages. And the best part is we can share these definitions across all of our services, whether it's our mobile app, our desktop app, our website, and more. Reach the 85% of the world that doesn't speak English at swyd.link/gt. Time for the crash out.
- 24:20
I'm going to do more of a detailed version of this in a future video, but I'm going to give you the CliffNotes of it now. The $500 plan was a mistake. Not the existence of a $500 plan. Doing it right now. Absolute optical show. To tell people that the 20x plan is now 10x, but don't worry, you can get the 25x plan if
- 24:40
you pay 2.5 times more, pisses everyone off, myself included. This was just a mistake. If they were going to do this, which obviously they had to eventually, they should not have done these things at the same time. It is such an unnecessary self-inflicted L that if I defend anything they do, you guys are going to call me a pay shill
- 25:01
and I understand because this is such an absurd blunder optically that it makes them look insidious. The reality is that these subscription plans are not how they make their money. They make their money off the consumer plans like the $8 and $20 plan and selling API access to the labs, not from these $200 subs. This is not real profit for them. These are all about marketing. So, if this $500
- 25:22
plan did not help them with marketing, it failed. And I can comfortably say it failed. There was a bigger mistake though. And this is the one that's going to get a dedicated video in the near future. Ultrafast. All of the things I have said up until this point made sense for one reason. 16k is so cheap that cutting your cash value in half does not meaningfully
- 25:43
decrease the amount of value you're getting from your subscription as long as you move to soul cuz the model is so much cheaper that 50% cut doesn't feel like a cut. Ultrafast is the opposite. Astra priced at a dollar per mil in cash, $10 per mil in normally, $50 per mil out. God forbid you have a bigger context window. Now you're paying
- 26:03
double. Then you turn on ultrafast. You're not hallucinating. That just went from $10 in to $60 in. The cash input just went to $6 in. Cash rights are now $75 and the output tokens are $300 and that's before the long context. God forbid you don't have a context window
- 26:25
limit on which thankfully we all do in Codex. Input's now $120 in. Cash input's $12 in. Cash rights are 150 bucks and the output tokens are $450 per mil out. I had Astra on ultrafast review two PRs. It cost $600.
- 26:48
$600. We're now in the category where it would have been roughly the same cost to pay an engineer. It would have been slower, but I could have paid him. So, Open AI, the same day, put out the Pareto frontier cheapest model for the value you get ever by far. Like, let me go back to Slopilitics and show you the
- 27:08
Pareto frontier really quick. They own it. Luna and Soul are so deep in the Pareto line that you can't even see them when I highlight it cuz they're just the Pareto line. Opus has a section that isn't that you see highlighted there when I hover because low and medium are not worth using compared to just using 61 Soul. So, they put out the cheapest model for that level of intelligence by far.
- 27:30
Didn't put it on Cerebrus cuz that's how ultrafast runs. It's on different infra that is built for crazy inference. And it is incredibly expensive for them to run it. They just don't have the computer available. And it's like special chips from Cerebrus. So, they're trying their hardest to make enough to support this. It's crazy costs. I get it. But they put out the cheapest thing ever and the most expensive thing ever the same day. And the day before announced a
- 27:51
cut a massive cut in how much cash value you get with your sub. This is all just pathetic. And the reason it happened or Astro ultrafast is Nvidia. I thought that they were used so they straight up just haven't gotten Cerebrus working. Have they just given up? Apparently, that's why they did this tweet because Cerebrus didn't work out for Astro ultrafast. Damn, I didn't even know that. That the cost makes even more
- 28:12
sense now cuz that means that they are like massively over provisioning in order to make it that fast. God, yeah, no. The margins on the ultrafast stuff are Like, way, way lower. Apparently, Jane Street bought all the Cerebrus chips and that's why they aren't able to use them. Fun. Good to know. Regardless, this sucks and makes the value of that $500 plan feel awful
- 28:34
because they're branding it as the plan with ultrafast. Ultrafast is an ultrafast way to burn through your limits. I did a bunch of math with my $500 plan using ultrafast. You get about $1,000 of ultrafast API value per week on that $500 plan, which means from our estimates 2.1 hours of generation. Yeah,
- 28:56
my weekly limit can be killed in 2.1 hours. This was a mistake. They probably shouldn't have put it in the subscription at all. And combining all of these things at once, the release of 61 Soul with way lower margins that people don't understand, combine that with the limits being cut in the accounts so drastically before people understand this change, combine that
- 29:17
with a new $500 plan that obviously people are going to feel pressured to upgrade to, combine that with ultra fast, which is the fastest way to burn the through that new plan, and any goodwill that could have been won by 61 Soul's efficiency has effectively been burned even faster than my limits. That's the problem. They did not release these things in an order that makes
- 29:38
sense so that we could learn these changes. And I don't blame anyone for not understanding that 61 Soul has way lower margins. And I'm also not surprised that there are people who upgraded to this new plan excited for the ultra fast and then were blown away when their account was locked after like an hour or two of prompting. All of these things suck. All of these things hurt the trust that OpenAI has. And I'll
- 30:00
be real, almost all of these things are OpenAI's fault. Doesn't mean I don't think they can come back, just means that Anthropic has put them in a tough place right now. The biggest benefit OpenAI used to have is that they were better at post-training in RL, so they could make smaller models that were really, really capable. Anthropic figured this out with the recent Sonnet and Opus releases, so that gap isn't really there anymore. And here we are.
- 30:22
I guess Sam was right. I'm an Anthropic fanboy in the end. But yeah, the value you get on your account is a lot lower. My ultra fast usage on the $500 plan gets me around $4,000 a month, whereas just a few weeks ago, my $200 plan used to get me $12,000 a month. So my price got two and 1/2
- 30:42
times higher and my usage got three times lower. That just feels bad as long as you're thinking of it that way. It feels bad. I went from months ago being able to use 5.4 or even 5.5 literally all day and if worst case I'd hit like 10 to 15% of my usage in that window to not even surviving half a day with my $200 or even $500 plans, it does feel
- 31:04
bad. And Open AI needs to fix it. They need to fix it by making things that are actually better than Astra. So, there's no reason to reach for that model. They need to make things that are as capable as Opus 5.5 that are as pleasant to use, that are way faster than they have currently for reasonable prices that can feel unlimited in our subscriptions. And they need to never ever ever brand the $500 plan as the ultra-fast
- 31:26
plan. That was a mistake from the front. I warned them ahead of time. I still think that was a bad decision. So, to answer the original question, does the $200 sub suck now? If you are thinking about it as the Astra plan, the way I used to think of Claude as the Fable plan, yes. If you're buying the $200 plan just to use Astra or you would buy the $200 Claude plan to use Opus and
- 31:47
Fable, yeah, go get the Claude plan. But if you can do both, I personally get a ton of value from having Claude spin up sole sub-agents to review its work because Open AI models I've still found are a bit more thorough with their analysis and will find things deeper than Anthropic models tend to. So, I do a wield and the vast majority of my
- 32:07
Codex usage is either me using it to like do my email and computer you or it's coming from Claude calling Codex via T3 code in order to review Claude's work. And for those tasks, I have found it to be incredible and I get plenty of value out of my $200 subs. But the Claude plan is absolutely mandatory in my opinion if you want to ship serious
- 32:29
engineering work regularly. It's just such a good value and you get so much. But I do suspect in the near future as Open AI and Anthropic get more and more competitive with enterprises and start driving those API prices down more and more, that the amount of subsidization we see here will go down and people will complain. There'll be lots of crazy graphs people are showing on Twitter how
- 32:51
much worse the value it is, but I promise you you will still get more work done with the new models on the same plan in a year than you're getting from it today, or something went really, really, really wrong. Hopefully you now better understand the change to that $200 plan. Do I think everyone should rush to go sign up for it right now? No. Do I think the Claude plan is a better
- 33:12
deal? Honestly, kind of yeah. But I also think 6-1 is all is one of the best values in models ever and you shouldn't discount that just because your discount rate is a bit lower than it was before. Let me know how y'all feel and how you think I'm still an OpenAI hardcore fanboy after a video where I tear them apart like this. I I know how this one will go. But until next time,
- 33:32
peace nerds.
Description
OpenAI's $500 plan's new Ultrafast mode means your weekly limit can disappear in 2 hours, so does that just mean the $200 tier is useless now?
SponsorThank you Browserbase & General Translation for sponsoring!
Browserbase: https://soydev.link/browserbase
General Translation: https://soydev.link/gt
Want to sponsor a video? Learn more here: https://soydev.link/sponsor-me
Check out my Twitch, Twitter, Discord more at https://t3.gg
S/O @Ph4seon3 for the awesome edit 🙏
#ai #coding #programming
