Nobody Is Competing With Anthropic Transcript, AI Summary & Key Points

00:00 I've been waiting for this release for far too long. I couldn't be more hyped. After Opus 5.5, Sonet 5.5, and Haiku 5.5, it just kind of seems like the 5.5 family is unstoppable. And that's why everyone is so hyped about today's release. Fable in Mythos 5.2. Huh? Thought we were waiting for 5.5. How do you go backwards when we were at Okay, this is going to be a more complex video than I was hoping it would be.

00:26 As y'all know, I was a huge fan of Fable 5 and even more so of 5.1 and almost seemed like I was the only person who noticed that Fable 5.1 was a really good model. And then Opus 5.5 came out and showed everyone why Fable 5.1 was so good by making it faster, cheaper, and slightly less annoying when it talks to you. And I was so excited to see those wins in Fable 5.2.

00:47 And let's be real, we are seeing a lot of them. Whether it's the state-of-the-art score on Artificial Analysis, putting it a full four and a half to five points ahead of Gemini 4 Carbon, which was the highest before. Mind you, high and X high also scoring way better. Nuts. Doing CAD at a level that we didn't think was possible for agents. Or my new personal favorite bench, M D-Bench, where I measure how many M dashes are in the announcement post.

01:13 As you can guess, this was state-of-the-art at 29 M dashes in the post. Jokes aside, there is so much to talk about here. I did my best to cover everything people ask about. So much so that it is currently 1:15 in the morning as I film this because I wanted to get this video out for you guys with as much detail as possible. I'm going to show you guys what this model does best in class, some of the places where it's been frustrating me, how much it costs for real, how much it'll use your subscriptions for real, which I

01:36 did a ton of analysis of specifically, and so so much more. This is an incredible model, but it really should be used specifically and carefully. I'll do my best to explain why after I cover the insane amount of tokens I spent today with a quick word for today's sponsor. Y'all mind if I brag a little bit for a second? See this number on the screen? 10 minutes and 42 seconds.

01:56 I have put a lot of work into that number. That is the time it takes for a complete release of T3 code to be built in rci. All the shorter numbers are checks to see if we're due for a new release or not. Ignore those. This is the number that matters. I have went to hell and back to get here. I built my own custom Chrome extension to make a better viewer for RCI so that I can see what's going on and where the bottlenecks are all on the fly as things happen.

02:19 And I've consistently found one theme in my pursuit of maximizing speed in CI. All of the slow jobs have something in common. They weren't using today's sponsor, Blacksmith. Seriously, it's insane how much faster Blacksmith has made things for us. The official Windows ARM GitHub runner was so slow that it was doubling our build times. The Blacksmith Linux runners are so fast that I was able to emulate the Windows environment in my Linux runner and still make it four times faster.

02:46 I was able to make even more improvements using Codesmith, their built-in CI expert agent. Not only can it speed up your CI, it can make it cheaper by finding boxes that are too big for the task at hand and massively decreasing your costs. To be fair, moving to Blacksmith will also decrease your cost because their prices are insane. Their machines are half the price of GitHub, and since they're also faster, you end up spending less overall.

03:09 You have no idea how much your CI is bottlenecking you until you fix it. Get it fixed now at soy.link/blacksmith. Let's start with the M dashes from Hellpost, which also, by the way, is quite bright, so early apologies for the flashbang warning. We've upgraded Fable and Mythos. Mdash models that are designed to tackle the most daunting technical challenges.

03:28 Technical challenges is a very specific way of framing things here that I will overread much like how I will overread the 5.2 instead of 5.5. Fable and mythos 52 show improvements in areas like coding, reasoning, and vision. During pre-release testing, they've helped their engineers complete otherwise infeasible software migrations, designed three working electronic devices from scratch, solved two open problems in the physics of black holes, made substantial progress on robotics evals, and formalized the proof of

03:57 Fermat's last theorem in just 30 hours, mdash, among many other applications we'll discuss below. You might have noticed that in this set of things, there was only one thing about software infeasible software migrations. Well, as you guys have seen, I have done some very infeasible software migrations already with Opus, like rewriting TypeScript in Rust.

04:17 So, this may or may not be a novel unlock. We will see some other examples as we go. They call out something that I have to remind you all of a lot. Fable and Mythos are the same underlying model. The way I have framed this is imagine there is a restaurant with two entrances. One entrance has a bunch of guards that make sure you are who you're supposed to be, that you're allowed, and all these other things.

04:38 The other entrance is much more lightly guarded, but you have to have special access to get there in the first place. That is what is different here. The restaurant inside is the same. The food inside is the same, but who's allowed in and out and what's allowed in and out are different. So maybe if you come in from this entrance, you're not allowed to take the food out with you, or only certain things are allowed to be ordered on the menu.

04:56 It's still the same restaurant, though. Yes, there are some things one can get that the other can't, but it's the same building. The models are the same. Mythos just is less likely to have a third party around it block the request or the response. Babel has strong cyber security and biology safeguards to guard against misuse, which allows us to make it available to all users.

05:17 Organizations pursuing beneficial applications in cyber security or biomedical research can apply for access to Mythos through their expanded life sciences and cyber verification programs. Cool. Awesome. I'm not even in those. Testers have found that Fable 52 and Mythos 52 communicate more clearly than past Fable and Mythos models. They generate outputs up to 30% faster than Claude Fable 51 and Mythos 51.

05:40 This is another one of those very interesting details that I will be massively overreading in a little bit. They're also cheaper, too. $8 per million input tokens and $40 per million out. Fable and Mythos 52 come with an optional fast mode. Another big change. Previously, Fable and Mythos had to run at their stock speeds. You had no way to go faster.

05:56 Now you can pay twice as much money in order to get it 2.5 times faster. Not actually that bad, especially considering how much less a bump normal fast mode is from OpenAI. You have to do ultra fast for a real bump. And then the price increase is so insane, it's like just not really worth it. They have a section about safety and alignment. What a surprise.

06:13 They care a lot about people not distilling their models and using them to do bad things. There's a very long detailed system card as always. I did a quick audit of it. Admittedly, not quite as deep as I normally go cuz this one was 264 pages. Yeah, I will continue to audit this in the future and see if there's any really big things that we missed, but I want to do a more general overview for this video.

06:37 On that note, if you're here looking for when to use Fable versus Opus versus Soul versus Sonnet versus Haiku versus Luna versus Terra versus Astra, if Astra is even a thing anymore, make sure you're subscribed cuz I don't have time to do all that in this video. Anyways, they have their own real brief when to use Fable and Mythos. So, I'll let you have that much at least.

06:55 They are the best for the most ambitious and open-ended technical work. For most coding and knowledge tasks, they recommend starting with Opus 55, which is their strongest daily driver, and then switch to different models as required depending on the scope and complexity of the work upward to Fable or Mythos or downward to Sonnet or Haiku. Fable 52 also can play the role of adviser, where other models ask it for help on hard problems, or orchestrator, where it can delegate simpler tasks to smaller models.

07:17 Cool. Awesome. If you want a tool that lets you do that across all the different models, I heard there's some YouTuber that sloped together something you might want to check out. We'll go much more in depth on where things should and shouldn't be used. I don't fully agree with this advice, having used the model all day. The one piece I absolutely agree with in here, which we'll talk a lot more about in other videos, is the idea of the strongest daily driver being Opus 5.5.

07:40 All I'll say for now is that I agree beyond hard about that. Something I've gotten out of this release is a deeper appreciation for Opus 55, not because it is smarter, but because of the relationship between these models and how unbelievably capable it is. On that note though, let's look at some benchmarks. My face is going to cover the word Gemini 4 carbon in the corner.

07:59 Not that you have too much reason to look there. And we can see here Fable got the best score on Agent Coding with Frontier Code 2, Frontier 2, crushing most other things. GBD6 Astra got a 65, it got a 70. GB6 Astra got a 42.8, eight and it got a 49.1 business workflows. It got the second highest after Gemini 4 carbon. Interesting. It won Agentic Science Research and this is a big win by the way.

08:24 It's a huge jump from Fable 51 which is a 52.6. It got a 74.6. Also a massive jump from Opus. But most notably the gap between Opus 55 and Astra here was massive. It was 10 points. Table 52 has now gone even beyond Astra from 68.1 to 74.6. It seems like a lot of the value in this model is in the sciences outside of computer science. Like they went super hard on bio on the types of chemistry and things in terminal bench science.

08:53 They're going harder on the sciences and it shows in these numbers. Cyber secbench it also crushed. They switched to mythos for that because of course as you guess normal fable will not allow for those tasks. Cyber security in the CVE rediscovery it was able to get around a 90%. protein design, it got an 85.1 again with Mythos because Fable will not do these things.

09:13 Win win. And then in the ELO in knowledge briefcase, it did slightly worse than Opus 55. Still an allout win for Anthropic. And then win win and win. Yeah, this is one of the most utter dominations of a benchmark suite I've seen in a while. But don't worry, there are other benchmarks, many of which I've made, and it does not do quite as well in mine.

09:38 Yeah, I have some benches that I've been talking about in other videos, specifically in my Gemini 4 argan video where I do things like have the model analyze a complex codebase. In this case, I have it analyzing TS Rust, T3 Code, Lakebed, reviewing a bunch of PRs, stuff like that. And I found that Fable 52 is actually underperforming Opus and Sonnet and massively outperforming OpenAI models in this particular eval set.

10:03 And it's not just in one of them. It's in almost all of them. It goes quite a bit below other things. What is it that makes it score so much lower? I have theories. The simplest I could put it is that Fable has that the wise owl problem where it's trying to be clever a lot and that has its strengths. It does things that other models don't as a result.

10:28 And it seems like it benches really well in other more traditional benches. It's hard to describe, but the the word that comes to mind is spiky, which I hate because I normally reserve that word for open AI models, which spike into utter unbelievable levels of stupid. Fable is more like the the smart person on the team that occasionally will not understand the full situation.

10:51 They'll be in the meeting for the whole meeting. They're only hearing half the words, then they'll just say something kind of weird in left field. And most of the time it makes some sense. Some of the time it makes literally none at all. and you question if they were even listening and then that small remaining sliver it's suddenly realized it's like oh that's super genius that's so much better than what we had in mind this model covers the whole whole range there believe me ah what else do we got in this post we got

11:16 some fun views with the different benchmarks they have here Frontier code where you see some interesting changes in particular X high on Opus and high on Fable 52 are not too far off in cost X high is the highest score crushing out Opus 55 on Max at not that big a price gap. However, Max is worse. This is the first of the many lessons of Fable we will be learning today.

11:42 If you are the type of person to watch a video like this, a video about getting the most out of your models as a developer from a channel that was originally a webdev video channel, you are the type of person that is using these for code. Max reasoning effort has always been an iffy balance for code where not only is it much more expensive, it often will go too far and make mistakes as a result.

12:02 This is now encoded in the benchmarks. Fable 52 Max will score worse on most coding tasks consistently and you should not use it as a developer. If you think that you know better than me, there is a high chance you are the exact type of person that should not use Max. If you are begrudgingly using it because you have very specific things where the model is stopping too early and your attempts to prompt out of that are not succeeding and you are actively evaling against smaller reasoning efforts, maybe you can use it.

12:32 But like if they're going to go so hard as to make you click a button in settings to enable the full access mode, they should do the same for max mode because it's as if not more dangerous and significantly more expensive. They show two others here where max does slightly better. What I find really interesting is how closely aligned the prices are on Fable 52 high and Opus one tier higher.

12:53 So Fable 5 High, Opus 55 XHigh, almost nearly aligned on price. Same with Fable 52X high and Opus 55 Max. But again, max mode is not a big benefit. It's better if you just don't use it. They also snuck in bench CAD here, which is for hardware engineering, and saw a pretty massive gain. This has been interesting. I've been talking to a lot of my electrical engineer friends.

13:13 It seems like we're now at the moment where the benefits we've seen for code and software are being seen more in hardware and electrical engineering. This is going to be one of those things that ramps really fast in the very near future. I believe here is where things start to get strange. The coding section where they say all of their code is written by claude like it is for us.

13:30 Like the vast majority of code in T3 code like 95% plus is cloud at this point which is why they so excitedly say all of theirs is oh over 90% of the changes merged into anthropics codebase are written by cloud. 90% a lot of the people in this audience watching this video are landing more PRs with Claude than Anthropic is here allegedly. For what it is worth, I have heard from Boris in some interviews and other places that the other 10% isn't not Claude.

13:55 It's not properly attributed and they'd rather give the low number than make the bold claim after the amount of times they've been burned by saying we're all just writing everything Claude. The things that they said 6 months ago that we all made fun of them for are now true for many businesses outside of anthropic. And now their claims feel reserved, which is really funny.

14:13 Team of agents running an earlier version of 5.2 migrated a critical production service from Python to Rust. The service is now more efficient, more reliable, and it starts 26 times faster. Wow. I rewrote a language in Rust using Opus, and it was 20 to 100 times faster. It's kind of crazy that these things that were so bold not long ago don't even phase me at this point.

14:33 Like the next one's hilarious. As an experiment, one anthropic engineer gave Fable 52 the task of designing a prototype operating system. Wow. Operating systems, those are basically not touchable now. Crazy to think an AI could do something as impossible as an OS. T3 OS is coming soon, by the way. Very excited about that project. It's actually way cooler than I expected.

14:54 To be fair, it is Ubuntu based. It's not a new OS. The point I was trying to make here though is I assumed Opus could do this. Hala assumed Astra or Soul could have done this. Like drivers are a relatively solved task and everything else has enough references. Like I'm not super surprised that it was able to in 8 days make an OS that had drivers display out working and enough network to have a little claude widget built in.

15:17 Cool. Good work. Somehow not that crazy impressive to me anymore. It's it's insane how fast we got used to this. like the this frog got boiled fast. Here we have a little science and engineering section, but I would like to remind you before we go into these boring details that the end is what a lot of y'all are here for, which is my fun code demos, the fish slop run, cuz of course we're going to run this through fish slop.

15:38 Come on, you guys know me at this point. We can't not do that. And my honest thoughts having coded with this all day and landed like 20 or so PRs with it so far. They might change over time, but I spent from 11:00 a.m. when this dropped until 1:00 a.m. when I'm filming testing the hell out of this. I think I have some takes on it now. Obviously, not quite as deep as they'll be in a few days, but it's a good model.

15:59 So, let's hear about the physics side. It was able to solve two open problems in theoretical physics, both around black holes in their effort to unite the two lynchpin theories of modern physics, which are relativity and quantum mechanics. Both rely heavily on holographic duality, the idea that certain universes contain gravity, which can precisely be described by a quantum theory containing no gravity at all.

16:19 In the first problem, physicists in 1983 had shown that as these universes heat up, black holes form abruptly. Um, I'm sorry. I couldn't stay awake in physics class. It's just it's not my thing. I prefer software. Oh, it also does protein folding. Awesome. There's been a lot of advances in protein folding. It's cool to see more of these. The more proteins we can fold, the more cures we can have to the more things.

16:42 We are very likely close to a cancer cure from AI and I am curious to see what that does to the world. I was also able to do a formalization of fermens last beerum which uh caused a lot of drama on the internet. Cool. It's good at math. Crazy that this doesn't even really phase us at this point. Got way better at robotics which is scary but quite exciting.

17:01 Both building its own robots and also controlling said robots. Absolutely slaughtering in Robo-Dogjo which is a benchmark on how well the models do with robotic tasks. You got research and analysis where a bunch of people said it's really really good at doing analysis of all sorts of ambiguous data heavy tasks. I have had mixed results with this both in using Fable 52 personally myself today but also with the Fable models in general.

17:24 I have found that they have more novel ideas on what to do with the data but they also make jumps sometimes that don't make real sense and they forget to check timelines outside of the data set to get a better idea of like confounding events. Like I've had to remind this model to go check when other model releases happened in order to use that to flag changes in the data as I analyze T3 code data with it.

17:46 Yeah, it still has ZDR off when you use it with enterprise things. So you can't use this model without anthropic getting the data, but they do have the new enterprise frontier safeguard beta which allows you to run the container to verify things in your own infrastructure instead of theirs to keep it a little bit safer. They already have this in p private beta with Bridgewwater, Salesforce, and Stripe.

18:10 One of the very interesting things I learned, I've been talking about a bit, but I saw the impact of it with Opus 55 is that the ZDR policies have an insane impact on adoption at companies. Like obviously that's the case, but Opus was the first like god tier model that behaves well that a lot of enterprises could use with their existing policies without losing ZDR, which resulted in just worldbreaking adoption of Opus 55 at a lot of businesses.

18:35 It is hard to put into words just how insane the Claude Opus takeover of everything was. The vast majority of turns in T3 code, as high as like 63% or more, have been using Opus 55. We had never before seen one model break 50% and Opus hit 60 a few days after it came out. Literally all Codex models was just barely over 50% before and then very quickly Opus took vast majority of the traffic.

19:00 I threw the data up on Slopolytics for those who want to see if you're not familiar slopytics.com. I take all of the numbers from artificial analysis as well as all of the numbers for T3 code and put them in one place so you can see how these things are actually being used in the real world. And as you see, we Codex had a slight lead for quite a while going as high as 70% at its peak, but usually like the 60% range and then ups and downs here and there.

19:24 Then Opus 55 dropped and an instantaneous flip to Claude happened. Now in the 78% range and still growing, by the way, this complete and utter endless domination from Claude is just kind of surreal how fast this happened. But now I need to start talking about actually using it, how it's worked in my code bases, and of course, fish slop. So, let's go over all of my evals.

19:49 Since I have all these giant code bases that were built by agents, I found analyzing them with agents to be a surprisingly useful bench, especially my review bench where I have the models review real code from my real repo and give me an idea of where the shortcomings are in this set of PRs. In the more general codebase analysis runs, I found that Soul and Astra were able to build more cohesive understandings of the codebase and to be frank say a few less stupid things.

20:18 Fable 52 still makes wrong claims here and there of the 61 claims it made about the codebase. Four were wrong, but only cost around $23 to do this thorough analysis of TS Rust. Remember, complete rewrite of TypeScript in Rust. Opus wrote the codebase and it still came in second place behind 61 Soul. 61 Soul made 38 claims, none of which were false. Did it in 29 minutes for $4.19.

20:41 Saying this for a bit, 61 Soul is a workhorse for analysis type stuff. Insane value for the money. How about T3 Code? T3 Code analysis. Opus came out on top. Terra somehow of all things came in second place, doing surprisingly well and insanely cheap. And Fable 52 was a nice third place, although it did make six false claims out of the 83 that it made.

21:00 And in the code review stuff, I found it to be quite good overall. It does take far too long to do these reviews, which honestly kills a lot of the value for me. The fact that it took 120 minutes or so of runtime in order to review 8 PRs is rough. Yeah, too much time for not quite enough of a win here. I don't have the numbers logged for the GPT models here, but they took a lot closer to like 20 to 30 minutes for this batch.

21:28 So, yeah, take that as you will. ups and downs. Overall, slow, but it is the fine-tuned comb going through for PRs at the very least. Generally speaking, I've had pretty good luck with if Fable thinks a PR is good to go. It'd be surprising if it isn't. There's a benefit here that I haven't seen much in other models. Other models are a little too quick to get caught up in nitpicks and point out things that aren't that important.

21:50 Fable 52 is much more willing to let those things slide. And this has benefits and negatives. It does miss some things, but it doesn't get stuck anywhere near as bad as other models do. And for reasons I cannot say yet, models getting stuck on things is going to be a very important topic throughout the rest of the next couple weeks. Let's take a look at some of the things I did with this model today.

22:11 The first was, of course, going through a bunch of math of my usage of Claude code inside of my day-to-day work to see how much this uses the limits. We'll talk a lot about that in a sec. But first, I want to complain about what Fable did. I had it verify the numbers Opus had run. And then I asked it to make the page that it had generated less ugly, and it did.

22:30 It also changed the screenshots here. So, I'm going to go to the original to show what I had from Opus, which is ugliest sin. I then had it do a redesign, which it did here, which I don't know about y'all, this does not very clearly visualize the gap between these numbers at all. like at all. Especially considering that the thing I asked for was to see how much expense and spend you get across different models in your $200 sub.

22:58 Rough. You know what? While we're here, we should talk about said sub. Here is the numbers I got when I drained a few accounts and checked what usage I got out of it. The first thing that I feel is very important to call out is that Anthropic is still doing the thing where you get half of your weekly limit for Fable. They show this in a strange way that confuses people a lot.

23:22 You get two limits on the $200 plan now. Three on the $100 plan and the cheaper plan because those have the 5 hour as well. There's no 5 hour on the $200 plan, but you get a limit for the week and separately you get a 7-day Fable limit. The thing that people don't seem to understand here because they think if they have zero in the normal and then 100 in Fable that they can go use Fable.

23:41 No, the Fable limit is an additional limit on top. Every two points you burn in the fable five limit counts as one point in the regular limit. So you can get the fable limit down to zero and still have 50% left in the week. This is the state I used to leave all my accounts in because Opus 5 was garbage and Fable 5 and 51 were awesome. So I used to have all six of these accounts get to the point where I had no limit left in my Fable, but I had half my limit left in the week.

24:08 That's not the case anymore. So I redid all of the math accordingly. On top of that, remember this model is slightly cheaper, which obviously boils down some amount into our day-to-day usage of it. I had to do a lot of math here. There is one other interesting thing they did, and we get to learn about this together. I'm burning a thing that's very valuable to me to test this with y'all.

24:27 They gave us a reset with this release. If your account is under 6 months old, you get one reset. If it's 6 to 12 months old, you get two. And if it's over 12, you get three. So, on one of my accounts, I got three and all the rest I got one or two. There's an issue, though. This is a model reset. What does that mean? It means it resets your limit with Fable, not your overall limit, just Fable.

24:52 But, as you see here, I have 22% left of Fable and 11% left overall. It also expires in 7 hours, so I could just let this reset. But, nobody has answered this question for me. I want an answer. So, we're going to figure it out together. Will it give me 50% back on my week? Because that's what it should do. Here it shows for the week I have 51% used and for fable I have 0%.

25:16 So it does indeed give you back half your weekly limit. Cool to see. I was curious how they'd handle that. I now know they do still do one of my favorite things which is that when you burn a reset, it doesn't reset the timer as well. Every time Tibo resets codecs, you end up with your timer reset as well. So if you were 3 days away from a reset and then TBO resets, your timers reset to seven days instead.

25:39 So if you have a specific cadence you work on, that can screw you up. With resets and claude, it leaves your timer. You just get more usage earlier. I have found this to be much easier to work and plan around and it screws up my schedule a lot less because with Codeex, you end up with this weird problem where all your accounts reset at once and you're just burning against this larger number.

25:57 Obviously, this is weird for me because I have multiple accounts, but I would imagine decent number of people do at this point. So, having a separate schedule you can plan around because it's consistent is actually nice if you're trying to get the most out of your accounts. So, credit where it's due. Claude does not do anywhere near as many resets. I would take the number we get from Tibo any day, but at the very least, they are easier to plan around.

26:17 So, how much do we get with that $200 sub? With just OPUS55, which means you can use a 100% of your week in Opus, it's $14,00062. 73 times more than you put in on the $200 plan if you max it out. This is for the month. You can do about 14.5 to 15 grand with API prices. And remember, the API prices are, say it with me, inflated. Everyone looks at this as the crazy subsidization of the subscriptions, and it is that.

26:47 But on the other hand, the margins that they charge are so insane on the API prices, they get away with it. Think of it as some part us paying less on the sub, some part enterprises paying a huge tax on top of the API prices in order to subsidize us. It's a balancing act here. Everybody is getting kind of screwed except for Claude, I guess. Anyways, the Opus gets you 14 12 to 15 grand.

27:09 But remember a thing I've been talking about in my other videos about the pricing of models. They said that this model was $40 per mill out instead of the 50 that it was before. And it's also faster. First thing to know is that Fable 5 was supposed to be priced a bit differently. When they announced Mythos originally, it was $25 per million in and $125 per million out.

27:34 Then they got the compute deal with SpaceX and dropped the surprise announcement that it would only be $10 per mill in and 50 per mill out cuz they had enough compute. That was them firing shots at OpenAI as well as dealing with the fact that ZDR not being optional for it meant that a lot of companies wouldn't be able to use it cuz they need zero data retention.

27:53 So they gave the model out cheaper and ate a huge portion of their usual margin in order to do that. As much as half their margin was cut in that simple act. They could have left that price and hypothetically did something like make the model smaller or cheaper to run in order to balance out that cost difference and increase their margin and they decided to make it cheaper.

28:14 Anyways, the fact that this model is 30% faster, I've ran the numbers. It's actually within about 5% of Opus' speeds, that was probably the biggest surprise to me because Fable was historically quite a bit slower. It was 81 to 84 TPS with Fable 51. And this one's 106 TPS versus Opus at 112. They are spitting distance from each other. Now, this model's not slow, which means we're in conspiracy land.

28:36 This is a smaller model. This model is without question smaller than Fable 5.1 almost certainly or they did some weird hacks or made the expert smaller or something. It is clearly cheaper for them to run. That is why they are doing what they are doing. That's why they're offering fast mode now. That's why they made it cheaper. That's why it is faster.

28:54 Objectively, this model is either smaller or for some other reason easier to host and run. Awesome. We benefit from that. That also means that the secret giant Fable model they have internally, it's probably not even Fable. They call it model 4 that they all use, is far from what we got today. It's the crazy mythical thing they hide internally. And they use that to make Fable 52 as well as the 55 line.

29:13 We'll get to the conspiracies around why 52 is not 55 in a bit. I just want to point out this model is definitely smaller. While I talked about the subsidization, I do all of that because it explains the 50% cut because they ate 50% of their own margin when they made the price go from 125 to 50. They cut their margins in half. So, they cut the amount you can use in the sub in half.

29:35 It kind of effectively lines up if you think about it. Don't think about it too hard cuz then your brain will fall out cuz the stuff is all randomly decided internally. So, let's zoom out and look at everything else here. You start with Fable 52 until you're out and then you use Opus for the other half. You get about $3,000 of Fable 52 for the first chunk of the weeks for the month.

29:53 And then with the other 50% every week, you can do about seven grand or so of Opus 55. So you end up taking a $4,000ish dollar hit on the total you can spend in a month if you max out your Fable usage. Could be worse. Obviously, that still sucks. And if you're only using Fable, your multiplier goes from 73x to 15X. It's an 80% decrease. Not great. Thankfully, Opus 55 is a great model.

30:22 Previously, Opus 5 was not, so that other half would just go to die. I've probably lost way more than a hundred of the 50% weak on Fable just across all my accounts because I couldn't use the other ones. Now, I burn all of my accounts to the ground relatively often thanks to Opus. You do get significantly more tokens for that money though because again, Fable 52 is cheaper.

30:40 It is more efficient. So, you get way more output tokens, which are the value that we get for the model. Even though the dollar amount is nearly the same compared to 5-1 for what you get there, the number of turns and threads you get is where things start getting fun. Though, you do lose a bit of turns with Fable versus just going all in on Opus, but it's only about 20% less instead of the much much bigger hit you would take otherwise.

31:03 And for my test, which admittedly, remember, I'm benching the model for a lot of today, it does seem that Fable plus Opus gets you more threads overall because Fable is more likely to finish a thread in a given turn than Opus, but also cuz a lot of the things I ran were one shots today. It writes way faster, too, as I mentioned, the TPS is way better.

31:19 I'm impressed. And as you can guess, this whole UI that looks very anthropicy was made by the new model. On that note, let's check out Dar's Witch AI and see how it does at front end. I hate this font, but the little animations for the cards, not the worst. This one is cute. I like the color and the animations, but remember this color cuz we'll go back to it in a little bit.

31:44 Then we got the subway style, which does not look as cool when it's not zoomed out enough to see proper. I love the animation on this one. The way it goes off the screen is actually quite cool. It'd feel better if the page was aligned a little better, which it struggles because of my smaller screen size. It was a lot better on my normal screen size earlier.

31:59 Bit rough. Still cool overall. And the next one's actually don't like this one that much. And have any like idea where it wants me to go or what it wants me to do. And the fifth one is here. There's cool ideas here. I like the pink. Actually, the fading in and out is not quite as cool as I would hope. But this is all with the design skill on. Ready to see what happens if we turn it off?

32:27 Yeah, you get Tailwind template slop again. There are some cool animations here and there, but like half of these are awful. God, I hate these star style ones so much. Everything sneaks one in. And this brutalist style I'm so tired of. I'm so tired of this. Yeah, and these cards here are cringe as [ __ ] It has its moments, but yeah, this model really needs the design skill or you're going to end up with fable slop.

32:54 Claw slop in general in the UI is a thing you have to beat out of it, and thankfully they provide that skill. It's crazy how a little bit of markdown can help, but it does a lot. That tangent started because I was venting about how much I was pissed at the changes it made. This picture is wrong. The previous one was the vertical one I showed you where there was no difference between the models, and I crashed out at it.

33:12 I said, "The top level screenshot's awful. It doesn't even visualize the difference." I gave it a screenshot here where it like shifted the UI a whole bunch in ways that were really cringe. It then proceeded to fix it and make it significantly less bad. But I have found I've had to be in the loop a lot more with this model. Some of it's cuz I'm not like giving it all of the council type stuff I do with the other models where I have it reviewed obsessively with other things.

33:35 But I also was hoping I wouldn't need to do that. It seems like this model almost needs it a bit more than Opus did. Like it needs you to handhold it a little bit or it can get itself stuck pretty quick. Not stuck as in like it can't solve the problem. It actually unblocks itself quite well. More stuck as in it will make a decision that's not super important and then spiral down it in ways that are obviously wrong to a human.

33:58 Yeah, I'll have a lot more examples of this in the future. There are silly reasons I can't share them all now. I promise it'll make sense soon. I put one other in here that I saw in Twitter where somebody pointed out that this model still does all these unnecessary eyebrow texts and like subheadings. And I had noticed that in a lot of my stuff too. So I asked what it calls it and why it does it.

34:18 It told me what to do. So I made some changes to my agent MD to make it not do that. And it hasn't done them as much since which has been nice. It also seems better at writing pros in general and talking to models. So it's been nice to work with it to overhaul a lot of the copy and the various things I'm working on including a bunch of skills. Yeah, TStack, by the way, is a thing.

34:37 I need to maintain it more and get more things in it soon. PM I was using Tara for some stuff a few days ago, just testing out its capability with managing Linux systems because I found that Tara's computer use has been quite good. But every time I asked it to describe why the storage on this particular machine was way slower than it should have been despite a huge file system overhaul I did, it was just not getting it and spitting out utter slop every time I asked it to describe what was going on and why it was slow.

35:04 It just kept spitting out the same charts and saying a bunch of details that were nonsensical. Cuff 2, which the name of the machine, took 3.2 to 3.3 times as long to set up a warm T3 code work as the three physical T3S machines. A lot of it was in PNPM, yada yada yada. Hard to read, obnoxious. I had just asked like six separate times in this thread as well.

35:23 So, I followed up by switching to Fable and said, "I feel like I'm going insane. Can you just tell me why it is slower?" To which it said, "Cup 2 is a VPS. The other three are physical machines with local NVME drives. That is the full reason. It causes two problems. This seems like a very simple response. The bad news is it's not because I did the same thing in the Terra thread to which it responded with more nonsense.

35:46 Cup 2 is slower because PNPM spends longer creating and linking thousands of small files. That step alone accounted for 5.5 of the 6.3 extra seconds versus the other machines. Two things are happening on this machine. Pulling the same job on plane ext4 saves almost 4 seconds. The new XFSVDO work volume is a major reason setup got slower after the reimage.

36:08 It also takes longer per operation. The small disc reads took about four times as long. Even with every file in RAM, it took longer. Oh yeah, the detail is hidden in the small sentence here. That's the whole thing. Yeah, I mean I get it. Terara is a small crappy model. I just expected better cuz I've been having a good experience with it overall. Obnoxious though.

36:27 This is just a simple good example of a big model doing the right thing. Opus also had a good response here. I just opening eye. What the [ __ ] I know Terara's small, but this is pathetic. What I'm trying to say is this model is pleasant to work with. It does what you tell it and it does it relatively well. I was working on some MCP stuff with a combination of Muse Wonderful and Opus 55.

36:45 And I was trying to simplify MCP guidance in general because I noticed models are really bad at configuring it. And the guidance they wrote required putting a link on the page that was seeable by humans that when clicked would do literally [ __ ] nothing but bring you to an endpoint that errors. Yes, it's sincerely recommended at the top of this. Literally, step one, find the MCP endpoint by putting a URL on the page to the MCP endpoint that a human can't see.

37:13 You could see the link. You can click the link, but the end point's an end point. It's not going to do anything. Yeah, this is sincerely what Opus and Muse recommended. After going back and forth for hours, I asked Fable to take a look at this and immediately was like, "Yeah, I have no idea what this other agents recommending. Obviously, your solution is great.

37:27 Only problem is what if the user's on a normal browser?" And I told it to go look through my history. And it was able to prove itself wrong. That that made no sense at all. It's actually quite nice. I felt like a co-worker that I was talking to here. And here's where I swapped. I told it to give me honest thoughts. Said that my idea is right. The guide is good but oversold.

37:45 It spent too much time on review and too little on actually building. I said that the device flow is only needed for CLI and remote agents. I told it to go through my histories to see if that is the case. It did and it confirmed I was right. You give it the sources for the data, it can usually disprove itself, but it's not scared to push back, which are the things I really like about it.

38:03 But also just misses things sometimes, which results in really stupid behavior. One that I'll show a lot more later is I was working on a dashboard that I wanted to have hide my email and other info when I was filming videos and it screwed that up so bad that I'm scared to even open it during my recording session because in that view it has a big red banner at the top that says you're in the safe mode email address.

38:24 Yeah, I just put the email address in the bar that says you're in safe mode so that nothing leaks and then it just leaks a bunch of things in that bar. It's so stupid. just the types of mistakes I would never have expected even Opus to do. This model will sometimes do. You still got to audit your outputs. I like the way one of my friends framed it who had early access, which as you couldn't tell, I didn't.

38:46 This model is much more capable, but it does not have much more autonomy. It's not better at like making decisions by itself in long runs. It's better at not getting blocked in long runs and solving hard problems in long runs. But things that need a human touch or a human thought or decision, it's not going to do much better there. Unless you're like a super expert in the field and it's doing like very specific technical implementations.

39:08 Yeah, sure it'll do slightly better, but like it's not going to magically make tasteful decisions where other models wouldn't have. At least other OpenAI models certainly wouldn't have, but other cloud models will do well in a lot of the same things. One of the silliest examples I will give for you here is when I asked the new model to audit my project lake bed, my cloud that really should be out by now.

39:29 I've been quite busy if you can't tell. I asked it to audit the project. Let me know what holes you perceive and what you think I should prioritize. And it gave really good feedback about how I'm not guaranteeing backups properly. I haven't set up free account limits properly. We haven't confirmed where the abuse email goes. There are code fixes that we need to land still.

39:47 yada yada. And if you're looking at my screen closely, you might have noticed something on the bottom right. Opus 55 high. I had forgotten to switch to Fable because Opus was my default and I initially ran on the wrong machine. So, I swapped machines and then didn't switch the model and didn't notice until after I read it and was impressed. This is going to be a common theme.

40:04 Opus 55 is still insanely [ __ ] good and you can get way more done with the limits you have available to you and if you're actually paying API prices even more than that. Opus 55 is a gift from the heavens. It is unbelievable how capable it is, especially for the price. Nothing has changed there. Oh, I do actually have one more dumb example here. I was trying to hand off my analytics internal tool Sloth House to another member of my team and I wanted them to better understand it.

40:30 I also wanted to test the design capabilities of the model. So, I had it hunt through my history, find a previous diagram and explainer for how this all worked so that I could have it tidied up, make it look nicer, update it with new info, and then send that to my coworker. It did that and it made something genuinely really nice looking. See the little like goo dip there and [ __ ] like it's cool as hell.

40:53 However, no diagram. The whole point of this was to explain the architecture and make it easy to understand. And I get that I didn't specify to use a diagram, but I told it to find the most recent one we made and update the current architecture and use that in the code as the source of truth. I then complained saying, "It's really cool, but it doesn't have the architecture diagram that I was looking for."

41:19 And this time it drew a diagram. Doesn't necessarily fit perfect on the screen and it's not my favorite diagram I've ever seen. But it added one when I told it I needed to because I need to have the data for what is coming into and out of all of my servers for my architecture here because it turns out uh analytics is not trivial. I have a lot of rants to do about Cloudflare in the near future.

41:40 I uh I've hit a lot of the edges of pipelines recently is all I will be saying. Yeah, we got one more thing that I know y'all have been waiting for. Oh, that's hilarious. This app is such vibecoded [ __ ] garbage. I [ __ ] you not this just happened live. I chose to use claw code as the tool to spin up fish because I was at the time working on cleaning up my T3 code setup so I could use the model properly.

42:05 I'm going to press the letter T once. It is currently doubling every key press. They did some stupid hacky [ __ ] to improve performance and in the process made it so I can perfectly double every single key. I know that a lot of this stuff seems rigged like I'm just trying to dunk on OpenAI Anthropic when I show these things. I don't like that. I have been doing my best to not just say T3 Code is the only usable [ __ ] tool in a lot of my videos recently, but I've been using all of the others recently so that I can

42:39 find the good features so that we don't run out of good ideas and stuff in T3 Code. Holy [ __ ] [ __ ] nothing else works. T3 Code is far from bug-free. It is far from perfect. It's far from even great in my opinion. There are so many things that I stare at every time I open it that I just want to fix constantly. But we don't [ __ ] double the keys when you type them.

43:01 I actually can't fathom that type of bug. My god. At least the model is capable of handling that as I finally now can open up fish slop. Here we are finally. The fish have never been sloppier. Mostly a joke. The fish are actually quite nice. There are a lot of little things I didn't get right on this in my opinion. And I also had to tell it to do a second pass because the first one had uh not the best graphics or the best controls.

43:28 It also had a problem where uh yeah, just the gameplay loop was rough. I did not enjoy playing this one as much as the Opus one, even though it did look quite a bit better and the models are significantly better. That was the one thing that like Astra and OpenAI still had is it could make better models in Blender at the cost of the game being a jinx.

43:50 low FPS [ __ ] show that looked good in screenshots but terrible to run. This runs a hell of a lot better than anything I've gotten out of an OpenAI model. And it looks better overall, too. The gameplay is still not as compelling as it was with Opus. And once again, I feel like a combination of Opus and Fable is the way to go. But I will explain all of that and more in an upcoming video where I go very much in depth on the strengths and weaknesses of all of these things.

44:16 Oh, great. This build broke the right click to attack on a trackpad. So, I literally can't shoot it. I got it to shoot once somehow. Oh, no. It's left click to shoot. That's why I forgot that they changed that in the second build. Again, the controls were jank, so I had to change everything. I also like the open tank style that it did here. Here we are in fish slop opus edition.

44:42 And as you can see, there's a lot of things in the graphics that are worse. There's also a lot of artistic decisions I prefer. I really like the manta ray and wish there were things more like that instead of just more jellyfish in the other one. I like the animations in this one a lot more, too. Also, like the like random objects they have here are cool, but obviously overall it's way uglier.

45:02 There's the bug where things are getting clipped by the ceiling there. And of course, the walls are actual hard walls instead of being a glimpse into a much larger region. But I also found the gameplay felt better in this one overall. And I actually ended up playing this one to completion because I found the gameplay loop to be actually quite fun. So yeah, take that as you will.

45:23 I find that Opus actually makes more fun games. Fable just makes them look a lot better and more importantly stays more coherent over long runs slightly. So where does this leave us? Is Fable 52 the greatest model of all time? Yes. But I would argue so is Opus 55. It is kind of insane the level of domination Anthropic has right now where they have the best big model and the best I I don't know how to call it than like real or practical model because it's not like it's small.

45:51 It's not that it's dumb. Opus 55 is just a workhorse. It does exactly what it is told to do. Whereas Fable tries a little too hard to be clever. It's like the really smart friend that sucks a little bit to work with but really knows what the hell they're doing and if you give them the right exciting problem, they'll come in and kill it. Opus is just the burly bodybuilder that's in there to move [ __ ] And it will move the [ __ ] out of things.

46:13 It'll move things I didn't even think were movable. So why is it 52 and not 55? There's a lot of ways we can read into this like what the role of model 4 is, what model was trained when, how they're thinking about pre- training versus post- training right now, what the five means. But I'm going to put it really simply. When Enthropic put out Opus 4.5, it became the model I used for everything.

46:35 When Anthomic put out Opus 5.5, it became the model I use for everything. Fable 5.2 just dropped and it is not the model I'll be using for everything. I am hopeful this is meant to be a signal to indicate that Fable 5.5 still isn't here and that in the future A.5 or A6 might be the model that I use for everything. 52 would have been 55 if the gap between Opus and Fable was bigger.

47:00 But the gap between Sonnet 55 and Opus 55 is significantly bigger than the gap from Opus 55 to Fable 52. The gap was the same. They would have maintained the numbers. The gap isn't, so they didn't. There are lots of ways to overread this, and believe me, I am tempted, but I am choosing not to because it is almost 2:30 in the goddamn morning, and the reality is that the names are all made up, and if the model was good enough to be Fable 55, it would have been.

47:25 So, sadly, no, we don't have Fable 55. But thankfully, yes, we have a new fable model that is able to do things that I never thought I would see a computer do, much less an AI running itself do. The future's insane. This is the worst it'll ever be. And I'm so excited for a month or two from now when it's four times cheaper, because that is consistently the pattern that we've been seeing.

47:46 Godamn, Anthropic, you cuckooked, and you deserve the wins that you got. That all said, I am exhausted. I want to get this hat off and go sleep. And I'm just going to hope that no new models come out this week cuz there's no way anyone's going to drop anything else after this, right? Especially with like GTA 6 coming there. No, def definitely not. I'm going to go sleep.