# Sonnet 5 vs Opus: the pricing twist nobody expected.

Channel: The Next New Thing
Video: https://www.youtube.com/watch?v=osawfmJzzFE
Duration: 27 min
Language: English
Words: 5021
Transcript page: https://viewrankai.com/tools/youtube-transcript/osawfmJzzFE

---

[0:01] It happens. Claude Sonnet 5 has released and it is by far the best bang for your buck in AI right now. Not everybody agrees with him. I've watched hours and hours of video about Sonnet 5. I've got some people who are making a really good case for why it is not the best bang for your buck. But we're going to watch his video and others here. I've got the best

[0:22] of them with the best of reviewers. Justin Brooke, he and I are going to go through them and he'll give his take because he uses Sonnet every single day and he's been testing it, too. Presented by Zapier, the AI automation company. Let's take a look at Alex Finn's 3D boat prompt. This is his first test. Really nice 3D boat simulator. I'm going to put the prompt down below if you want

[0:45] to copy it and run it yourself as well. I'm going to hit enter on that. I All right. And he hit enter on it and then he showed it in chat GPT. 55. This was chat GBT55's 3D shift simulator. In the prompt, if you take a look at it, it says, "Allow to configure the rain, the wind, the waves." Uh, a little disappointing. One,

[1:07] I can't move the camera at all. Two, the ship is not moving. Three, the water is not moving. Okay, let's take a look at it in Sonnet. And then Justin, what do you think? Let's take a look at what Sonnet 5 did here. I like it better. The waves are moving. The ship is moving. Is there any rain going on at the moment? No. You can kind of see thunder there. But I like the way the ship and the waves look a

[1:31] lot more. What do you think of this test, Justin? I you know, I I' I've seen some of these and I feel like people are using Sonnet for things that I wouldn't ever use Sonnet for. I mean, it it has the we can see that it has the capability here. I just don't know that I would use it for 3D model rendering. Uh but yeah, I mean it looks like it's comparative to 5.5 and when you look at the various test scores, which I don't normally look at those, but I mean it is scoring relatively uh comparatively to 5.5 and also Deepseek V4. So not top of the

[2:13] line, but you know, just a step. What would you use it for? And then we'll go on to World of AI. I want I want you to see what they said. Yeah, I use it for writing. You know, it's it's called sonnet. You know, it was originally designed as a writing model and it shines for any kind of

[2:27] blogging, scripting, reports, analysis. If you look at their anthropics own reports, they say that it's meant for knowledge work, you know. All right, fair point. Let's take a look at this. This is from World of AI, and I like what he did. But buried in the footnotes is something that is quite doubtful cuz the Sonic 5 switches to the Opus 4.7 tokenizer. And this means the same price of text can tokenize into roughly 1 times to 1.3 point times more tokens than before depending on the content. In other words, this basically means that pricing is lower, but some prompts may consume tokens, effectively increasing the overall cost in certain

[3:14] areas much more than what you would actually expect. So, this is what I've seen from a lot of people. I like Mo's take on this because he always has a fun, sarcastic take. Here's how he said it. Most computers cost like millions of dollars and now you can buy a computer for like $500. Yeah, AI is nothing like that. They gave Sonnet a new tokenizer that uses up to 35% more tokens. But see, Enthropic did something here and it's why I love them so much. I mean, is it normal to have a crush on a company? I have a little corporate crush here, man. I have

[3:46] a crush on a C corp, bro. Anyway, the new sonnet ends up costing more with the new tokenizer. So, to get you hooked on the new model without you getting buttth hurt, they're lowering the price for a couple of weeks. After that, it's going back to normal sonnet pricing, but with 35% more token usage. It's like if your landlord kept rent the same, but started counting

[4:10] the months in dog months. That's essentially the take everyone's had, but he's had the funniest approach. You've noticed that, too. What do you think? Yeah. Yeah. So, I'm with him. I like I'm perplexed as a business owner. I'm looking at what Claude or Enthropic is doing and I'm like, okay, there's kind of some smart pricing strategies and launch strategies. So, I don't know that I would say it's a couple of weeks. It's till August 31st and so the input is $2 and the output is $10 when it when Sonet 4.6 was 3 and 15. So, it looks cheaper, but it's only cheaper for a little bit

[4:51] of time. And then also like they said there's this tokenizer that adds 30% of tokens which I've been talking about be before because they're anthropic is also fighting for HTML code instead of markdown code and I'm like that's just more tokens you know for the same thing. Like it's kind of smart they're getting us to spend more tokens which is good for their profits and their eventual IPO

[5:17] but as a user I'm like hey stop that. All right, we'll see more examples, but I should say also Peter Steinberger, the creator of OpenClaw, who now works for uh OpenAI, says, "Watch out, price per token does not equal cost per task." And that's essentially it. All right, so we have that. Let's take a look at what it

[5:33] could do. We actually have already generated a Mac OS clone with the world of AI benchmark tool. So he has a benchmark tool where he creates a Mac OS clone. Let's see how it performs there. What it had done. So, this is the first login page for the Mac OS clone. You have a notification system, which is nice. The top bar does seem to work, but there are slight errors to it. You can change the display. You can also change the appearance from light to dark mode,

[6:03] change the wallpaper. Okay, we don't need to go through all the details of it. It does look nice, but let's look at the analysis of the creation. Essence, it did do a pretty decent job with this Mac OS clone. And just to add, it took approximately 40 minutes within the workbench and it spent many tokens on max mode to get this done. And I can tell you that it was not efficient at all to get this output. But regardless,

[6:28] it did get the job done. Okay, he says this over and over again. It is token inefficient. Token inefficient. Let's go into more detail from Theo. I love his uh breakdowns. Look at cost per task. you'll see something a bit terrifying. By five on XH high ends up being cheaper in realworld work than Sonnet by more than 2x. In fact, Opus 48 is also cheaper than Sonnet 5. But ready for the craziest part here. If we cover the cost to run the entire benchmark, this is the total cost. I'm making an assumption here. I haven't talked to the artificial analysis guys about this, but I'm guessing that the cost per task is

[7:05] averaged with extremes filtered out. And the total cost does not have those extremes filtered out because Sonic 5 is the most expensive model they've ever run through the bench at $6,000, popping even Fable 5 at $5,600. Okay. He gets so angry here. He's even like throwing his business partner or his podcast partner under the bus for

[7:27] not recognizing this. This is an issue. What do you think? Yeah, I I do think overall this is an issue. We're coming into a world where these kind of tests are very important. We have to understand that these companies are trying to max their profits and that may be against our own advice. I I worry about these tests because it's like well your test is on your use but if I use it on my use I'm going to get different results. So it's like you have to do head-to-head comparisons. Like if you're a G I think he's do using this for game coding and so and I use it for blogging. So you

[8:05] want to make sure that whenever you do these you test them side by side. Though never does eat up a lot of tokens. It's once you start creating things that are more resource intense that it becomes a bigger issue. And we could take a look actually here at um here let's look at prompt engineering. He I'm going to throw in GPT 55 medium and low in here. Let's do this. Oh here we

[8:27] go. So this is aentic search performance by different effort levels. Now, Entropic is providing this. The orange one is Opus 48. The gray one is Sonnet 4.6. And then we have Sonnet 5, which is substantially better compared to Sonnet 46, which is expected, but it's actually lower in performance compared to OPUS 48. I would say it's probably OPOS 47 level performance. However, if you look at the price point, this is where it doesn't make any sense. So, let's say if you use Opus 48 on high settings, it's actually lot cheaper compared to Sonnet 5 for a very similar level of

[9:14] performance. All right, I'm not going to belleaguer it, but he has done this with Anthropic's own data and it's still coming out that way. Let's take a look at a hard test. This is, like you said, Theo tested it for game. Here is a setup slow as balls for realworld work, which I experienced myself trying to recreate my fish web game. I pointed at the original repo and told it to rebuild it from scratch. You can use some of the assets. You can reference the code, but

[9:39] I want a new game built from the start. I tried this with three models. I tried this with Opus 48, GLM52, and with Sonet 5 all at the same time. And I have a lot of thoughts. Let's get into his thoughts. This is Opus 48. Has some rough UI quirks here. here. It doesn't do padding and things right, but once you're in the game, the controls are solid. Core mechanics work. And the most surprising thing to me is it did a really good job balancing the economy of the game. Like, it felt good to play for a while and like rank up, get more

[10:08] money, and like actually play the game. Okay. What were you thinking about that? I I think it looks nice. You know, that's that's beautiful. I've been loving GLM 5.2 in my Hermes agent. I love when people use them for comparison 52 version which was interesting. I don't have costs for the other runs because it's not the easiest thing to

[10:29] calculate. They don't give it to you. But I used open code for 52. So I do have a cost. It cost $8.30 to create the following porch. This one's interesting because uh it changed the UI more, but I don't necessarily like the things it changed. It also has super choppy movement, a really weirdly rendered background. Its economy is garbage. So, it just doesn't feel good to play. There's a lot more time sit spent sitting and waiting and nothing really happening. And most egregiously, when you click for the gun to shoot in a direction, it almost feels

[11:00] like it picks a random. This is actually a really important thing to bring up that he he went into later on in the video. GLM52 does not have vision. And I've actually run into this problem and I regret in the video that I made on GLM5 that I didn't hammer that in even more than I I I brought it up. I should have hammered it home. I use it for my Hermes agent. If I need to send it images, it can't handle the images. It doesn't understand them. And that's a real obstacle for both game development like he's doing, but also for me sending a picture to my Hermes

[11:30] agent and asking it to understand what I've sent over. What I have to do is use a fallback model for those images. And that's that's an issue here. That's an issue that's going to keep coming up. But now let's look at Sonnet 5. The same thing. Here's Sonnet's version. Play. Sure. You might be able to see it's a bit of a mess. They got rid of all the hot keys for buying. There are no fish in the tank by default. Just this very poorly rendered pet. I can click the button to buy things, but when I do that, it also

[12:02] shoots. And the economy is garbage. It just doesn't actually feel good to play. At least the bullets go in the right direction, even if they go when they shouldn't, like when I'm clicking other things in the UI. Okay, I won't get into his nitpicking. I will say this. One thing that really frustrated him was that it took over 2 hours for Sonnet to do it versus Opus 48 26 minutes. It did go through a lot of tokens and I love his breakdown of what happened. Um, but you know, before I even break down give his breakdown of how it operates, I will say this, Justin. I could just as easily find

[12:35] people who created games and I have them. I don't know that it's worth going through where the Sonnet 5 was maybe better than the 4.8 or like it's these are small differences where it does seem to keep the issue that everyone seems to have is it does chew through a lot of tokens for resource intense work. Yeah. I mean, we're coming into a world where like, look, if you're just using the latest greatest model and you're not paying attention to token usage, you're the one that's getting wrecked by all these reports. And like that's how I when I see all this data, what it's telling me is like look, if you're not

[13:16] actually getting skilled at AI today, you're going to be just pillaged. Your your wallet, your credit cards are going to be pillaged. But if you start understanding, look at these tests and and understand, oh, this one's better for gaming, this one's better for writing. Uh maybe this one needs more instructions, this one can use just a oneshot prompt. Then we start understanding the right models and the right tasks tactics with those models to get the most efficient cost. And that's that's where the pros are going to be

[13:47] going forward. And you're saying sonnet great for writing. Here's where it is. Yeah. like handles writing no problem. And if you're good at giving it a stepby-step instructions, like that's the problem. Like what I see as a as a daily Sonnet user and I look at his what he did here and I'm like it looks like you gave it a one-shot prompt and that's what Sonnet will do with a

[14:11] oneshot prompt. It's not good at that. But if you give Sonnet some examples, you give it step-by-step instructions, it's really good at working its way through those instructions. and I don't know what his prompt was. So, I will say this, Alex Finn did do that at the end of his video. He said, "Look, the way to use this really well is to use Opus 48 as the thinker, as the organizer, and then after that pass it on to uh to Sonnet to actually do the work." I didn't include the video here because he didn't, as far as I could see, he didn't show the the outcome of it. So, there wasn't anything

[14:44] beyond a strategy there that I just showed. Yeah, that's exactly what I do is I use Opus 4.8 is kind of like my side my day-to-day sidekick. Helps me come up with the big research, the instructions, the skills, and then Sonnet goes and does all the work because it's cheaper to do all the little steps with Sonnet

[15:02] than with Opus 4.8. By the way, I should say that Zapier has their own benchmark where they test to see how these models perform at real business scenarios. Like for example, update Salesforce account tiers, send a launch announcement to enterprise contacts. Sonnet finished it end to end, but Sonnet 5 did. Sonnet 4.6 stalled out on the same task. And what they've found is here, I'll zoom in. Uh clearly Opus 4.8 is doing better. It's 15.4% uh versus Clawed Sonnet 5 at 13.5% success rate, but the cost is now more significant. For a task like the one I described, it would cost 2.0 Oat sense to get in sonnet 5 versus opus 4.8 at

[15:47] 2.36. And these are very clear results where you're trying to connect multiple tasks with multiple tools and sonnet is much more capable and the cost is lower but not dramatically lower. Not as low as I would have expected. You know what then? Maybe we do want to take a look at this one. Um check this out. This is from AI foundations. He did the same

[16:08] test on both platforms. This is on Opus 48 and Sonet 5 five with the exact same prompt. I'm not changing anything here. We are using Opus 4.8 on extra mode. Okay. What I could show you is the finish. I know what these are, but try to take a guess as I'm playing which one is which and try to tell the differences. So, I'm going to hit restart. And now I'm going to be moving around the canvas. Oh, that was difficult. Okay, maybe I don't know how to play the game. Oh, this is very

[16:38] hard to control. Oh wow. So you go straight. Okay. Now let's look at the other one. Here's version number two. Same mechanics. What you notice with version number one is you can see which direction you're supposed to be going because that fire that was shooting out of the rocket was was an indicator. Here you don't have

[17:00] that. The closer I get to that thing, that blue ball in the middle, the more it wants to drag me away. Oh, I figured out how to turn around my guy. Okay. Um, you and I both know, so I'll reveal to the audience. This one was Opus. I know what these are, but try to take a guess at and I want to actually go through and scrub this for a second to show you

[17:23] something. Differences. So 9,800 total. So is so that should be interesting to design goal is just going to keep working in see the difference between them. Okay. One interesting thing right off the bat, Sonnet 5 is flying. It's already running tasks, installing projects while Opus is taking a different strategy. Opus is thinking through this task a lot more. And so it says initiating npm right away installing Chromium. It's all that's what I wanted you to see that we could look at the final output and we could decide we like the Opus version or the sonnet better whatever. The thinking here is different. One is spending a lot

[18:04] more time thinking and the other is just going into doing. And it's the opus that's more thinking. Sonet that's more doing this thing uh without even thinking that much. We're both on extra mode by the way. Uh while Opus 48 over here is really thinking through this. So, okay. And it just keeps going and he's

[18:20] playing the thinking. Let's look at the cost. It actually used the real price of what Sonic 5 actually will be. I don't think it used the intro price, but for output it was 0.13. So that game cost me 13 cents to make. Opus 4.8 it cost me 25 cents to make. Okay, so roughly 10 cents more, but it's considerable when you're talking about

[18:48] that. It's like 60% more to use opus. The difference though is what do you have to go back and say I do want that fire behind. I do want you to think a little bit deeper and have more adjustments. Um, so the cost here I think um is is an interesting thing to notice. But the the reason that I included AI foundations in this mix is he was showing how each model was acting. One spent more time thinking, the other jumped right into doing. I also want you to take a look at this from Theo. The um this is his analysis of how they work. As I was saying, if

[19:20] the $100 an hour employee takes 10 hours to solve the task, and that would be like grand. The $20 an hour employee only takes 10 hours, then it costs a lot less money. It's only 200 bucks. What if that $20 an hour employee takes a bit longer? Let's say they take a h 100red hours and there isn't a senior engineer popping in

[19:38] to check in. Then that task cost $2,000. And now we have a more important question. For that junior engineer, do you think more time spent increases or decreases the likelihood they get the correct answer on any given task? If that task is within an engineer's capabilities, they're probably going to solve it pretty fast. But if it isn't and they try, they're going to take a long time. And that's how we end up in the situation where there are so many of these unnecessarily expensive runs. It's not because they built the model to be expensive. It's because they built it to go and go until it gets an answer, even

[20:09] if it's not smart smart enough to do it. Okay. Makes sense. I like that analogy. I think he's explaining himself well. Yeah. Also, it's brand new and they usually roll out a couple of enhancements, you know, so like as videos like this get made, I think it will get a little bit better over the next couple of weeks. But yeah, I think he's been very fair and I think he's accurately showing the problem and why we need to understand that we should have senior dev agents and junior dev agents and you know I don't want my junior dev just going and going and

[20:47] going trying to figure out the problem. I want to use a senior dev, you know, uh, which is one reason why I like Hermes is because you can just do slashmodel in the middle of a of a chat, switch out to your senior dev model, you know, your opus 4.8, solve a problem quickly, and then go back to a cheaper

[21:05] model to do the extensive work. And I also know you you have Hermes agent that's a writer, which could be on sonnet, and then another that could be on a different model. Let's look at one final video from AI Code King. And then also behind the scenes, I've been running a head-to-head comparison that I want to show you as well. Oh, good. All right, we'll finish with this and then we'll take a look at your

[21:24] head-to-head comparison. Um, he starts off with price comparison. Costs $9 or40 and only scores 55.71%. This model is quite bad. This model is quite bad. Um, yeah. Yeah. I mean, you're really looking at Kimmy being the cheapest here for the total run versus GLM 5.2 being reasonably close and obviously Opus 48 is the most expensive, but Sonnet is actually and it ended up for him being more expensive and being not the dumbest of all of them, but close. And but here's the other thing that um I thought was really interesting, the weirdness that came up for for him. More stuff that is extremely weird with this model is that

[22:06] it tries to do a lot of stuff in the root directory temp folder. For example, let's say that I give claude code with sonnet 5 a task like a math question. Then instead of maybe making a folder inside the current working directory and working in that, it goes into the root directory of the machine and uses that temp folder to work with it which is very weird. This makes it ask for permission all the time as well because it is going outside the current working directory. The model is also pretty weird in a ton of places when using it with something like open code. And it is

[22:37] even more weird. Like it doesn't follow the system instructions well either. I don't know what that means at all. All right, I'm going to stop it there. That is interesting. I will also say the other thing that's weird is the way that AI code King just zooms in and out all the time. I don't know if it's me, but it makes me a little nauseous. Having said that, this is the only review that I saw that had that kind of detail about what it's doing, where it's going, why

[23:02] it's constantly asking for permissions. Okay, that's my roundup of the top videos that I've seen uh from reviewers. You said that you would run a test yourself. Let's take a look. Yeah. Yeah. So, I know that there's the popular generate an SVG of a side view of a Mercedes prompt. Uh I don't know where people come up with these, you know, uh example prompts to use, but uh I did I Let me go ahead and share my screen. Okay, so what you're looking at here is inside of Mind Studio. They're like an NAN or a make.com um a agent builder, but they have this tool called the profiler, which allows

[23:41] you to head-to-head test prompts in different agents. So you can see across my screen, we've got Claude five, Sonnet 5, we've got Gemini 3 Pro, and then we've got Claude 4.8 Opus. And I it's the exact same prompt at the top. You can't really enter a different prompt. Uh so create an SVG of a BMW M4 CS side view. Um and then it ran all of them and then it shows us the cost comparison. So before I go, I have the outputs set up to the side. I'll show the outputs here in a second, but it's cool because it also shows you how long it took for each task. So we can see

[24:23] done in 22 thou what's called 23,000 milliseconds. Mhm. So I think that's 23 seconds if I'm I don't know milliseconds. U and then the other one over here, claude 4.8 took about 27,000 mill. 22 seconds. Yeah. Okay. Turns out uh so roughly the same amount of time sonnet 5 versus 4.8. Okay. Yeah, this is a fairly straightforward thing. Okay. Although though it took me with opus a lot longer

[24:54] when I tried this. Yeah, tell me. Okay. And then the costs. We've got 6 cents for claw 4.8 opus and then we've got uh 4 cents for sonnet five. Okay. And then let's take a look at Gemini Pro, who actually failed the task. Like it didn't actually output something I can use. At least not in the tools that I use. Somebody smarter would probably be able to do it. But this took uh 68,000 milliseconds and but it only cost

[25:29] 0.3. So not even a whole scent to do. Um all right. So let's take a look at the sideby-side comparisons. I got to click share this tab. It's the sonnet version that was that was being made fun of on Twitter. Okay. Yes. So, this is the sonnet version. Definitely not right. That's pretty tough. Um, and then let's take a look at the Opus 4.8 version, which is better, but I don't know that

[25:58] it's ideal. It's strange. I ran it um on my system. Here, I'll share my screen. And I'll show you what it looked like when I did it in in Opus. This is the problem that we see with with these types of comparisons. No two results end up looking the same. Yeah. But I did it and this looks decent. I actually don't know what a BMW M4 looks like, but I know what it doesn't look like. And all the examples I've seen

[26:26] don't look like it. Yeah. Okay. The big takeaway here is it is capable. It does tend to be slow for big projects. It is surprisingly more expensive. We're used to sonnet being the cheaper alternative and now we can't think of it as the cheaper alternative. We have to think of it as the tool for certain jobs, not not the the mini version of Opus 48. All right, if you found any of these helpful, let me know in the comments. Let me know with a thumbs up. And now that you've seen this, Justin and I both love our Hermes agent. We recorded a video where we did

[26:59] something like this for how to use a Hermes agent, how to set it up. Go watch that video right here. Okay, I'll hit stop.

---

About this transcript
Read from YouTube's own caption track and laid out by ViewRank AI (https://viewrankai.com).
ViewRank AI finds the videos already beating a creator's own average on Instagram, TikTok and YouTube Shorts, transcribes them from the audio itself in more than 60 languages, and turns what worked into new ideas and scripts.
Free transcript tools, no account needed: https://viewrankai.com/tools
How to read any video this way: https://viewrankai.com/llms.txt