Claude Haiku 5.5 Is INSANE – This IS the BEST Cheap Model Yet!

0:00 The entirety of this testing used 1% of our weekly usage. Insane in all caps. A sign maybe. Anthropic has added another addition to the 5.5 model family, leaving one left,

0:14 and that is Claude Haiku 5.5. Now, this is their cheapest model, and historically, it was not very capable because this also hadn't been updated for the longest amount of time in all of

0:26 their models. We were still on Haiku 4.5 up until a couple of hours ago. Now, this model seems to properly earn this 5.5 model designation being that every other model we've seen so far, Opus and

0:38 Sonnet, are absolutely fantastic compared to their predecessors. Now, the big takeaway here is this is a direct competitor to GPT6 Luna in that sort of class of models, and it's really

0:50 absolutely demolishing it, at least based off of these benchmarks right here. I would imagine that this will probably prompt some form of GPT 6.1 Luna as a response or something else

1:01 just to kind of take away from the fact that the anthropic models lately have been very very very impressive. So before we get into it, do feel free to subscribe so we can get that 100K

1:10 plaque. And let's take a look at Claude Haiku 5.5 which is prominently listed right here as designed for high volume costs sensitive tasks. So, this is a cheaper model that's not going to have

1:22 as capable performance as Sonnet or Opus or Fable. However, it is really darn cheap and pricing is the huge takeaway here. Now, I don't recall if we already looked at this, but this is basically a

1:33 quarter of the price of Sonnet 55. So, 50ents per million in and $2.50 per million out. Additionally to that, this is also half price of its predecessor right here, Haiku 4.5. and the Haiku

1:46 family of models. There were a lot of folks speculating, is this even going to be updated or are they just going to kind of drop them? Reason for that is take a look at this history of releases

1:56 for the Haiku family of models. Basically, one a year for the past 3 years. 1 2 3. Yeah. So, like this had not been updated since October 15th, 2025 up until today. So, this was very

2:11 long in the tooth. and then its predecessor in October of 2024. It's cool to see this updated now because the Sonnet 55 and the Opus 55 models have been such absolute monsters that I'm

2:23 very very eager to see if that capability has translated to Haiku 55 as well. Now, if we look at some of the benchmarks right here, there's hopefully a bit of justification as to why I'm

2:33 opting to run this on max effort right now. Because while the cost per task does go up quite a bit, there we can see there's a pretty significant capability increase between running this on max and

2:44 running it on X high. So it's not like the drop off we sometimes see like with Opus 55 or Sonnet where it may be better to just run it on high or something. I want to see max performance and

2:54 capability of this model. This is I almost feel bad about this. Look at the Hiku 4.5 dot on this chart. It's just like it's just like why why why do that to

3:08 it, you know? They did it dirty. You know, I haven't even realized I made a mistake here in looking at this pricing table. Four prompts under 100k tokens right here.

3:19 This is 10 cents per million in and 50 cents per million out. Now, it's likely a lot of the agentic coding tests we're going to be performing in today's video will keep us on the higher side of this

3:30 pricing chart. However, that is pretty interesting and I realized that because I was going to just take a quick peek at the technicals of the model and I saw they mention the input pricing here and

3:39 output pricing is from 10 cents per million in and 50 per million out. So, that is pretty interesting. Additionally to that, the context window is 1 million and the maximum output tokens are 128K.

3:51 And the knowledge cutoff right here is June 2026. Now, all of these models are aligned currently for their knowledge cutoff and that's only a few months ago. So good to see. Okay, so first and

4:02 foremost, our browser OS that worked for 43 minutes has hit a response exceeded the output limit. Now, my concern here is I don't know that this is going to continue where it left off. I don't know

4:16 that it had done anything. Um, yeah, it hadn't done anything. That may just happen again. So, at least for the browser OS, we'll be trying it on high. So, here is our browser OS. Oh my

4:28 goodness. Okay, welcome to Nova. Definitely some elements of like a small model smell here first out the gate. Now the big one actually we do have a clock and date in the bottom right in our

4:42 correct time and date in the local there is a right click. H when you click it makes like a you know ripple is the word I was looking for. New pattern. Oh,

4:55 okay. Let's try next wallpaper. Equally disgusting. All right. So, basically were to pick one of these. Probably just close the laptop and walk

5:07 away. Now, next up, we also have mood engine refresh desktop. Let's take a look at our Gemini logo start menu. Okay, we don't have a search option, so that's out the gate a ding on this.

5:19 Here's our mail app. that's actually fairly fullfeatured and packed and the traditional what we see here where there's just references to apps on this browser OS as well as just some random

5:31 things like hey are we going for our trip this weekend good and we did get a response always nice to see and we have the notification in the top right next up okay so the GTA clones all the way

5:42 down there I'm going to click on that now because oh no I think this has muse spark 1.3 dust where the map is. You know, that's okay because it's

5:58 translucent enough that we can still make out some of the game. Oh, I saw a pedestrian. All right. Okay. And we do have the ability to collide with said pedestrian.

6:16 This is simple. Now, of course, the map thing here is a big detractor, but to be honest with you, I think our testing and time is better utilized doing more results than having it fix this specific

6:27 issue cuz we can see enough of the game that this is just actually these cars are pretty good to drive. Like the handling, we have three police moving in on our location. All right, let's see.

6:38 Do they snap to our location and just stay there? No. So, that's actually okay. And then our health goes down. Can we get out of the car? I don't think we can. Let me

6:48 just try the traditional keys. No. Acceptable, especially for this class of model. All right, we have a files app. Similar to the mail app. Not bad. It's basic, but

6:58 it's not bad. It's just basic. This will show us our note app. We can adjust the font size. Save as. Okay. And it just saves it to the file system on this system. It doesn't save

7:10 it to our local system, which is acceptable. Notepad. We did just see terminal. This this background is just it's just it's quite something. Help. Uh okay. So we had caps lock on and we

7:23 can see sometimes the models will still work if we have caps lock on. Not bad little egg looking thing and some cleverness implemented as well. We have a paint app. That's actually a

7:34 fullfeatured paint app. I'm quite pleased with this and it scales nicely so there's nothing funky there. I heard a beep. I don't know what that was for though. That was the same sound that was

7:44 made when we gun an email, but I'm not seeing those fake notifications. Can we save this to our local system or does it save it in the file system? Storage full. So,

7:54 that's actually cool. It saved it in our pictures folder in the fake file system. Not bad. Following that, we have settings. Pretty simple. Meatballs, plasma, and flow. All equally as hideous

8:07 as one another. Mood engine wallpaper reacts to how you work. That's interesting because I don't quite know how to parse that. Maybe this one would I wonder if it's just that. All right,

8:20 we have a mood engine. Focus. The wallpaper is following you. Nova listens to the shape of your activity, how many keys you press per second, how fast the pointer moves, and

8:31 how long it's been since you touched anything, and turns that into a live energy value from 0 to one. Interesting. Oh, and then finally, how could we forget our second game, which is again

8:41 probably actually that's kind of cool. Look at that. It's It is 3D and it at least at first it looked very retro 2D. I was kind of okay with that. Not bad.

8:53 Okay, the effect actually when our ship gets hit is pretty good. Shot down. All right, not bad. Overall, simple. And again, we have to remember that this was not at its maximum potential performance

9:04 because of the issue trying to run it on that. And there was a pretty big leap in its capability between like high and x high or max. So this is not its best work. So also on max, our C++ skate game

9:16 has completed in just under 38 minutes. It mentions the file was modified on disk during the session. I didn't touch anything, so we'll just assume let's just ignore that. So, let's

9:28 see what we got for our C++ skate game from Haiku. I will say this is for a Haiku class model. This is Oh, is the keyboard plugged in? Okay. It didn't do WD for movement. It did arrow

9:43 keys and I got confused. What I was going to say, I think, is that this is definitely better than what was done by GPT6 Luna. And I think a lot of folks are going to be interested in seeing the

9:56 comparisons between those two models because they are competitors. Okay, so I want to see if we can do a kick flip and see if the board moves independently from the player. That

10:08 small model smell right there is when the trick actually moves the entire human. Can we interact with pedestrians? No. Okay, so this is definitely very basic, but it almost just does feel like

10:19 a boiled down version of what we get from Sonnet or maybe less so Opus. They do definitely have similar elements. And again, this is C++ with no dependencies. So this is not necessarily a trivial

10:32 task. Oh, there's also the replay feature. So this is the skate test where you can press Z and it will give you a cinematic replay of the trick you've performed.

10:44 And it even gives us like a trajectory there. That's very interesting. Now, I definitely believe that this is something that could benefit from an improvement. So, we'll give it some

10:53 feedback and we'll also maybe give it a reference photo. So, I've given it a follow-up just saying, I placed a reference photo in this folder. I'd like you to improve the visual fidelity of

11:02 this game as well as niche details like having the board move independently from the player model when making tricks. And the photo here that we've given it is just a screen cap from the Sonnet 5.5

11:13 review. So, it obviously won't replicate this. This was still like a freak result. It was fantastic, but we'll see how much influence it can take from this and if it puts water in in a boardwalk

11:23 and things like that. All right, in 23 and 1/2 minutes, we have our skate game overhaul. And this is one where we gave it this Sonnet 55 reference photo as well. So, okay, I definitely see

11:33 inspiration taken here. And the actual movement of the pedestrians definitely is far more in line with what we saw with that. Now, regrettably, I can't actually make this window any bigger.

11:44 Probably something I should have included in my follow-up to this. Let's just see. Now, one of the big ones was make the board move independently from the player when doing tricks. Oh, but

11:53 the legs moved with the board. That is one of the more disturbing things I've seen. I mean, that is actually pretty darn sick for a Now, I understand it did get

12:06 an image as a reference point for what I wanted it to look like. However, for a haiku class model, prior to that, I was going to say, "This is really sick." It probably still is. Can I get rid of this

12:20 help menu? Yes. Good. Good. So, traffic does work. I mean this is like this is one flexible individual. Ah oh. See we can replay the wipeout and

12:38 it does still have that trail there. I see some inverted text on the storefront. So sadly this model is not good. No. All right. Good. Look at this skybox. It did put the boardwalk though

12:54 which is what we got from the reference photo. Ollie, perfect. All right. All right. What happens if we just go into the water? Mesh collider or splash? Splash.

13:05 Good. And it just respawns us. Ah, there's no replay for that. So, but the pedestrian movement, that's not something that it would have figured out. Well, no. Maybe it is from the

13:21 photo because they were made of independent like pieces. So the legs have multiple pieces and then it could have discerned from that. Okay, they probably have some form of simple

13:30 movement animation. Overall, I'll say this is quite quite impressive, especially considering this class of model. Let's see if we can grind at all. That's one thing I have not yet seen. Oh

13:41 no, get out of here pool party game. I was trying to replay the nice kind of Oh, all right. Interesting. Definitely better. And still this compared to GPT6

13:54 Luna is quite different even prior to having it have the follow-up prompt. Next up, I've put this on Ultra Code just because I want to try at least one thing on there and we're giving it the

14:05 Blender and GDAU backyard pool party test where it needs to create a pool party game using Blender and GDAU. This is one that's very recent, so it's far less likely to have made its way into

14:15 any training data for these models and it'll be interesting to see how it performs in this task. All right. So, right now on Ultra Code, it's just working through the pool party result

14:24 and it's spawning sub agents, but it's just testing the water effects and the splash effects and stuff. Look at that. And I mean, like, it's pretty good. Oh, not bad. Maybe

14:37 that's a screen of some sort. All right, we'll see what we get. But so far, this is looking quite promising. And I'm going to step away from it now cuz I don't want this to be spoiled. I

14:47 like seeing these like for the first time together. Our Ultra Code pool party result has completed here. This is Blender and GDAU. And the reason I wanted to try this is

15:03 because it's less likely to have been like present in any training data because it's a relatively new and different problem. Okay, backyard. Look what it when it when it when the

15:15 camera orbits and the L goes in front of the doors there. It said backyard poo party. I need to I need to allow it to do a full rotation. I'm sorry, but we're

15:34 going to wait. All right, here we get ready. It's coming around again. So, right there. And wait for it. Right. It was funnier the first time, but

15:49 wanted to just dicks. Big dicks. Mia flips. Grandma Rosie or Shay. Go with Grandma Rosie. Oh, okay. So, unfortunately the I have no idea how to play this.

16:08 the UI though and stuff like this isn't this it's very like Wii Sportscoded in terms of the UI and this is just an okay look at the dog.

16:30 What? Oh, the dog jumped in the pool, too. Insane in all caps. a sign maybe. Next up, Shay.

16:52 That's like a three star result. Look at the dog in the water and the water effects. That's That's just nice attention to detail. I still have no idea how to play this, but I'm just

17:06 going to spam stuff. Oh, okay. So, we have to get the perfect pump. Let me quit to main menu. Let's just do practice. Oh, cool. We can actually like

17:20 space to pump. Oh. Oh, that's actually kind of cool. The inair movement ability. And we have to press space when that's in the yellow or in

17:34 the in the green. This really is not bad considering it used Blender and GDAU. Yeah. All right. Good. So, it does respond to that. That was

17:47 even look at the options screen and things like that. It did a nice competent job on this, I'll say. Aside from some of the background coloring, this was

18:00 I think it was actually pretty good. The water effects and stuff, the dog jumping into the pool, like it, the string lights. This was This was good. Next up, we're going to be giving

18:13 it on max effort just here through the Claude Code terminal app. The subway FPS. This is the newer one where when we have cleared a wave, we enter a train that pulls into the station, which

18:23 brings us to a next station and the next wave. All right, in 35 minutes, our Subway FPS has been completed. And it gives us here just the share link to run it, which I guess is fine. Okay, last

18:35 stop. Not bad. So far, not bad. Oh, okay. And now it's bad cuz the mouse maybe we just that is definitely an issue where the weapon doesn't move with the Let's see

18:51 if Okay, so the glitches in terms of the weapon rendering and not the actual shooting logic. This is really for the class of this model. This is not bad at all. We should have Okay, good. We have

19:03 and we have the right click ability. Yeah, this is Look at that. Benchmaxed. No. All right. Now, the train will come. I actually I want this to be fixed because I'm angry.

19:17 Maybe we can also include the train being a little overhauled than that. Actually, that's not so bad when it shows up. For a haiku to have done this like,

19:28 yeah, it's probably a fairly known prompt at this point on account of the, you know, influencer status. Uhoh. It's frozen. All right. I've given it

19:39 the feedback just mainly about the arm being messed up and then when we were riding the train to the next station, it froze up. And then I said add some more details. So, we'll see what we get. But

19:48 really, that was not a bad result for a Haiku model. But again, this is something that's been around for quite a while. So, be I want to try something that is very less likely to be

19:58 potentially included in training data. All right. In only 8 and 1/2 minutes, our Subway result has supposedly been fixed. So, let's just open it again here in this little handy artifact link, I

20:09 guess. Now, the big one. Good. It did fix that where we had the hand was floating and not corresponding with the camera movement. And there was like just the tile.

20:22 It's just well done. It's clean. Now, the ammunition tracer staying there is pretty darn cool, but unfortunately, that is an issue. We also told it to increase the visual polish a bit. I

20:33 don't know that I can tell off the bat if it has, but I'll say that it's definitely gone kind of hardcore. And the ammo tracers are just kind of cool. Uh-oh, now it's

20:45 freezing up again. Yeah, unfortunately, it just bugs out. It seems like when it's time for the train to come, which is weird because in a way it could be thought of that it regressed a bit. The

20:55 last thing I'm going to do because we have time to just see how far we can push this is give it the specific error I saw in the console and then tell it it freezes. All right, so I'm going to

21:04 quickly just run through this to see if it was able to fix the issues we were noticing with this game freezing. I am just running it in the browser natively instead of the shared link which I had

21:14 tried previously as well and that didn't make a difference with the freezing. So technically now, well now the train should come. And so far it does seem like the freezing has stopped

21:28 at least. Okay, I can't shoot while on the train. That's interesting. I've not seen any other models do that. Cool. And we're at an entirely different

21:38 station now. All right. I'm glad we did this. This was definitely worth spending a little bit of extra time on. And then the train leaves. I'll go to one more and then we'll

21:54 Oh, look. The next station changed, too. That's cool. It put a screen in the train showing us that info. That's just well done. All right. And and all the stations so

22:05 far have had different aesthetics. I like that. So, it would have been real easy for this to just have the same exact station each time, and I would have forgiven it because of the size of

22:14 this model, but it didn't do that. And I think there's just maybe Oh, I think there's just maybe more hostiles. Oh, all right. And then we'll we'll see what

22:34 happens if we lose. And there are bigger ones, too. You died on the platform. That was well done. Third times the charm. So, next up on Ultra Code, because I wanted to try

22:50 something that's again less commonly seen, I made it do a pixel perfect replica of PvP at the Grand Exchange in Runescape 2007. We'll see how Oh, wow. Okay. This is definitely maybe one of

23:03 our limitations shown. Although I can't fully like hate on this. It's simple weapon, bronze sword, and it goes back into our items here. I don't know that the PvP is necessarily noob.

23:20 Not bad. Okay, there's no like NPCs that I see. Examine water. This is the only model so far to do this test that has just given us like completely beginner stats. Most

23:35 of them make you like maxed out and everything. This is a level three with 10 hit points. Also, on max effort, I gave it the traditional watch website test, including the one where it needs

23:44 to have the behan where it puts this photo as the dial of the watch. So, let's see see what we get here. We also do need an exploded view. Okay. Slapis so far. Ah, very good. That's well done.

23:59 It has the issue we notice sometimes where the bands are just really, really, really narrow. But the rest of this, I will say, is actually quite nice. All right. So far so good. Well, aside

24:11 from the super narrow band. Good. We have different elements of it as well. We have like a steel one with the blue. Looks good. And someone had said in the comments, "Watches are shown with the

24:23 time here because it's like the most aesthetically pleasing way, but also you can see like the top and the bottom of the face. It's not obscured by the dials.

24:32 The special edition 99 and they all have an issue where it's like almost blurry 8,900 for the Bejan. Bit on the lighter side and it did actually pull details for the strap from

24:46 that image which is told to do in the prompt to inspire the aesthetic of that watch by that photo. Now we need to see our pull it apart slowly feature. Not bad. And the crown actually came out

24:59 to the side as well. H that's annoying though is we can't manually interact with this. Like I want to be able to click and drag. It's not denoted in the prompt that it needs to do that. But

25:09 most models do do that. So also though like I have to remember this is a haiku model. So it's easy to forget that with some of these results which I would say bodess well for the w for the waddle. I

25:21 was going to say watch and model for the model. And that took 22 minutes and 9 seconds just on the highest effort level that it had available in the claude code CLI. So max. So the final thing I've

25:33 given this is this Gemini generated image. So nano banana of this demolition derby. And I've told it to make a pixel perfect replica of this. Additionally, because we're not going to get a pixel

25:44 perfect replica, the cars need to have soft body physics, the ability to have different camera views, such as cinematic, firsterson, and third person chase camera. Parts need to fall off.

25:54 There needs to be suspension, and it should just be a fun little demo derby game using Blender and GDAU. This is another one that's fairly new, so less chances of this having been seen

26:04 anywhere by the model. So, here's our It's It's not a hor Yeah. Okay. It's It's not horrible, but it's like Just maybe turn that all the way down. There is suspension. We do have the

26:41 cinematic and firstperson view camera and I think it actually did a decent job with the arena and stuff like that and with the suspension. It's not horrible.

26:52 It's just the cars are quite hilarious. And it even replicated the smoke particles that happened from the collision in the reference photo. Those are actually seen there. So, I give it

27:04 props for that. It's a dirt lot. There are clouds in the sky. Let's see if we can get like a I'm not sure that I'm able to discern any softbody collision physics here.

27:17 Actually, I think the bumper just bent when we backed into that vehicle. Yeah. So, it did kind of like it kind of implemented softbody physics in a way. And the suspension's actually

27:30 not bad. Handbrake. Okay. Interesting. I didn't know what to expect. So, regardless, that's actually in How long did that take? I don't see a specific time that it's listed here, but

27:55 I don't really think that took very long. 1 hour ago and then 44 minutes ago, but 1 hour ago could be like an hour and 10 minutes. Still, I think this took less than 30 minutes and that was

28:05 on max. So that was not using ultra code or anything. A very interesting result. All right, so I didn't know how I was going to title this video because on one hand, the results were not amazing

28:17 compared to Sonnet or Opus or Fable or GPT6 Astra. However, the pricing is really, really, really cheap. Now, the thing that's pushed me over the edge to probably put insane in all caps in the

28:29 title is the fact that the entirety of this testing used 1% of our weekly usage. I took a screenshot of my usage prior to beginning to film this video, and it was at 10% used for the week.

28:42 Now, it does reset in like a couple hours, so I should probably just do some random stuff with like Opus or Fable. However, 1% for the weekly usage, that's very,

28:53 very, very, very low. And the takeaway here is this model is probably going to have folks come out and say, well, yes, but on higher reasoning efforts like this, it's too tokenheavy. It's

29:04 inefficient. And if you see some of the benchmarks there where the actual cost per task may be far more inefficient than GPT6 Luna, there's whatever the Luna number is, it's yeah, six.

29:16 However, I'm doing this all on the subscription. This is the $200 a month plan subscription. And the entirety of this testing used 1% of the weekly usage. That's insane, I think. So,

29:28 overall, that is going to conclude our first look and test of Haiku 5.5. I think in lie of a traditional results overview, cuz these were pretty close together. It didn't take super long to

29:39 do all of these tests. The backyard pool party is probably one of the best showcases of its capability just because it's very unlikely that this could have been benchmaxed in any way and it had

29:50 the dog the UI here looked pretty good. It was funny. It was just a fun game and we have to remember it made all of these assets itself in Blender and put the game together in GDAU. So, it was cool.

30:02 I guess actually as like a mini results overview, the one that was also pretty darn cool was the C++ skate game. Now, we had a couple different iterations with this, just giving it a reference

30:12 photo as well of the sonnet result, but it really did a nice job of properly replicating that reference photo. This for a haiku class model right here is quite quite nuts. This is

30:26 state-of-the-art Frontier. Like I mean this would have almost been comparable to maybe I won't say the original Fable result but cuz that was done with the California Boardwalk

30:40 version and not the NYC map but like I can tell you this is close to state-of-the-art from like 6 months ago if not fully. So that's just awesome. The buildings look good. It's clean.

30:53 There's movement. You can hit the pedestrians. And this is where I'll probably leave this off at. The skybox was hilarious, though. So, this model has some pretty

31:02 insane capability, but the biggest takeaway for me personally is the fact that the entirety of our testing right here used 1% of the weekly usage. That's fantastic. So, overall, that's

31:15 going to conclude our first look and test of Haiku 5.5. There's only one 5.5 remaining, so we'll see when that comes out. If you have any questions, leave them in the comments.