I Tested DeepSeek V4 vs Opus 4.7 vs GPT 5.5

[00:00] In the last 24 hours, we have had huge

[00:02] updates to two of the biggest AI models

[00:04] on the planet. First, we got the release

[00:05] of GPT 5.5, which is boasting certain

[00:09] benchmark scores that beat out Claude's

[00:11] Mythos. Secondly, we got the release of

[00:14] DeepSeek V4, which is an open-source

[00:17] openweight model that has benchmarks

[00:19] that rival these frontier big players.

[00:22] So, with all these new models to choose

[00:24] from, what are you, the average user,

[00:26] supposed to do? Well, today I'm going to

[00:28] help you answer that question as I pit

[00:30] Opus 4.7, GPT 5.5, and DeepSeek V4

[00:34] against one another so you can see which

[00:37] one actually makes sense for you. Now,

[00:39] before we kick off this head-to-head

[00:41] test between GPT 5.5 inside of Codeex,

[00:45] Deepseek V4 inside of Open Code, and

[00:48] Opus 4.7 inside of Cloud Code. Let's

[00:51] first take a quick look at the

[00:52] benchmarks, especially these two latest

[00:54] models that dropped in the last 24

[00:56] hours. Now, let's first talk about cost.

[00:58] Now, DeepSeek V4, as you know, is an

[01:00] open- source openweight model, but that

[01:02] does not mean you can run this on your

[01:03] computer because this thing is huge. I'm

[01:05] talking 1.6 trillion parameters. You

[01:08] need some serious hardware to run this.

[01:10] So, we still got to pay for it. We're

[01:11] still going to have to use the API, but

[01:13] it is infinitely cheaper than the

[01:15] competition, about 8 times cheaper. And

[01:18] of the three models, the brand new GPT

[01:20] 5.5 is actually the most expensive,

[01:22] which is kind of surprising because by

[01:24] and large, OpenAI has been cheaper than

[01:26] its anthropic competition. In terms of

[01:29] what it will cost you per 1 million

[01:31] tokens of output for GPT 5.5, it's going

[01:34] to be $30. For Enthropic, it's going to

[01:37] be $25. And for DeepS, it's going to be

[01:39] $348.

[01:41] Now, if we're talking about input

[01:43] tokens, which is a smaller part of the

[01:46] hole, GPT 5.5 and Opus 5.7 are the same,

[01:49] it's going to be $5 per 1 million input.

[01:53] And for DeepSeek, it's about like $1.70.

[01:57] So, way cheaper on the input and way

[02:00] cheaper on the output. That being said,

[02:01] when it comes to 5.5, this is like twice

[02:04] as expensive as 5.4. However, OpenAI

[02:08] claims that it actually uses way less

[02:10] tokens due to its power. So, while it's

[02:12] double the price of 5.4, they say in

[02:14] terms of actual token spend and actual

[02:16] cost for the same task, it ends up only

[02:19] being like 20% more expensive when it's

[02:21] all said and done. So, just have that in

[02:23] the back of your mind. So, we've talked

[02:24] about the cost. Now, let's talk about

[02:25] the benchmarks. How good are these

[02:27] models on paper? I know we're all kind

[02:28] of numb to benchmarks in general. we

[02:31] need to take them with a grain of salt,

[02:32] but it's still worth taking a look,

[02:33] especially when we're looking at the

[02:35] numbers that are reported by each player

[02:37] on the same benchmark. So, there were

[02:40] three in the coding category that all

[02:42] three reported numbers that was Swebench

[02:44] Verified, SweetBench Pro, and Terminal

[02:46] Bench 2.0. Now, for SWEBench verified

[02:49] and SweetBench Pro, Opus was the winner

[02:51] there. On Terminal Bench 2.0, GPT was

[02:55] the winner by far at 87.2, 2, which by

[02:58] the way is a higher number than what

[03:00] Anthropic reported for Mythos, Mythos,

[03:03] sorry, which is kind of crazy. You know,

[03:05] the super secret model they can't

[03:06] release apparently does worse on

[03:08] Terminal Bench 2 than GPT 5.5. Now, the

[03:11] Terminal Bench 2.0 is the biggest

[03:12] outlier here. Opus 4.7 and V4 Pro are

[03:15] way behind, but take a look at Opus 4.7

[03:18] versus V4 Pro. It's less than two points

[03:21] while being eight times cheaper. And you

[03:23] see the same sort of story here with

[03:24] SweetBench Verified and SweetBench Pro.

[03:26] Yeah, Opus wins, but when we compare the

[03:30] second place with the third place, and

[03:31] V4 is always third place, there isn't

[03:33] the huge gap you would expect. I mean,

[03:36] five points isn't nothing, you know, on

[03:39] Sweetbench verified 85 to 86, but again,

[03:42] eight times cheaper, open- source, you

[03:45] know, there there's some actual

[03:46] trade-offs here that we can make if we

[03:48] don't need the most power. Another thing

[03:49] that's interesting to talk about is long

[03:51] context where oddly opus 4.7 is really

[03:55] bad by the numbers like significantly

[03:57] worse than 4.6 which kind of blows my

[03:59] mind. And when we're talking about long

[04:01] context where we're trying to retrieve

[04:02] things between 500,000 tokens and 1

[04:05] million tokens 4.7 is actually terrible

[04:08] and does way worse than deepseek and GPT

[04:11] 5.5. Now you can have a whole discussion

[04:13] about why are you even in the 500,000 to

[04:16] 1 million token range to begin with? how

[04:18] many people are actually operating there

[04:20] because we are hitting context rot no

[04:21] matter what at that place no matter what

[04:23] model you're using but it is interesting

[04:25] that for whatever reason we've seen some

[04:27] regression when it comes to the

[04:28] anthropic models but big picture I think

[04:30] the takeaway is 5.5 is really strong it

[04:33] beats opus 4.7 in certain metrics loses

[04:36] in certain metrics but it's an extremely

[04:38] robust model and on top of that while V4

[04:41] Pro is kind of you know lagging behind

[04:44] by and large it's within striking

[04:47] distance while being infinitely cheaper,

[04:48] which again is a great option for your

[04:51] average customer because right now it

[04:53] feels like you don't have a lot of

[04:54] options on the open source side that

[04:56] actually can compete. Now, let's jump

[04:57] into the actual head-to-head test with

[04:59] all three of these models. And we're

[05:00] using a harness for each of these

[05:02] models. With 5.5, it's going to be

[05:04] codeex. With Opus 4.7, it's going to be

[05:06] Claw Code. And with DeepSeek V4 Pro, I

[05:08] am using Open Code. And for the first

[05:10] test, what we're going to do is we're

[05:12] going to have them create a flight

[05:13] simulator for us in 3JS that runs in the

[05:16] browser. You can see the prompt right

[05:17] here. I'm saying I want it to feel good

[05:19] to fly. I want it to have some weight to

[05:21] it. I want some strong visuals, and I

[05:24] want it to use whatever structure and

[05:25] tooling it thinks is correct. So, it's

[05:29] straightforward enough that they know

[05:30] what to do, yet there's enough leeway so

[05:32] we can see some divergence between the

[05:33] models. And while we are going to look

[05:35] at what they're able to oneshot, we are

[05:37] going to go through multiple iterations

[05:38] of this and have follow-on prompts

[05:40] because as cool as it is to see how well

[05:42] it does on one shot, that isn't how we

[05:45] really work in real life, is it? Right.

[05:46] I want to see how it does when I give it

[05:48] follow-on prompts and how quickly it

[05:50] takes to get it to something I like. And

[05:52] when we compare these three models,

[05:54] there's really four things I'm going to

[05:55] look at. It's going to be time. How long

[05:57] does it take to build this? Cost, how

[05:59] many tokens are we using? Quality, how

[06:01] good is it? And then four is sort of

[06:04] vibes, and that sort of relates to

[06:05] quality. It's very subjective. Which one

[06:07] do I actually like more? And also of

[06:09] note, all three models, all three

[06:11] harnesses are also using the exact same

[06:13] skills. So, let's begin with Deepseek in

[06:15] the questions. It's asking us, it's

[06:16] asking what sort of flight model we

[06:18] want. Let's go with full sim. It's

[06:20] recommending oceans and islands for the

[06:22] terrain. We'll go with that. Let's see

[06:24] how. And then it's asking camera

[06:25] preference. Let's do both. Let's see if

[06:26] it's able to give us a toggle for both

[06:28] first person and third person. We'll go

[06:30] with its recommended tooling preference.

[06:32] And we'll just go with a low poly model

[06:33] for the aircraft and visuals itself.

[06:35] Now, moving over to Codeex. Same sort of

[06:37] questions, although it's only asking us

[06:39] three. Saying what kind of flight should

[06:41] this plane optimize for? Let's go with

[06:43] hard simulation. Which playable

[06:45] experience matters most for the browser?

[06:48] Let's do island takeoff loop. It is kind

[06:50] of interesting how they all have the

[06:51] same one. And what camera and aircraft

[06:53] presentation? I'm going to do toggle for

[06:55] this as well. And for cloud code, we'll

[06:57] do study sim learning. For the field,

[07:00] ocean, and islands input, we will do

[07:02] keyboard and mouse. It will let it go to

[07:04] work. So plan mode by the large very

[07:07] similar across all three. pretty much

[07:09] the same questions of like what do you

[07:11] want the physics to be? What do you want

[07:13] the terrain to be? What do you want the

[07:14] camera angle to be? So, no huge

[07:16] difference there. And let's see what

[07:17] they come back with in terms of a plan.

[07:19] All right, so all three plans are

[07:20] complete. So, let's go through each of

[07:21] them pretty quickly and see some of the

[07:23] differences. First one we're looking at

[07:25] here is Deepseek and it's pretty bare

[07:27] bones in terms of the plan it lays out.

[07:29] So, gives us the project structure and

[07:31] then talks very quickly about light

[07:33] physics, environment, camera, and HUD

[07:35] overlay and really just a few bullet

[07:36] points. On the other hand, when we're

[07:38] looking at 5.5 inside of codeex, it

[07:40] gives us a summary, key changes, goes

[07:43] into implementation details, the test

[07:45] plan, and as well as the assumptions. It

[07:47] spells all that out for us. And then we

[07:49] have Claude codes plan, which took the

[07:50] longest, took it about 5 minutes, but by

[07:52] far is the most thorough because it's

[07:53] the context, the stack layout, talks

[07:56] about the flight model. It's going into

[07:58] like the actual different moments,

[08:00] talking about stalls, like the stall

[08:01] buzzer, like it's it's going very very

[08:03] detailed. goes into the controls, the

[08:05] world, the mod, the actual aircraft

[08:06] we're going to be using, performance,

[08:08] and just keeps going on and on. So, very

[08:11] detailed. So, now we're going to have

[08:12] all three implement their plan, and

[08:14] we'll see what the final result looks

[08:15] like. So, GPT 5.5 inside of Codex was

[08:18] the first to finish. So, let's see what

[08:20] it looks like. So, here's the flight

[08:21] simulator it got us. We have, you know,

[08:24] some clouds in the sky. We have what

[08:28] looks like an AOA indicator up there. We

[08:31] have our speed down below. And let's see

[08:34] if we can actually get this thing off

[08:36] the ground. I will note there's no more

[08:37] like runway. It's just like straight

[08:39] grass and it said it was going to be

[08:40] like an island thing. Although when the

[08:44] camera kind of spazzes out, you can see

[08:45] the runway down below there for a

[08:47] second. All right, we're stalling out

[08:49] and we just we can't even get off the

[08:50] ground. All right, so this this one's

[08:53] actually just a little is actually kind

[08:54] of difficult. Um, so what I'm going to

[08:58] do is I'm going to give it a second

[09:01] prompt asking it to make it a little bit

[09:02] easier to fly because it has a lot going

[09:04] on here, but this is this is tough. So I

[09:07] wrote, "It is really hard to fly. Can we

[09:09] make this easier to use?" Aka a little

[09:11] bit more arcadey and also the graphics

[09:14] could use some work. So let's see how

[09:16] that does. Now, of note, um it took 5.5

[09:20] about 7 minutes to create that first

[09:22] pass for us, and it took 63,000 tokens.

[09:26] All right, it said it made it a little

[09:27] bit easier to fly and updated the

[09:29] graphics. So, let's see what the second

[09:31] pass looks like. So, here's what we got.

[09:33] Graphics definitely look better, but

[09:34] let's see if we can actually get off the

[09:36] runway this time. So, all right,

[09:38] throttle's at 100%.

[09:41] 50 60. What's the rotation speed on a

[09:44] Cessna? All right. 70 80 90. We got to

[09:50] be able to get off the ground now. Okay.

[09:52] Wrong way. Let's go. Get off the ground.

[09:55] Get off the ground. Nope. This is

[09:57] probably going to stall me out, isn't

[09:58] it? Yeah. Stall. Okay.

[10:01] This still needs this still needs some

[10:02] work. So, let's let's give Codeex one

[10:04] more shot. Let's give 5.5 one more

[10:06] chance to make this actually playable.

[10:08] So, I told it I can't even get the

[10:09] aircraft off the ground and enter

[10:11] flight. We definitely need to make it

[10:12] easier to take off and actually fly the

[10:13] thing. Okay, so it says it fixed the

[10:15] takeoff problem. Apparently the brakes

[10:17] started locked on before. I don't know

[10:19] if that's why we weren't able to do it.

[10:21] Oh, we it didn't automatically set it to

[10:24] takeoff flaps. Yeah, this was we had

[10:26] this on like super super simulator mode.

[10:29] Here is attempt number three at our

[10:31] flight simulator. Let's see how we do.

[10:34] So, can we get off the ground? Oh, we're

[10:36] bouncing on the runway with this time.

[10:38] That's something. All right, cool. We're

[10:40] off the ground. We're actually moving.

[10:43] Let's see if we can get on one of these

[10:45] rings. I mean, the graphics aren't that

[10:48] bad, you know, for something just

[10:50] generated in less than 10 minutes. Um,

[10:53] it seems to be pretty accurate in terms

[10:55] of, you know, it's giving me like my

[10:57] vertical, you know, feet per minute down

[10:59] at the bottom, my actual altitude, the

[11:02] knots, heading, AGL. So, like it's

[11:05] relatively sophisticated in terms of

[11:07] tracking everything. I mean, this little

[11:09] indicator in the front, I mean, looks to

[11:11] be like an angle of attack, you know,

[11:13] indicator, which is kind of cool. So, it

[11:15] has some good stuff going on. The the

[11:19] actual like controls are a little janky.

[11:21] As you can see, I can't control this for

[11:22] anything, but by and large, not bad. You

[11:25] know, we can kind of like kamicazi this

[11:27] and see what happens at, you know,

[11:28] 18,000 ft per minute.

[11:31] But yeah, you know, for 66,000 tokens,

[11:36] about 10 minutes, 15 minutes or so, give

[11:39] or take, you know, with the back and

[11:41] forth. I don't think that's bad at all.

[11:42] So, now let's take a look at DeepSk. It

[11:44] took about 10 minutes to do this. And in

[11:46] terms of tokens, 63,000

[11:50] 44. So, 44 cents, 10 minutes. And here

[11:53] is what Deep Seek came up with for us. I

[11:57] have no idea

[12:00] what I'm looking at.

[12:04] This is supposed to be third person.

[12:06] This is supposed to be the cockpit. And

[12:08] obviously,

[12:09] our first pass with Deepseek was an

[12:12] utter disaster. So, I'm telling Deepseek

[12:14] the simulator is a complete mess. The

[12:16] graphics are completely buggy and I

[12:18] cannot fly anything. Please fix. And

[12:21] here's what our second pass looks like.

[12:24] I still have no idea. Absolutely no clue

[12:29] what the heck Deep Seek is. Oh, hey,

[12:31] there's a plane. You know, there's

[12:33] something.

[12:35] I Yeah, this is This is brutal. And to

[12:38] be honest, I feel like even giving it

[12:40] another prompt to do this, I would need

[12:43] to start getting very, very specific

[12:44] about what we're trying to do, which

[12:47] again like falls pretty short of what we

[12:48] did with Codex. Like it was very, you

[12:50] know, kind of bland prompts. was able to

[12:51] get something at least close even on the

[12:53] first pass. Like this clearly it's

[12:56] completely struggling with the graphics.

[12:58] We are just I I don't even know how to

[13:00] describe this. But hey, it was super

[13:03] cheap. So now let's take a look at what

[13:06] Claude Code was able to give us. For

[13:08] reference, it took 13 minutes to

[13:11] actually execute the plan. The plan

[13:12] itself took 5 minutes. So let's call it

[13:15] 20 minutes to come up with the first

[13:17] pass. And then for total tokens, this

[13:19] run took about 15% plus the 5% before

[13:21] for plan. So we're looking at well sorry

[13:24] we are looking at 11% context plus 5%

[13:28] before. So call it 20 minutes 150,000

[13:31] tokens for Claude Code which is

[13:33] definitely the most expensive and

[13:34] slowest out of all of them. And here is

[13:37] Claude Code's attempt at this. Um for

[13:41] whatever reason we are instantly in the

[13:42] air. We are stalling. We are an IFR. I

[13:46] don't know what's happening. We are

[13:48] about to crash something. Let's Can we

[13:50] save this? Can we pull this out of a

[13:53] dive? No, we're stalling. No, we're

[13:54] dead. Okay, that's interesting. Um,

[13:57] again, it instantly slingshots us into

[13:59] the air. We are in the clouds. We are

[14:02] stalling. I don't know what is

[14:04] happening. We need We need a second

[14:08] pass. So, I wrote, "Upon loading, I'm

[14:10] instantly thrown into the air. It's hard

[14:12] to control. I want to start on the

[14:13] runway, and I want it easier to fly. Oh,

[14:15] and by the way, improve those graphics,

[14:16] too. So, it took about four minutes, but

[14:18] it made some changes. We're going to

[14:20] spawn on the runway. It changed the

[14:22] gear. So, now it's tricycle gear and a

[14:24] few other stuff. So, let's see what it

[14:25] looks like. All right, so here it is

[14:27] again. We are thrown immediately into a

[14:28] fog bank. I'm trying to control this

[14:30] thing. And I just Yeah, there's there's

[14:32] no controlling this at all. All right,

[14:34] we are going to give we're going to give

[14:35] Cloud Code one more chance here. So, I

[14:37] told it's still instantly slingshotting

[14:39] me into the sky. I said, let's go with a

[14:41] much more arcade type feel with the

[14:42] controls. I think we probably should

[14:43] have done that with the initial prompt

[14:45] for all three. I think going for a more

[14:48] realistic sim type thing, it it really

[14:51] struggles to

[14:53] I think do that in a way where it's

[14:55] still user friendly. I think it's

[14:57] probably doing a good job under the hood

[14:59] in terms like okay like angle of attack.

[15:01] All right, you're stalling at this you

[15:02] know angle versus the speed and all that

[15:04] but actually manipulating this from the

[15:07] computer is basically impossible.

[15:09] Although I I think the fog stuff is

[15:11] really strange. So, let's see if after

[15:14] the second round of prompts, it's able

[15:15] to do a little bit better because right

[15:17] now GPT 5.5 did much much better. So,

[15:20] Cloud Code made some more changes, made

[15:22] it more user friendly, and let's see if

[15:23] I'm still going for my instrument rating

[15:25] this time. So, yep, we're still going

[15:28] we're still going for instrument rating.

[15:30] RIP men's here, but you know, I can kind

[15:32] of see it. You know, I can I can check

[15:34] my instrument panel. All right, we're

[15:36] coming off the runway.

[15:38] Um,

[15:40] yeah. Okay. Can I Why is there a tree in

[15:43] the runway? Trying to trying to go up.

[15:46] Can I go up? Can I pitch? Click canvas

[15:50] to lock mouse. I What? Oh, we're in the

[15:54] We're in the air.

[15:56] Nope. Nope. We died. So, yeah, I think

[15:59] this one is pretty clear. Uh, GPT 5.5

[16:04] easily the winner. I think Cloud Code

[16:07] was second place. I would give it second

[16:09] place. You know, it definitely struggled

[16:13] even with the prompts we gave it. We

[16:14] didn't give it great prompts. Let's be

[16:15] totally honest. I think given more time,

[16:18] better prompts, a few more back and

[16:19] forth, we could have got it to where we

[16:21] wanted to go. Like, it was at least it

[16:22] had an aircraft, it had a runway, it had

[16:25] trees in the runway, but it had the the

[16:28] actual things we needed versus Deepseek

[16:31] with Open Go. I have no idea what was

[16:33] going on there. That was a complete

[16:34] mess. I feel like I would have had to

[16:36] start over from the beginning, like give

[16:37] it a very specific prompt. like it

[16:38] wasn't even close to being messed with,

[16:40] but GPT 5.5 right off the rip, you know,

[16:42] with pretty vague prompts. I thought it

[16:44] did really good. 5.5 also used a total

[16:47] of 66k tokens. We're looking at over

[16:50] here with Opus altogether about 200,000

[16:53] tokens. So, a quarter of the tokens

[16:55] essentially quarter the cost and it was

[16:57] a bit faster. I mean, at this point, I

[16:59] don't even care about how open code

[17:01] actually took longer than GPT 5.5 as

[17:03] well. And it just sucked. Let's just be

[17:06] honest, it just sucked. Now let's move

[17:07] on to test number two. This time we are

[17:11] going to be asking them to create a

[17:13] landing page that shows off web GPU

[17:15] shader work using 3JS. Now web GPU

[17:19] shader work is the kind of stuff you see

[17:21] on awards websites. I'm talking websites

[17:24] like Igloo. This kind of thing like very

[17:27] high-end graphics. It looks like a video

[17:29] game. It's essentially using your

[17:31] computer's graphics card to render all

[17:33] this stuff. Now, I don't expect any of

[17:35] these to get anything even close to what

[17:37] we see here, but I want to see what they

[17:39] can do using essentially the shaders

[17:42] technology. This is definitely a step

[17:43] above your basic SAS templated landing

[17:46] page. I want to see what they can do and

[17:48] push them to the limits in the world of

[17:50] web design. Now, I've given all of them

[17:51] a skill that actually breaks down how to

[17:54] do this sort of thing. So, it's not like

[17:56] they're completely in the dark and one

[17:58] also doesn't have an advantage over the

[18:00] other. The only thing I've told them is

[18:01] I want it to feel modern and visually

[18:03] striking. something you would see on

[18:05] awards and to make smart use of GPU

[18:07] compute. So they can pick whatever stack

[18:09] and project structure they like and use

[18:11] good judgment on hero concept UI and

[18:14] interactions. And just like the first

[18:15] test, they're all on plan mode. So let's

[18:18] get started. Okay, so they all finished

[18:19] their plan and funny enough, none of

[18:21] them asked me any questions even though

[18:23] we put them in plan mode. So let's take

[18:24] a look at GPT 5.5 first. So it's telling

[18:28] us it's going to do a fullbleleed

[18:30] interactive GPU driven hero. The concept

[18:32] will be a living signal field with some

[18:35] like dense particle thing it's going to

[18:36] do. We'll see what that ends up looking

[18:38] like. And overall, it's a minimal awards

[18:40] style landing copy. Fully interactive

[18:42] web GPU scene with pointer reactive

[18:44] compute simulation.

[18:46] All right, for Deep Seek, it's a pretty

[18:48] short and sweet plan just like we saw

[18:51] with the flight simulator. Hopefully, we

[18:53] get a better output this time. But a

[18:55] hero section with 75,000 GPU computer

[18:57] particles. I am kind of guessing that

[19:00] all of them are going to go for some

[19:01] sort of like particle theme on the on

[19:04] the hero.

[19:05] So, it's gonna have mouse interaction

[19:07] integration. It'll have a one-time

[19:09] initialization. And then we should see

[19:11] stuff like bloom, chromatic aberration,

[19:14] a custom vignette, and some film grain.

[19:16] So, we'll see what that actually ends up

[19:18] looking like. And then we have Opus

[19:20] 4.7's playing again going for this

[19:22] particle thing with bloom, and it's

[19:24] going to be interactive with the mouse.

[19:25] So, we'll see if any of these actually

[19:26] look different because on the surface

[19:28] all their plans sound very similar. So,

[19:30] the first one done was 5.5. It took

[19:33] about 6 minutes. And in terms of tokens,

[19:35] we've used 107K. So, let's see what it

[19:39] built us. And here's what it created for

[19:42] us. Now, this is very bright. Um, so

[19:45] it's hard to even see the actual

[19:47] particles, but you know, as we scroll up

[19:49] and down, it does have an animation

[19:52] going on in the background as well as,

[19:54] you know, some subtle color changes.

[19:56] Looks like right now our mouse is

[19:59] supposed to attract the particles and we

[20:01] have I'll move this over here. It gave

[20:04] some options for like repelling it

[20:06] versus drift, but again, it's kind of

[20:09] tough to to see it um due to how bright

[20:11] it is. So, I told that it's hard to

[20:13] actually see the particles due to the

[20:14] brightness. It also takes over a lot of

[20:16] the hero text. So, can we turn down the

[20:17] brightness a bit and also push it to the

[20:18] right a bit more because right now it is

[20:21] kind of overpowering. You can't even

[20:23] really read the text over here on the

[20:25] left due to just how freaking bright

[20:26] these particles are. And here's the

[20:28] update after the second run. It's a

[20:30] little bit better. It isn't as

[20:31] overpowering and leaves some room for

[20:34] the text. Um, although I will say it

[20:37] it's kind of blurry almost, but you

[20:40] know, it's not bad. Like it it set out

[20:43] to do what we told it to do given the

[20:46] somewhat vague prom. So, I'm not blown

[20:47] away by sort of the design it came up

[20:49] with, but I'm not like upset about it.

[20:51] Now, let's take a look at Clawude Code

[20:53] because as we've been doing all this,

[20:55] Deep Seek is still over here in the

[20:57] trenches trying to figure this out. And

[20:59] here's what Claude Code gave us.

[21:03] So,

[21:05] kind of nothing.

[21:07] I'm not sure if it's saying the

[21:10] background. I guess the entire

[21:11] background is supposed to be

[21:16] the WebGL. I'm I'm assuming it's very

[21:19] understated,

[21:21] which I guess is something you could

[21:23] totally do. I mean, like on screen it

[21:25] doesn't look like it looks kind of cool,

[21:26] but I I'll be honest, I was looking for

[21:28] something a little more flashy. So on

[21:31] the second pass when I told it to make

[21:32] it a bit more flashy, there wasn't a

[21:34] huge difference. Although like it's it's

[21:37] really subtle. There's kind of like this

[21:39] film grain almost like this blur that

[21:42] goes from bottom to top. So it's a

[21:44] pretty subtle thing. And you can see

[21:46] here on the bottom it tracks like the

[21:48] frames per second. It's using 250,000

[21:50] particles. So I mean honestly it looks

[21:54] cool. It's just not super flashy. So

[21:56] it's definitely like a taste thing. Now,

[21:58] total tokens on the cloud code side was

[21:59] about 175,000 and it took just slightly

[22:02] longer than 5.5 inside of Codeex. Now,

[22:06] let's take a look at DeepSk, which has

[22:08] taken 116,000 tokens at this point. It

[22:11] took the longest um as well, but total

[22:13] cost, we're talking again under a

[22:14] dollar. And here's what it gave us. So,

[22:18] it's

[22:20] kind of this particle field thing that

[22:22] uh somewhat follows my mouse.

[22:26] Interesting. I think it might give you

[22:28] like an epileptic seizure,

[22:31] honestly. Um, beyond that, it's pretty

[22:33] bland. Um, the flux, you know, text

[22:37] right here kind of changes colors, but

[22:40] yeah, pretty much just created this this

[22:42] thing. And after telling Deep Seek to do

[22:44] another pass, it then came back with

[22:46] this where now it kind of has like some

[22:48] weird parallax thing. It's got some like

[22:51] blue stuff going on in the background.

[22:53] And now this thing looks like a UFO,

[22:55] which kind of responds to your mouse,

[22:58] but

[22:59] yeah, it's it's it's something. And

[23:02] overall, the token count from Deep Seek

[23:04] was 130K tokens, coming in at a $143.

[23:08] So, after all those tests, where does

[23:12] that really leave us? So, now let's talk

[23:13] about the final results. When it comes

[23:15] to test number one, which was the flight

[23:17] simulator, clear winner. That was GPT

[23:19] 5.5 inside of Codeex. It was quicker

[23:22] than Opus 4.7 inside of Cloud Code. It

[23:25] was also faster and the end result was

[23:27] by far the best. Deepseek did terribly

[23:31] in the flight simulator. It wasn't even

[23:33] close to what we were trying to do. I

[23:34] would have had to continue to prompt it,

[23:35] prompt it, prompt it to even get it to

[23:37] like close to the first pass from 5.5.

[23:41] And Opus 4.7 and Cloud Code was like,

[23:43] eh,

[23:45] it wasn't awful. like it really didn't

[23:47] work at the beginning, but after a

[23:49] couple prompts, you could tell we could

[23:50] get it to a place where it was

[23:51] equivalent to what GPT 5.5 was doing,

[23:54] but that would have taken more prompts.

[23:55] It would have taken more time, and

[23:57] ultimately would be more expensive. So,

[23:59] clear winner for 5.5. In terms of the

[24:02] web GPU landing page, again, DeepSeek

[24:04] struggled here. I was not a fan of this.

[24:06] I don't really know what this is

[24:08] supposed to be. Sure, I didn't give it a

[24:09] super great prompt, but like is this

[24:12] what we're going to be getting as a

[24:14] baseline median outcome if I don't like

[24:17] grab Deep Seek by the reins and really

[24:20] force it to do something? I guess so.

[24:22] Now, when we compare Opus in 5.5, I

[24:24] would have gone with Opus 4.7 and Claude

[24:27] Code with how it handled the web GPU

[24:29] thing. I think that has to do with sort

[24:30] of a taste kind of deal. Yeah, you could

[24:32] argue the 5.5 was flashier, but I

[24:35] thought it was kind of ugly. Um, again,

[24:39] in all these tests, we kept the prompts

[24:40] rather vague to see what sort of path it

[24:42] would go down. So, I would definitely

[24:44] give Opus the lead here, although it was

[24:48] more expensive and it also took slightly

[24:50] longer. So, if they were given a more,

[24:53] you know, hands-on prompt that was very

[24:55] specific about what she wanted to do

[24:57] because 5.5 did what we wanted it to do,

[24:59] like it did create a web GPU landing

[25:01] page. I just thought it was ugly. So, it

[25:04] still completed the task. It just didn't

[25:06] complete it as well, I think, as Opus.

[25:08] Now, big picture, what does it mean if

[25:10] we take all that together? Well, I think

[25:11] it means great news for anybody who's

[25:15] using agent coders. We have options,

[25:18] right? You can use Opus and Claude Code

[25:20] or you can use GPT 5.5 and codecs.

[25:23] You're not wrong with either. I think

[25:26] it's totally a personal preference at

[25:27] this point. And the best part is if you

[25:30] go down the cloud code route, it pretty

[25:32] much all applies to Codex. If you go

[25:34] down the codeex route, it pretty much

[25:35] all applies to clawed code. So, I don't

[25:38] really think there's vendor lock in the

[25:39] sense like, oh, I've only learned about

[25:41] clawed code, like I can't go to codeex

[25:43] or vice versa. That's not the case at

[25:45] all. If you're doing this the right way,

[25:46] what you're really learning is AI

[25:47] fundamentals and how to build things.

[25:49] And that applies to both of these guys.

[25:51] And the more competition, the better it

[25:53] is for us, the consumer. Now, as for

[25:55] DeepSeek, n I don't know. I wasn't very

[25:59] impressed. This might be a situation

[26:01] we're like, okay, like Deep Seek makes

[26:02] sense if we're doing simpler tasks where

[26:04] we just don't need the power of

[26:05] something like Opus or we just don't

[26:07] need the power of something like GPT 5.5

[26:10] because remember we're talking about

[26:11] something that is eight times cheaper.

[26:13] Sure, I didn't like the web GPU landing

[26:16] page this thing came up with, but was it

[26:18] eight times worse?

[26:20] Maybe, maybe not. Kind of hard to

[26:22] actually, you know, articulate that and

[26:24] quantify that, but obviously that's

[26:26] something we need to take into account.

[26:27] So, you know, I don't think it's really

[26:29] competition to be frank with 4.7 or 5.5.

[26:33] I think though if you're doing simpler

[26:34] task and you're like very token

[26:36] conscious, very cash conscious, then

[26:38] hey, maybe Deep Seek makes sense for

[26:40] you. So, that's all I got for you guys

[26:42] today. I hope that sheds some light on

[26:44] these three models and how they kind of

[26:46] stack up to one another. I think it's a

[26:47] great time to be in the space. More

[26:49] competition is better for everyone. So,

[26:52] as always, if you want to get your hands

[26:53] on the Claude Code Masterass, make sure

[26:55] to check out Chase AF Plus. There's a

[26:57] link to that in the description and I'll

[26:59] see you