(Transcribed by TurboScribe.ai. Go Unlimited to remove this message.) In this video, I am putting the newly

released GPT 5.3 codecs to the test

in a new benchmark that I am calling

BridgeBench.

BridgeBench is going to put GPT 5.3

codecs to the test from giving it tasks

like building an FPS game inside one prompt

to building the Windows 10 interface all in

one shot.

If you guys would like to check out

the prompts that I'm going to be using

in BridgeBench and even test it out yourself,

you can just go to bridgemind.ai and

go to BridgeBench.

There'll be a link to this down in

the description below, but you'll be able to

actually see and copy the prompts that I'm

using for this test that you're going to

see in this video.

Before we get too deep into the video,

we are only 25 members away from hitting

5,000 members in the BridgeMind Discord community.

So if you guys are watching and have

not already joined the fastest growing vibe coding

community on the internet right now, make sure

you check out the description below and join

the BridgeMind Discord community.

Also, I do have a 200 like goal

on this video.

So if you guys haven't already liked the

video or subscribed and turn on post notifications,

make sure you do so.

And with that being said, let's get right

into the video.

Let's now open up BridgeSpace, which is the

ADE, the agent development environment that we've been

working on that's going to ship in the

next week or so.

And what I'm going to do now is

I'm going to open up a new workspace

in the BridgeBench directory.

So let's go over to my desktop and

let's select BridgeBench.

And we are going to launch six codex

agents inside of this workspace.

So I just select the directory.

I select how many terminals I want to

open up and I select which AI agent

I actually want to use.

And now I'm going to create this new

workspace.

And just like that, we have six codex

agents that are ready to go.

So I'm actually going to close one of

these because we only have five prompts here.

So we need five codex agents.

So what I'm going to do is I'm

going to give a quick prompt and I'm

just going to I'm going to add the

prompt one.

And I'm just going to say, I want

you to I want you to work inside

of the GPT 5 3 codex directory and

complete this project.

OK, so we're going to input this prompt

and we're going to use this same exact

prompt.

I just use that was BridgeVoice that I'm

using for my my prompts there.

That's another product in the BridgeMine suite of

Vibe coding tools that we have releasing.

So we're just going to add all of

these prompts that are in the BridgeBench and

paste in the same exact prompt here.

And we're going to do number four now.

So it's prompt number four and it is

now prompt number five.

So here we are.

We have all of these ready to go.

And just like that, let's just submit all

of these.

They're all in extra high.

They're all going to work in the GPT

5.3 codex directory.

And we're going to let these work because

as you guys know, one thing about GPT

models is that they are a little bit

slower.

They think for longer, especially when you're in

this extra high mode.

So we are using GPT 5.3 codex

extra high for this.

But while these actually work, you can see

that each of them are working in their

respective directories here.

But what I want to do now is

I actually want to take a look at

the benchmarks while these are working that OpenAI

released when they released GPT 5.3 because

I think that benchmarks are a very valuable

thing to look at for us to understand

what this model's capabilities are.

So while these all work and give us

a good look of what this model is

capable of, let's now take a look at

the benchmarks while these complete in the background.

All right.

So one very important thing that you guys

need to know is that GPT 5.3

codex released the same day as Opus 4

.6 did.

They basically released this at the same time

for composition sake so that people would be

able to be like, oh, which one's better?

Which one's better, right?

Now, a lot of people are saying that

GPT 5.3 is better.

But one thing to know is that GPT

5.3 codex is not available via the

API.

You can only use this inside of OpenAI's

codex right now.

So if you're using the codex CLI, you

can use GPT 5.3. That's the only

place that you can use it right now.

So it is a little bit unfortunate because

we also cannot see GPT 5.3 codex

inside of LM Arena.

So if we go to LM Arena real

quick, because it's not in the API, we

are unable to see it inside of the

LM Arena.

So you can see here Opus 4.6

is absolutely crushing it on LM Arena at

1576.

So it's dominating GPT 5.2 high.

Now, we don't know how that will rack

up once GPT 5.3 codex is actually

available inside of the API and is able

to be benchmarked on these leaderboards.

But Opus 4.6 is actually crushing it.

But in terms of GPT 5.3, it's

not yet available in the API.

So we're unable to see it on the

benchmarks.

But let's go over here and see the

SWE Bench Pro.

So this is how it compares to GPT

5.2 codex.

So you can see that that's actually a

bad chart.

Like here is one that's very, very interesting.

The Terminal Bench 2.0. So this is

one of the biggest jumps that OpenAI did

with this model.

So you can see that on the Terminal

Bench, it jumped from 64% in GPT

5.2 codex to 77.3% with

GPT 5.3 codex, which is actually a

huge jump.

And it does beat out Opus 4.6

in the Terminal Bench by quite a bit.

So that is a big deal for us

as people that are using CLI tools.

So here's a little example that they gave

on what the capabilities are of these differences.

They basically gave it the same prompt.

So interesting there to be able to see

that.

But let's check it out down here.

OS World Verified.

You can see a huge jump there.

So that is an agentic computer use benchmark

where the agent has to complete productivity tasks

in a virtual desktop computer environment.

So this is one thing that I want

to kind of highlight here because I've talked

about it a little bit on stream, but

I think that computer use is going to

start to become a bigger deal in 2026.

We saw that with CloudBot or OpenClaw coming

out and people using it for computer use.

So that's big as well.

It saw a huge jump there.

They have a lot of text here.

Can we get?

Okay, here we go.

Finally, in the appendix, we have some of

the information here.

So one thing to note, and this came

with both the Opus 4.6 drop as

well as the GPT 5.3 drop.

In both of these models, we did not

see a significant improvement in the SWE bench

marks.

So this is the SWE Bench Pro benchmark,

but you can see that it hardly improved

from GPT 5.2 to GPT 5.3.

We saw a 0.4% improvement on

the SWE Bench Pro.

Now on the terminal bench, we saw a

huge improvement, which is great.

And in these other areas, we also saw

a pretty big improvement.

But with SWE Bench Pro and SWE Bench

Verified, which is also an important benchmark, we're

not really seeing the massive jumps that we

saw.

For example, from Opus 4.1 to Opus

4.6, there was a huge jump in

capability.

Same thing from GPT 5 to what we

now have, GPT 5.3, a pretty large

jump in coding capabilities.

But this just seems like a very small

iteration when you're actually looking at the benchmarks.

But I will say after having used the

model for hours and hours and hours on

end, I will say that that small jump,

that small iteration has made a difference as

well as some of the other improvements in

the models.

In my personal workflow, I have noticed it

one-shot things that I've been struggling on

doing with the previous models, GPT 5.2

or even Opus 4.5. I did notice

that GPT 5.3 was able to one

-shot things that it was not able to

do before.

So I will say that, but that's one

thing I want to highlight is that in

terms of the benchmarks and the coding benchmarks,

we're not seeing some massive jump with 5

.3 or Opus 4.6 for that matter.

But specifically with 5.3, we're seeing only

a 0.4% increase here.

So not going to be a massive improvement

in coding capabilities, but still a small improvement.

So maybe that small improvement helps us do

a lot.

So that's the benchmarks, that's what you guys

need to know.

Just to summarize, not super big jumps in

coding, but some pretty big jumps in other

areas like the Terminal Bench 2.0 and

OS World Verified.

So those are some pretty impressive jumps, but

coding, not that great.

But with that being said, let's go check

up on our coding agents and see if

they're done yet.

All right.

So all of our agents have now finished

working.

So just as a reminder, we are building

an FPS game, a Windows 10 interface, the

flight simulator, and bridge trade stock app, as

well as a new bridge mine landing page.

So let's actually now pull these up and

see what was created starting with the new

Windows 10 interface, which was created by chatGBT,

or sorry, GBT 5.3 codecs.

So here we are.

So we have a Windows interface.

Let's just click around a little bit.

Let's check out the command prompt and see

if this works.

So let's do an LS.

So we have desktop.

Let's CD to our desktop.

That's interesting.

Okay.

So let's CD out.

So how can we...

So this looks like it works.

The command prompt works.

It looks like we have a calculator.

Does the calculator work?

Let's do six times seven equals...

Okay, 42.

Okay.

So that looks like it works.

You can see the calculator.

When I dismiss it, it stays down here.

If I close it, it gets removed.

That looks nice.

What's this?

This is a notepad.

Hello there.

Can I save this?

Let's see if I can actually save this.

Let's save as, and we'll save it to...

Let's save it in our desktop, and we'll

just title it test.

And if this shows up on a desktop,

that's actually interesting, which it doesn't look like

it does.

So it looks like we just saved it,

but it didn't actually save.

Let's see file explorer and...

Oh, here's test, and here's hello there.

Oh, wow.

Okay.

So it did do that correctly.

There's the test.

Now, for some reason, it doesn't actually show

up on my desktop here.

Can I drag and drop this into other...

Okay.

I can't drag and drop.

Can I just...

Okay.

I can delete these.

I can rename them.

Remember, this is just in a simple HTML

file.

So it was able to do all of

this in one shot.

If I go to my recycle bin, here's

what that looks like.

You can see that test.txt that I

had was now moved to the recycling bin.

So, I mean, that's pretty impressive.

Microsoft Edge.

I mean, I don't think this is going

to work because it's in an HTML file.

So, yeah, it's not going to work.

So that's decent.

Now, photos.

What is up with this?

Photos is all right.

Control panel.

Here's everything in the control panel.

Can I actually change this to...

Oh, wow.

Okay.

So I can adjust the taskbar color.

I can adjust the window color, which is

pretty cool.

I can turn on dark mode, dark mode,

light mode.

I mean, this is pretty basic, in my

opinion.

It's all right.

You can adjust the brightness, resolution.

I don't notice any difference with the resolution.

But settings.

Settings is interesting.

Paint.

Can we paint?

We can paint.

So that's cool.

And remember, you know, this is all in

just one shot.

We can erase.

We can fill.

Green.

Lines.

Rectangle.

I mean, this is...

Wow.

This is pretty cool.

So that does work.

Now, with this, I mean, I guess this

is all the functionality that it has.

So if we do the windows, like, here's

what it has.

The weather.

I doubt that the weather works.

Yeah.

So, I mean, this is just in a

basic HTML, but it did do a decent

job.

I mean, it's got the time correct.

You can see it down here.

1103272026.

It's got the got the Wi-Fi.

I don't think it's actually connected to my

Wi-Fi, but we can adjust the volume.

Battery.

There's not actually any battery, but I mean,

for a one shot, this is okay.

Obviously, you can see, like, it's not like,

I mean, it's not Windows 10, but, you

know, it did a pretty good job for

one prompt.

So I would give this maybe like a

6 out of 10.

You can move everything around.

I don't know what else you guys would

like to see, but, I mean, it did

build the Windows 10 interface.

Not exactly Windows 10, but definitely was able

to build something pretty, pretty unique here that

reflects Windows 10.

So I would give it a 6 out

of 10.

All right.

The next thing that I want to test

out is this FPS game.

So first of all, okay, so it's called

Neon Breach.

Browser FPS built with a pure JavaScript raycasting

engine.

Controls are WASD.

Mouse look, shift, sprint, 1, 2, 3, switch

weapon, R to reload, left click to fire.

Okay.

So let's go and first of all, like,

I do want to pull up cursor and

I want to see how many agents, how

many lines of code did these agents actually

create?

So for the Windows interface, let's see this.

So that did, the one shot for the

Windows 10 interface was 5,308 lines of

code.

For this FPS game, it's only 1,623.

So let's click to start.

So let's just mouse look, shift to sprint,

1, 2, 3, switch weapon.

Okay.

Let's start.

So one, whoa.

Okay.

First of all, the, there's a map in

the left corner here.

There's, you can see the map in the

top left corner here.

Oh, am I taking damage?

Oh, shoot.

I'm literally taking damage right now.

The, okay.

First of all, the styling isn't that great.

What is this?

Okay.

Shotgun shells.

How do I get a gun out?

How do I shoot?

Whoa.

Okay.

So I'm shooting, but I'm also getting hit.

Am I getting killed?

What, how am I getting, okay.

Game over.

Enemies killed.

So I arrived for 29 seconds.

Accuracy was 17%.

Let's restart.

Okay.

So this is actually like not that good.

So shotgun shells, can I, I can't even

shoot.

I'm, I'm, I'm doing, okay, wait, hold up.

Oh, whoa.

Okay.

So I think I did shoot it.

Okay.

So I'm sprinting, I'm pressing shift.

Okay.

I think I did.

Was that me killing it?

Okay.

I think that was me killing it.

Okay.

I mean, the graphics are horrible.

The animation's not good.

There's health over here.

I mean, there's a map in the up,

upper left-hand corner.

The styling that it did, like the spacing

that you guys see, not good.

They're also like, I just don't even understand.

Like, it just wasn't very creative, if I'm

being completely honest.

Like, hey, the map is nice.

There's like a ton of people in over

here.

Can I like get over here?

Okay.

I mean, there's no animations.

There's like, I wish there would be like

some sort of, okay.

It's reloading my shotgun.

So reloading works.

Okay.

Oh, I mean, there's, I wish there was

like some type of, you know, gunshots or

something like that.

Here's health.

Here's, what is this?

Machine ammo.

How do I?

Okay.

Okay.

I switched my machine gun.

Oh, whoa.

Okay.

Game over.

We survived for 79.5 seconds.

Apparently, my accuracy was 217%.

So I don't know.

I would give this like a two out

of 10.

Um, this was not very good.

And you know, this could go back to

our system prompt that we gave it.

And just that it, uh, was maybe like

misunderstood our system prompt, but you can look

at this and like, obviously the animations and

what it created here, not super impressed by

what it did.

So that could tie back to our system

prompt.

We may need to, you know, update that

system prompt for future bridge bench tests, but

I am not impressed with what Codex did

here.

I'm next thing that I want to check

out is the front end styling.

So this is the remake of the bridge

mind website that it made.

So, you know, one thing that GPT models

have been notoriously bad at has been styling.

Now, as I've used to GPT 5.3

Codex in the past couple of days, I

will say that when given the correct information

and prompts, it has been doing a much

better job than what we had with like

GPT 5 for sure.

But what I will say with this is

it's not that great.

What it created here.

I mean, let's, um, let's try shrinking this.

How's the, I mean, I would say that

the responsiveness is actually pretty good.

Like the responsiveness and spacing is, is pretty

good.

Um, it could do a couple of things

better, like maybe making the, not those buttons,

not wrap and just making them, you know,

compact a little bit more, um, and adjust

to the screen size, but, um, overall, definitely

better than what we've seen with previous GPT

models.

I will say that, Hey, with like one

shot with no styling guide, um, it's doing

a pretty decent job.

It just created this landing page.

I believe.

Yeah.

Like if you, all these buttons will just

link to, uh, to different parts of that

page.

So I would give this a four out

of 10.

Um, I'm not super impressed by like the

background gradients.

Like this is like kind of your standardized

AI look.

I don't feel like this is really unique.

You know, sometimes when you're using Gemini three

pro, you can get these really unique UIs.

Uh, but with GBT, I would say that

it didn't do a great job on this.

Now with that, I will say like, uh,

for example, with bridge voice, I was having

GBT 5.3 codex, make a couple updates

to bridge voice and all the UI that

you guys see here.

Um, like even like these, these theme changes

and whatnot, these were all made by a

GBT 5.3 codex.

And this one had a little bit more

of a structured styling in place to begin

with.

And GBT 5.3 did do a really

good job making it like very uniform.

So that's just one example that I will

show you.

Like from my personal work, when you're getting,

when you're actually using 5.3 in practice

and you're giving it the correct style guides

and you're giving it the correct references and

you're prompting it really well, uh, 5.3

can do a really good job at improving

your styling.

It definitely did in the case of bridge

voice here, but for just a one shot,

like off rip, just singular prompt, not super

impressed with this one.

I'm going to give it a four out

of 10 for the styling on this website

with one shot.

Now for another cool one, let's go to

the flight simulator.

So here it is the flight simulator.

So it says procedural terrain, full flight model,

dynamic weather and waypoint challenge.

Um, so here, so to pitch w okay,

so it's WASD.

Um, so pitch roll, we can do rolls,

a arrow left, down arrow, right.

Okay.

Um, rudder.

So there's some like actually some decent sorts

of throttle shift up control down camera chase,

uh, cockpit.

Okay.

So let's, uh, let's try this.

So camera, see chase slash cockpit.

Okay.

Whoa.

Okay.

All right.

So here is the look here.

So let's try moving around a little bit.

Okay.

So first of all, whoa, uh, codex, what

are you doing?

Okay.

So one thing I will say is we

are completely inverted.

I mean, I don't know if this is

like straight out of flight, if you guys

have seen that movie, but, um, I don't

know why the mountains are above us, but

that does not look right.

There's the weather change.

There's the daylight change.

Um, so control C is going to change

the view here.

Um, you can see that we have like

an altitude checker over in the top left.

What is that?

It's VSI waypoints.

Uh, I have a buddy.

That's a flight.

He is a airline pilot.

So, you know, he may understand how to

fly planes.

I personally don't, but when you're in this

mode, it looks like it's, um, man, there's

like a lot of glitches with this.

This is not smooth at all.

Um, let's go back to here.

And we were like literally in the matrix

right now.

Uh, we're in a bunch of green, very

glitchy.

Uh, there are like some interesting things like

on the left and right.

I mean, it is like, looks like a

flight simulator, but why are the mountains in

the sky?

Why am I in a green blob?

I'm going to give this a two out

of 10.

Uh, this is not, this is not good.

Uh, you know, we may want to, at

some point, adjust the prompts to be a

little bit less specific and maybe let codex

have its way with it.

Um, so let's actually try that.

So we're going to give this a shot.

We're going to try and rebuild this flight

simulator, but rather than giving it such a

distinguished particular prompt, we're just going to give

it something basic here.

Let's exit.

Let's pause this real quick and let's actually

pull bridge space back up.

And what I want to do is let's

actually launch a new workspace, uh, inside of

let's, let's pull back bridge bench up.

And, um, let's just launch a singular codex

instance in this workspace.

And what I want to do is I'm

going to just give it this prompt.

I want you to create a new flight

simulator that does not follow the readme or

the prompt in the existing flight simulator.

This is to be a completely newly invented

flight simulator.

All I want you to do is create

a flight sim that is simple and accurate

and allows me to fly around a map

in a plane.

Okay.

Let's just try that.

And we'll, we'll see if that's a little

bit better.

Um, you know, and we'll see if it,

you know, using its creativity, if maybe the

system prompts that we're giving it is actually

causing it to like hallucinate.

So we'll let it kind of be a

little bit more creative on its own and

we'll give it less of a, uh, less

of like specific instructions that we want it

to follow.

And we'll more so just kind of let

it run and do its own thing.

So let's say I want you to put

the code in, in, in a flight sim

sim to dot HTML file.

Okay.

So we're going to give this, and then

we're going to review the, um, stock trading

application that made, and we'll let this one

work.

And then we'll come back to this and

finish off the video, but with a review

of this, but so far with that existing

flight simulator, yeah, that was, this is not

good.

This was not good.

I'm going to give it a two out

of 10.

Literally the mountains are in the sky.

Uh, yeah, two out of 10.

All right.

So here is bridge trade.

So this is what it created for, uh,

basically the prompt that was requesting it to

create a website where I would get real

time stock analytics and be able to see

like stock market prices and stuff.

And right off the bat, I'm going to

say like, this is a pretty big fail.

Um, and I'll show you why like, Hey,

like this is not the price of Apple.

It's updating live.

It's literally Saturday right now.

So I don't know why it's actively updating.

So obviously this does not connect with like

any free, there's like a bunch of free

API APIs that you can integrate with.

So I don't know why Codex wasn't intuitively

able to search that up, identify, you know,

what it would be able to do to

build this and then be able to build

that in.

So that's a little bit discouraging there.

And then in terms of the styling, like

this just has absolutely zero creativity at all.

It tried to create like some logo that

looks bad.

Um, so I'm going to give this a

one out of 10.

I think this is a complete fail on

this prompt.

Um, and uh, yeah, that's, that's not good.

We're going to give this a one out

of 10, but let's go check back out

on the flight simulator and see how it

did.

All right.

So here's the conversation for the flight simulator.

So this actually completed very quickly.

So this one ran and it only created,

let's see here.

So this one was 600, 721 lines of

HTML and it was able to only work.

It didn't work for very long.

So let's like compare that first off with

the other, other that it created.

Right?

So let's go here.

You can see this flight simulator, the one

that it created initially, this one was 2

,500 lines.

So this one, you know, was 2,500

lines, about what?

Close to like three and a half times

larger.

Let's now go and let's do flight sim

two and go to flight sim two here.

And um, let's see, what is this?

Whoa.

Oh my gosh.

Okay.

Uh, so this one, yeah, so this, this

one, I don't even know what that is.

All right.

That is not good.

Okay.

So this one gave us a different approach.

It gave us a different look and I

don't even know what I'm looking at here.

This is a, this one's like really, really

bad.

So, um, yeah, I mean, I would say

this is obviously a one out of 10

as well.

So I, I don't know if, how, if

we could potentially be making, sorry, I've just

put the OBS, but I don't know if

we can make our bridge bench better to

really be putting these to the test, but

I just wanted to put GBD 5.3

to the test with some like system prompts

to be able to test like, okay, in

one shot, what is it able to create

in terms of like the windows interface.

But you know, in practice, when I've been

using GBD 5.3 in my normal workflow,

like you can even see this here, it

has been able to do a very good

job in creating real SAS products for me.

And the styling, I have found that the

styling is a little bit more creative.

Um, you know, I think that for us

to be able to look at like some

of these things like these here, these, these

sections here on my website about bridge code

and about the bridge mind MCP and about

the bridge space, the agent development environment and

bridge voice, you know, these were created using

codecs.

So I think that for styling and when

we're actually applying it, it's like, Hey, when

you're actually using the tools in practice, it

is a little bit different than just saying,

Hey, go build a flight simulator in one

shot.

Right.

It's like, I think that, you know, I

think it'll be nice to have some type

of bridge bench and I want to build

this out.

Maybe the community can help, um, as we

create some better system prompts to be putting

these different models to the test.

Um, but you know, I think this was

just a fun way to be able to

put it to the test, but all in

all, some of these things that codecs created,

I'm actually like not super impressed with.

Right.

I mean, what the best thing that it

created was that windows interface.

And then like some of this stuff just

wasn't that good.

Right.

So I don't know, I don't know how

I feel about that.

I think that with the Opus 4.6

with a one M context window, there's going

to be room for both in our workspace

and our, in our workflows.

Right.

But I will say this, this styling that

it did here, I really liked it.

And I think it did a good job.

Like even like, Hey, the bridge voice tab

here.

Right.

Like I had to create a bridge mind

theme.

And what's really cool about this is like,

you can see when I hover over this

now it's that bridge mine gold.

Right.

So I just thought that that was cool.

Like it did a good job, like picking

up on like really, you know, unique details

like that, as well as like, you know,

building stuff like this collapse feature, like, right.

Like that looks really good.

And Hey, the real state of software development,

when you're using vibe coding and using AI

models to help you is that it's not

a one shot.

A lot of people think about vibe coding.

They think, Oh, like the AI models can't,

you know, it can't create a flight simulator

in one shot, or it can't create, you

know, the winners windows 10 interface in one

shot, but really like what people are using

by coding for right now, or at least

what I'm using by coding for right now

is you build and then you iterate, right.

So you actually get the products working right.

And you kind of just go one step

at a time and you slowly iterate and

improve coding with the AI models via vibe

coding, but then you ultimately get something like

bridge voice, or you get something like bridge

space, which is a great product, right.

And these products are going to be launching

soon.

So I think it's a great opportunity for

us to start to build out something like

the bridge bench.

You know, I don't know how I feel

about these initial five prompts that I tested

with it.

I may want to change that over time,

but I think that with GBD 5.3,

you know, it's not like a substantial improvement

in coding, right?

Like we saw that with the benchmarks, but

even here, like, Hey, you have a flight

simulator, right?

And the mountains are in the sky, right?

And then you have a FPS game and

there's no animations for the bullets.

And there's no like animations for the characters

when you actually shoot them to like show

it that you actually like killed them or

whatever.

Right.

So it's like, you know, there's certain details

and intuition that GBD 5.3 codecs, it

does not have, um, that may need some

coaxing from the developer, right?

And that's why, Hey, that when you are

vibe coding, it's very important that you need

to understand, Hey, AI is just an extension

of the developer.

So if you understand, um, these different concepts

and you're able to prompt it correctly, then

you're going to be able to use a

model like 5.3, that will do a

good job for you.

Now, one thing I will say that I

did notice during this test is that these

prompts actually finished relatively quickly.

Like right here, this wrote 4,338 lines

of code here in HTML to build a,

what was this?

This was a, the, the bridge trade, right?

And it did do it in like less

than 10 minutes.

So I think that the substantial improvement that

we're seeing is like, I think that I

am seeing a speed up with 5.3.

Now this could be because they switched and

they partnered with Cerebrus and now they have

better and higher, you know, faster inference.

But I will say that I definitely have

noticed a speed increase.

I also have noticed some other intuitive things

such as it being able to run terminal

commands a little bit better and being able

to work in and launch different sub-agents.

You know, you can see that this one,

you know, worked for another three minutes on

doing a final polish pass before it, you

know, it did this polish pass.

So, you know, there's certain things and it's

right there.

Sorry about that.

So, you know, there's certain things that I'm

seeing with this model that are definitely an

improvement that are that like, Hey, I'm happy

that we have the model, but this model

is not the substantial jump that we've seen

with other models.

For example, like moving from O3 to GBT5,

right?

Like that was a substantial jump or 4

.0 to O1.

That was another substantial jump that OpenAI did,

right?

So I'm more looking forward to the next

iteration of whatever they come out with, whether

it's GBT6 or GBT5.3, because we didn't

see that substantial jump in the coding index.

And that's really what we need to look

at is, Hey, SWE Bench Pro, SWE Bench

Verified.

These are the coding benchmarks that we need

to be looking at.

But I will say, obviously it's a slight

improvement.

It is a speed up.

I'm going to be using Codex more on

my workflow.

I did purchase the chat GBT Pro plan

for $200 a month.

So, you know, sometimes I think that people

are asking the question, they're saying, Hey, which

model is better?

Opus models?