The cheap model won! I ran the same prompt

through 6 different AI models, and

GPT-5.2-Low just beat Opus 4.5 for a code

review.

I needed a code audit on YeeBall, which

is my multiplayer soccer game, so I had

Opus write the perfect review prompt.

About 52,000 tokens of instructions. And

then I fed that exact same prompt to every

major model.

So check it out. Gemini went rogue. It

started editing files when all I asked it

was for a code review.

But the real surprise was the model with

the lowest reasoning effort found things

that Opus 4.5 missed.

Crazy, right?

So GPT-5.2 low caught an interesting

issue with the way that sprites were

rendered in my app.

So for procedural tasks, you may not need

to be reaching for that expensive thinking

model if you just need some clarity.

Let me show you what exactly happened.

Code review handoff.

First, let me review your current work

documentation.

I've created a comprehensive handoff

review doc.

This prompt will then guide the reviewing

agent to the core objective, what's

included.

Opus is so good. So I had Opus

orchestrate this. This only cost me about

52,000 tokens and now I'm going to have

this handoff document and just switch this

over to

Gemini. Okay, so what's included? Full

audit checklist for phases one through

seven.

Cool. So key areas of focus are here. The

document also plainly marks what to

ignore,

and so on.

And of course, the power of having a

smart model like that in your codebase is

massive.

So we're going to test out Gemini 3

Flash.

How do I prompt it?

Let's see.

So cool.

Please work on this.

Gemini 3 Flash has a million token

context window.

This is going to be a really nice

extensive review and you can see Cursor's

already put

all of those things in here into to-do

lists and it's working extremely fast with

reading

all these files.

Very, very quickly. Whoa, does Gemini

already finish? No way. This is moving

fast.

Okay.

It's getting to work.

Gemini 3 Flash is fast, but if you have

Opus handing off the prompts to

Gemini 3 Flash, it's amazingly fast. Ooh,

I've updated to-do lists and the progress

made. Okay.

So would you like me to proceed by fixing

or would you like me to start working on

this animation thing?

I have completed all to-dos now, and I'm

ready for the next task.

It's interesting how Gemini 3 is kind of

talking to itself out loud here.

Actually, looking at the handoff, it says

you're tasked with conducting a thorough

code review and

work on this. I should probably fix the

kick animation now since it's broken. So

this is

interesting. The Google Gemini Flash

internal dialogue is sort of being exposed

right now in

Cursor because it's like, it could either

be the system instructions from Cursor or

Gemini's instructions to want to like

really get the work done but this document

only outlies the

code review task right so this is already

marked my to-do's is complete okay let's

do this i'll

add it to do this oh my gosh oh my gosh

so it just went to start fixing the stuff

wait oh no oh no

let me reevaluate okay so i think it

basically ended up it wanted to fix stuff

but i don think I don think it did

anything It very ambitious at wanting to

do things This is really weird I think I

got into a funny state

Gemini 3 Flash High likes to write files.

It already did this for me.

Jesus. I can't trust you, Gemini 3 Flash,

if you're gonna be doing this to me.

I just asked you to review the code, you

were like, yo bro, I got you, don't worry.

Oh, by the way, I did everything else for

you, bro.

Nah, bro, you need to chill right now.

Gemini 3 Flash, I know you want to show

off.

I know this is your moment right now.

A lot of people are getting your shine

right now, but I need to pull back a

little bit.

Just a little bit.

A little too aggressive right now.

Ay-yi-yi.

Okay.

Yeah, so you see, this is what Gemini

did.

It just said, like, I'll update

everything.

And it's like, did you?

Is it?

I'm just going to discard your work.

Discard. You know what I'm going to do is I'm

going to do this audit with instead of

Gemini Flash,

Loki I think Codex Max with smaller

reasoning window. That'd be interesting.

So Codex Max,

and then I'm just going to do low. I'm

also going to actually this is probably a

good time to see

if I can hand this off to Droid and see

if it behaves a little bit differently.

I'm really,

really curious. Yeah. Okay. So I got

Factory pulled up. We're going to give it

that same task

I don't think I need a lot of reasoning

effort because it kind of overthinks to be

honest.

I'll do low.

Okay, so it's applying the new changes.

We have flash low.

Okay, I'm curious to see how Gemini 3

flash low is going to work.

The reason why I chose a lower reasoning

level is, so this is something that Eric

Provencher

covers in the Rate Limited podcast, and

he says that sometimes the models with a

high thinking can overthink.

I think I'm kind of seeing that in Cursor

right now because you see the models like,

oh I should then write the code now,

right? I did the review, let's go write

the code. Like,

and it was just kind of fighting with

itself. I'm just trying to remove that

noise from its thinking

and just kind of be more procedural

based. Yeah, so here's Droid and did Droid

already finish? Dang,

how did it finish so fast? That's kind of

what's weird. I have completed the

comprehensive code

review already. Dang, did you review the

code? Okay, so I'm going to do the same

thing but I'm

I'm also going to just try a different

model. I'm just really curious. This is

how I discovered the models are kind of

funny.

Models. Okay, so we're going to try 5.2.

I'm going to give it the same. We're going

to say low.

And I'm going to try this. Right? Just

trying to make it not over complicated. So

this is low.

We're going to do Droid again. So 5.2

high. Let's just see if there's any

difference here.

And then we're going to do Droid again.

We're just going to compare a lot of them.

and then we're going to do CodexMax GPT

5.2 low. We have GPT 5.2 high. And then

now we're going to do CodexMax high.

And then slash model, Droid. Just for

cost, because I know people are trying to

try different things for cost.

Between Gemini Pro and Gemini Flash what

would be better if we just do Flash on

high So Gemini 3 flashed on high because

this is what Cursor is using I curious to

see what they going to do So let go ahead

and see here

And here we go. So I confirm the project

such as bun. Okay. Yes, I conducted a

review.

Specifically, I verified these things.

Okay.

And I had Opus generate this prompt. So

this is code review handoff.

And we're giving this one to the GPT-5.2

low, which is doing its thing.

This is GPT-5.2 high. It's going to read

the files. You can see the different lists

here.

And then this one is GPT-5.1 CodexMax.

And this only generated a smaller to-do

list.

And then I think what we have to do now

is just see if Opus is going to do

anything.

roid, slash, model. And I'll do Opus.

Like, maximum intelligence. What do we get

for maximum intelligence?

2x the cost, right? Is it worth the 2x

cost, right? Are you going to spend like

the same amount of tokens eventually to

get to the same place? And just kind of

compress time? I'm just really curious.

Okay, so we have so far is codebase is in

excellent health and deployable. Ghost

features documented but not built. So none

found, all documented. And then security

assessments. Okay, looks cool.

Undocumented features. Goal celebration

sync, lobby timeouts, maintenance mode,

user XP fields, actual state sprite

integrations.

Better than documented, all five OpenAI

characters have full rotations and running

animations. The kick animations exist but

aren't triggered during gameplay.

-huh, that's what we gotta fix. Security

concerns and the codebase is deployable.

Yes, okay. This is cool.

I like Opus's review. Codex Max is still

going and it's just reading files right

now. So this is pretty good.

As far as low, low kind of did a similar

thing here. So if you're using GPT-5.2

low, this is pretty good.

This actually falls on par to what Eric

Bravonchet said, that GPT-5.2 low is very

similar to Opus 4.5 in some cases here.

So I might have to start reaching to

GPT-5.2 low for a lot more types of things

because of cost, but also thoroughness

too, right?

That's amazing. Wow. GPT 5.2 low? Okay, I

see you. I see you. If you give it a good

prompt, it cooks.

Oh, wow. Frames are loaded. I like the

details. These are really good details.

Okay.

This is pretty good. Compare that to

Opus. Opus is kind of, it's basically the

same. Opus is a little bit more succinct.

It covered more features here that I had

in my document.

But for 5.2 low? For the price? Gemini 3

flash low. I was hoping you would do some

more work here, but compared to ranking

all these right now, it makes sense.

Like GPT flash low probably didn meet the

benchmarks anywhere I mean I don see it

anywhere in a lot of benchmarks Gemini 3

flash high I see why it the default in

Cursor because it could get stuff done I

haven checked how it compares to Haiku to

be honest But as far as our winner here I

think it GPT and Opus 4 Then it goes in

terms

of matter of cost. GPT-5.2-Low is nearly

on par with 4.5 for this type of task,

where I have a

very procedural document from Opus that's

generated about what specific things to do

to go review these

different files. Opus 4.5 provides a

little bit more detail as far as the

nuances and little

All therefore, I leave to vàü Fighting

process on MVPF.

Please turn off subtitle and use CC by

error.

This is an interesting callout.

Active public testing, no live batch.

Okay.

And then Autostart is client host only.

This is a good catch.

So just by looking at this, you already

know it's live.

So this is where GPT-5.1 Codex Max

doesn't really get that.

It's like, your document says live and

I'm literally just taking it literally and

it says it should be live.

It needs to be live in here.

I don't see live, you know?

And like Opus 4.5 and all the Sonnet

models are like, chill bro.

Like, isn't it obvious? Like, there's a

countdown, there's a score, there's stuff

happening.

So, that's the difference between some of

these two models, and, you know, that's

GPT-5.1, CodexMax.

GPT-5.2 didn't really get that pedantic,

as we can kind of see, it didn't call this

out. But it's just interesting that I tried

this prompt, the same prompt, across all

these models.

And especially, like, the levels of

thinking too, and like, what it chooses to

hang on to and say,

and so on.

So, it's really hard for me to kind of

discern which one I should lean on.

I think running both of them could be

helpful, but if I were to just choose one,

I think

it's going to be Opus 4.5.

It costs more, but I do always keep

wanting to try these things to test them

out. And I think you get to learning from my

experiments and cost here.

So for procedural tasks like audits,

reviews, and checklists, the low reasoning

is actually

going to beat overthinking.

And I want to let you in on the quick

verdict here between all these different

models.

Gemini Flash.

and so on.

But it's actually too aggressive.

All I asked was for a code review and it

just started writing files.

And I can't be having that.

Opus 4.5 is actually the smartest.

It caught the little nuances, but it

costs quite a bit of money.

GPT-5.2 low though, that was nearly on

par with Opus 4.5 for this type of task at

a fraction

of the cost.

So next time you need a code review, just

try the cheap model first.

All you have to do is just write a good

prompt and give it low reasoning and see

what happens.

And look, this field is changing every

single week.

There's new models, new workflows, and

all that crazy stuff.

If you want to figure out this stuff

together, I'm doing live coding sessions

on the weekends

at Start My AI.

It's completely free and all you have to

do is just put your name on the list and

I'll let you know when I'm streaming.

It is your chance to code with me live

and ask questions while we work through

this stuff

in real time.

Drop in the comments which model you're

using for code reviews.

Alright, let's keep cooking.