The AI Models, Agents, and Subs I'm Currently Using

I get a lot of questions about what

model am I using, what harness am I

using, whether or not the new $200 sub

from whatever provider is actually worth

it. And I wanted to make a video going

over my current opinions as of March

30th, 2026 on all of this stuff. I'm

including the date in this intro because

unfortunately, I don't know how well

this video is going to age, but I think

I want to try and make this a recurring

video entirely for the fact that things

are changing so damn fast. But it's how

I'm filming now and I want to go over

it. I have this thing to point to of

these are the models, harnesses, and

subs that I am currently using and I

would recommend to any of my friends or

anyone I'm working with. I put together

this site called State of AI. It is just

AI.davis7.sh.

It'll be linked down below in the

description. And all it is is a snapshot

of how I felt at this moment in time.

Starting with the models, my main

default model is GBT 5.4. If you look at

my token consumption over the course of

a week, I would guesstimate that 80 to

90% of it is going through GPT 5.4. It's

incredible at instruction following. It

has very up-to-date and recent training

data. That was a problem that OpenAI

models had for a long time that it just

doesn't really have anymore. It's

incredible at like computer use type

stuff. It's incredible at running within

a normal agent harness type thing. And

it's able to solve problems that I just

can't get any other model to solve. GBT

5.4 for basically everything is the king

right now. There is nothing better. I

really didn't like GBT models last year.

I was a huge claw guy. I loved how they

felt to use. But after about a week of

using 5.3 Codeex a couple months ago, I

finally got it and it clicked. But once

you get there, it's an incredible model.

There is one thing that it absolutely

sucks at, and that's UI, which is why

Opus 46 is on this list. I pull out Opus

46 when I need to touch front-end code.

The UIs that GBT 5.4 spits out are

genuinely horrific. I don't know what

happened, but it is a massive regression

in UI capability compared to even GPT5.

it's just bad at making anything

involving front end. Opus isn't. You

give it the front-end design scale and

it will just do the job. It is

incredible. It actually feels really

fast to use, even faster than GBT 5.4.

And I think a lot of the reason for that

is because it's very prone to action.

Something you'll notice with Claude

models is that they feel great to talk

to. The voice of Claude is probably the

best uh at least of all the Frontier

Labs. Like I said, it biases towards

action. So, it feels faster and it feels

better in a lot of ways than GBT does

because it'll just do things, but it

lacks the discernment and wisdom that

5.4 4 has where when I was working on

the original version of BTCA, which was

a massive TypeScript monoro, it would

make very subtle bad changes that would

end up having cascading effects in

completely different packages across the

monoro in weird ways that you wouldn't

realize just looking at the code it

generated cuz the code looked right, but

it wasn't right in the broader context.

That's the thing that 5.4 has over opus

and the reason why it's my daily driver

and not opus. It's that discernment. And

also another big problem with it is it's

really [ __ ] expensive. Like really

really expensive. Another good anecdote

from chat, which I do record these

videos live. You should come by

sometime. But in his case, he was trying

to do some cyber security stuff and ask

it to sniff some stuff on his Wi-Fi. GBT

just flat out refuses, but Claude will

do it. I think that this is again just

kind of the way GBT is. It is very

strict about security stuff and is very

careful to not make breaking changes and

not to do anything dangerous, quote

unquote. And in this case, I think

that's exactly what that is. So yeah, if

you need to do something like that, claw

does make a lot of sense. I don't hate

this model by any stretch of the

imagination. It's just not my main

coding model. 5.4 Mini to me is a

massive sleeper pick. It feels a lot

like a sonnet type model even though it

says mini on it. It's way smarter than

you would think it is. It is really,

really [ __ ] fast and it's great at

tool calling. It keeps a lot of those

characteristics that GBT 5.4 has where

it's good at following instructions.

It's good at working in a loop. It's

good at high level discernment and not

making dumb mistakes. But of course, the

trade-off to all that is that it is not

nearly as smart as GBT 5.4. But if you

need to do sub agency type tasks like

searching a codebase or searching the

web or making some small change on your

computer, 5.4 Mini is actually a

surprisingly useful model that I would

recommend picking up more than you

probably are. The Gemini models I really

don't have too much to say about other

than they do have a place. I was

benchmarking different models for taking

a blob of something, so in this case it

was emails, and converting them into

JSON that parses out what they're for,

how important they are, all that kind of

thing. And I put together a benchmark to

see which one was the best at that

specific task. And Gemini won pretty

handily. It's a great multimodal model.

It's just not a great coding model

because it cannot run in a coding agent

harness to save its life. The only place

where I would say the Gemini models are

usable for coding is cursor. And that's

just because they spent a ton of time

and effort refining the system prompt

and tool names for Gemini models so that

they wouldn't make mistakes. The last

model I put on this list is Composer

2.0. It is actually a RLED version of

Kimmy. 25% of the compute went into

making Kimmy 2.5 which is the base

model. Then 75% of the compute for

Composer like actually training it went

into just refining it through cursors

data. And it's a really good model. It's

absurdly fast, way better at front end

than you would expect. And anytime I'm

doing pair programming type stuff where

I'm reading the code, I'm in the editor,

I'm thinking specifically about it, and

I just want to have it make some

specific change in a file, composer is a

great model for this. The problem is you

can't use it outside of the cursor

harnesses. They have a lot of really

good harnesses, but still you're limited

to where it doesn't work in something

like a PI or Open Code or even over API.

And it's not nearly as smart as an

OpenAI or anthropic model, but it's

still a good model. And you'll probably

notice that I don't have any of the

Chinese models on this list. And the

reason for that is because I'm only

covering the stuff that I'm using

heavily day-to-day, and I have real

experience with. I've tested a lot of

these other models. I think they're

cool, but I don't have enough experience

to say anything intelligent about them

other than paring other people's

opinions, and I just don't think that's

useful. So, that's why they're not in

here. But, there is something else

that's in here, and that's today's

sponsor. There's a really weird split

that I've been seeing where on our

machines we have these insanely powerful

agents like codecs and cloud code, but

then in the cloud we're still just

running like normal Verscell AISDK

stream text loops. A lot of the reason

for that is because codecs and cloud

code benefit a ton from having a real

computer to work with, which is why

sandboxes are so damn valuable right

now. And that's why I'm really hyped

about Upstash's new box product. And you

might be thinking, wait, isn't Upstash

the company that just has that really

nice Reddus instance? And yes, they are.