So, huge news. Grok 4.6 just got

released as I'm recording this. But,

here's like the big headline. xAI caught

up with Anthropic. Now, the benchmarks

for Grok 4.6 High are showing that

they're equivalent or or close to Fable

5 Max. So, this is the first Grok

release that is at the time of release

is just neck to neck with the best

OpenAI model and the best Anthropic

model. 4.5 closed the distance and 4.6

but I'm right there neck to neck. Now,

of course, this was foretold, so to

speak. This was kind of predictable.

xAI's or Spec Space xAI's purchase of

Cursor probably had a lot to do with

this. In fact, as they say here, the

biggest difference between 4.5 and 4.6

was a much longer supplemental training

run. So, we believe that 4.5 and 4.6,

they are kind of the the V9 model or

base as it's referred to. It's a 1.5

trillion parameter base model and 4.6

had a much longer supplemental training

run versus 4.5. So, they used a lot of

the data from Cursor. And as they say

here, curated model generated data for

reasoning and advanced technical

concepts, high-quality engineering data,

and an improved optimizer and training

recipe. This produced a stronger

foundation for the supervised

fine-tuning and reinforcement learning

stages that followed. Now, of course,

most of us at this point view benchmark

numbers with some suspicion cuz we've

seen some places before kind of like

just trying to game the system, get high

scores on and those benchmarks, but the

models themselves kind of failed basic

tasks when people actually got their

hands on them. Is that the case here? I

don't think so. I've only played around

this model for a few hours at this point

and I got to say from my first initial

experiences just right out of the gate,

it's a beast. One of the first prompts

that I did was I asked to kind of do a

clone of Portal 2. So, kind of being

able to fire different portals, blue and

whatever, orange, right? You walk into

one, you walk out the other. Something

that a lot of the previous models up to

very recently just weren't able to do. I

used Grok Build to do this. It worked

probably for two to three hours or

thereabouts, and I don't want to say one

shot cuz I had a few suggestions and I I

steered it a little bit during the

process, including asking it to add some

sound effects and voice overs with 11

Labs, but pretty much yeah, it one shot

one sort of level from Portal 2. Created

all the mechanics, created all the

models, which you're going to see in

this demo in just a second. You know,

all the three objects it's created. The

conservation of momentum, the ability to

see through the portals, the ability

ability to see the character on the

other side of the portal. It's important

to understand that like this is not a

low-end model. This feels like a

high-end smart model. It created the

entire room in one shot, including the

puzzle that you're supposed to solve for

multiple elements. So, take a look. So,

this is chamber 07. Grok 4.6 build,

basically a replica of Portal 2.

>> Chamber 07, dual gate certification.

Please proceed.

>> And it's looking pretty good. We've got

reflections, we got lighting. Test

chamber 007. Blue gate.

>> Blue gate established.

>> Select the blue gate, and here is the

>> Both gates are live.

>> Let's see here. Oh, wow. All right. So,

we have a character model with

reflections. We're carrying that gun.

Uh that is pretty nifty. Okay, let's go

through here. Oh, wow. Okay, that's

pretty smooth, I got to say.

All right. So, if we wanted to go

through here and have my character here.

Wow, okay. So, I can see myself there.

So, what happens if I walk through it?

Oh, hey. That is

>> one, emerge from the other. Momentum is

conserved.

>> is conserved as we go through it.

>> Orange gate established.

>> Oh, look at that. Okay, so that thing

should

I did not expect that to happen.

I mean, it's supposed to happen. I just

didn't expect it to.

Okay, so this is our

>> Wait, you mad at Cuba quiet. Do not

introduce it to the acid.

>> Do not introduce it to the acid. The the

green bubbly stuff is the acid. Okay, so

basically I guess what we want to do is

slide this kind of like what is that

game where you slide the thing across

the ice? Loogie? Loogie? Something like

that. Okay, we go through here and as it

slides

>> Yes.

>> Hey, that is pretty good. I got to say.

I'm statistically a success apparently.

Well, thank you. So, there's a lot of

really incredible and amazing things

there. It's very impressive. Of course,

keep in mind this was just the first I'm

pretty sure this is the first prompt I I

threw its way. I was trying to see if it

works in Grok bot, which is the other

big release from yesterday. But, I

wasn't able to confirm that it was yet

available in Grok bot. So, I did Grok

build instead. That's where it built all

of this. And it was built with Grok 4.6

on the high settings. Later I realized

there's an extra high setting. So, I I

do want to test around it with that, you

know, turning up to 11. But, this was

Grok 4.6 high and this is very similar

to what I would expect Fable 5 to come

up with on on the first attempt. So,

again, Fable 5 level model from xAI.

But, wait, there's more because pricing

starts at $2 per million input tokens

and $6 per million output tokens.

Additionally, there's a fast variant,

which is twice the price. My next test

that I'm running within I'll post it

more about this later. So, I'm trying to

get it to build from the ground up fresh

without reusing any of the existing code

cuz I tried this before for a different

model. So, this is a starting fresh.

Here I'm using cursor by the way to to

do this. Important point on cursor and

Grok build because right now during

launch week, so that's right now, you

have double the usage for these new

models if you're using cursor or Grok

build. So, if you have those, do take

advantage. And so, here in cursor I used

a Grok 4.6 extra high fast. So, again,

fast is that model that it's twice the

speed, but it uses up twice the usage.

Right now, you get, you know, twice the

the usage, so it kind of evens out. But,

here I'm trying to get it to do a fully

AI automated run kind of streamer. So,

in this case, it's going to be playing

Pokémon Red and playing the game using

voice to narrate what it's doing.

Probably a HeyGen avatar kind of a video

avatar that's going to explain

everything while it's playing the game.

It's also going to be able to talk to

chat. And so, for that, I'm making sure

that it adds some sort of a censorship

so that people don't jailbreak it, which

could be bad. So, we'll see. We'll see

if this model survives contact with, you

know, you all people. So far, the

results are good. So, here's kind of

what we have so far. Just keep in mind,

this is a bare-bones prototype. If I

click start game, so notice it loads up

Pokémon Red. It goes quickly through the

initial screen, and I know it's got past

this right now. And so, right now, it

doesn't have all the logic and all that

stuff quite yet. Oh, the last time it

did get out of the house. Right now, it

seems to be stuck on something. But, the

point is, on its first attempt, it was

able to take a Pokémon Red emulator,

hook it up to make sure that it's able

to have enough of a harness to control

the game. It created a fake chat. It

created text-to-speech. So, that little

model that's over there, as it's

talking, that gets gets a real

voiceover. And so, all this was built

within the first hour or two of Grok 4.6

going live. So, again, I'm not claiming

anything about this model yet. I'm

saying just straight out of the gate,

this thing is feeling pretty beefy. So,

looking at these numbers, I'm not

doubting them. I think this is kind of

on the level that it's going to operate

at. This is Cognition posting about

Frontier Code 1.1. So, notice Grok 4.6

is right behind the Fable 5 and Opus 5

series of models. It's it's up there.

What caught my eyes, Elon responded to a

few of these messages saying that Grok

4.7 will exceed all current models. So,

the next release, 4.7, which we're

expecting in 3 to 4 weeks, will be

number one on the [clears throat] chart

assuming no one else releases new models

which is not going to happen. But as of

right now if everybody else was stuck in

time for three to four weeks, it would

be number one on the leader boards. And

for that model they're adding a massive

amount of SpaceX company data in

supplemental training. Okay, so so far

the only thing that's kind of like is

confirmed so far is what we know about

Grok 4.6. Some of the things that are

coming in the future, these are

speculations of certain announcements by

Elon, some of the kind of rumors that

are floating around. But we're expecting

4.7 to be a 2.1 trillion parameter

model. Again, that's not verified by

official sources as far as I know, but

here is the important thing to

understand. So, 4.5 wasn't that long

ago. Now we got hit with 4.6. We have

4.7 coming out in say three to four

weeks, and Grok 5 is targeted to be

released before the end of the year. So,

as you can imagine, that's a kind of a

crazy release schedule. Why is this

happening? xAI reorganized engineering

to ship major updates every two to three

weeks, it seems like. So, kind of keep

all this in mind, right? So, they caught

up, they got to Feat of Strength level,

and they are just gunning to release,

release, release. The pricing is $2 per

million input, $6 per million output, so

it's a half the price of comparable

OpenAI and Anthropic frontier models.

And we also have this SpaceX company

data that's going to be used in

supplemental training, but there's more,

and this is kind of the the other side

of the coin. This is the other big

thing. So, this is Grok Bot. And if I

understand correctly, 4.6 isn't running

this quite yet. I I did a few tests

here. I I don't think 4.6 went live on

this thing yet. Now, as you recall from

before, we talked about the different

benchmarks. One of them was GPT Val. So,

GPT Val takes real jobs, real projects

that's done by real people out there

that work in specific industries, and

these models are asked to complete those

projects. And then real humans sort of

like verify. So, these humans have a lot

of background working in those

industries, and the industries are from

engineering to like AutoCAD to some

hospitality and travel industry like a

travel agent. There's finance, video

editing, like it's a broad range. It

really represents a large part of the

job market. And the 4.6 model beats out,

you know, Fable 5 and everything else.

Why is that important? Because it goes

hand in hand with this with Grok Bot.

So, Grok Bot is xAI's always on, always

online, always available agent. So, it's

on desktop apps. I believe it's on

Linux, Mac OS, it's on Windows with

Android versions coming soon. And the

whole pitch is these are your AI

teammates. So, this is not like a

chatbot. These This is a team of agents

that work for you. One of the first

things that it kind of suggested you do

is you create, you know, let's say a

chief of staff. That's the role that it

kind of recommends. And so, I I jump in

here and I give it this this task.

Initially, I was want to see how far I

could take that AI streamer. So, I told

it, "Do this. Create whatever agents you

need." It does. So, everything here,

most of these agents except one, it

created automatically by itself. Then it

begins running the various tasks and

handing off the sort of the delegates

work to those sub-agents. So, you're not

talking to the million of AI agents

working in the background. You're

talking to your chief of staff or

whatever you want to call it. It

delegates, it thinks about it, it asks

you questions, but those questions are

might be the ones that are surfaced by

the agents. You're You have one point of

contact. Or you can have multiple points

of contact, but the point is that you're

not talking to every independent agent

on here. If you're wondering, "Well, how

could this be always on? Isn't this on

your desktop?" Well, no. Each of these

agents, Grok Bot, it has its own virtual

machine in the cloud that it can use 24

hours a day. You can actually see the

screen here. So, here it is. This is

what's available for it to use. So, you

can open up Chrome, terminal, whatever

you want. What's interesting about this

is there's actually a button to teach it

a task. This is kind of brilliant

because there's going to be a lot of

things that these AI agents have to

navigate that they're going to struggle

with. I had one of my AI agents not

through Grok, it was I think it was

OpenAI or Claude, I I forget which, but

I had it go online and go through an

online messaging app like WhatsApp or

Slack or what whatever it was to try to

see if somebody had mentioned something.

While I was doing that, it accidentally

sent a message. Now, this wasn't a big

deal, it caught the error immediately,

deleted it, and the message was just the

command it was run to to try to search

through all the messages by keyword. So,

it was something like search whatever

keyword, right? So, it wasn't a big

deal, but it could be bad in in a

different circumstance, right? Because I

didn't ask it to message people, but

with these teach a task, I can actually

show it, it gets recorded, and then

Grok, these bots, they're able to

execute those skills. Now, of course,

this is great if you're trying to work

with it, and it's kind of great for XAI

because it's basically a lot of people

kind of contributing data to these

models. And by the way, of course, you

have the choice to to opt out or not, so

that's entirely up to you, but for the

people that choose to opt in, of course,

their data can be used to further

improve these models. One other agent

that I built was my X news real-time

person that went through and basically

checks real-time news on X. So, X slash

Twitter, it goes out there and it starts

searching for me. It's super fast at

finding those news, so I just ask a

question, you know, like, "Did So, I

just ask it questions like, "Did anybody

talk about this? Did Elon say something

about this? Did somebody blah blah what

whatever you want?" It goes through, it

searches it, and it just explains it to

you. So, one of the things I asked it

is, "Grok 4.6, is it live on Grok bot?

Is this where we're using it?" And I

asked the other model, too, but it

wasn't able to say which model it's

running on, so I don't think as of that

time Grok 4.6 was live here yet as far

as I know. So, it found the official

posts, it found some comments by Elon,

and it said that so we're not 100% sure,

it's likely coming very very soon. Do

you want me to watch X for when they

turn it on?" I said, "Yes." It

automatically created routines and it's

going to keep checking, you know, three

times a day and for the next 10 days,

after which he'll notify me one way or

another. So, this is kind of awesome cuz

now I have this bot whose job it is to

keep searching X and when we get new

updates, it notifies me. Since this is

always-on, on 24/7, it's not like when I

shut down my computer, it goes to sleep

and forgets it. This will notify me for

sure. So, the big point here is that

these are meant to be more like your

colleagues and coworkers that you

delegate work to. So, this is kind of

digital labor, right? So, this has been

positioned by XAI as digital labor as

actual workers. So, a company that has

access to this basically is able to get

a lot of the cognitive labor handed off

to these agents that are online 24/7,

that are able to do research, you know,

when you're sleeping, put together

projects, hand them back to you. And

this is kind of I would say how they

design it is innovative. Now, not like

where they're breaking new frontiers or

whatever, but there's a lot of little

things here and there that just seem

like when you start using you're like,

"Oh, yeah, this feels right." One is

kind of they pop up with these boxes

more and more where you get to choose

your own answer. Your initial impression

might be like, "Oh, that's for, you

know, people that are just starting out.

Maybe they don't quite know exactly what

they're trying to do." But working with

this, I realized that there's a lot when

you try to go fast, this is very

helpful. Cuz when you give it a project,

it might have a bunch of questions for

you. Each time it asks you a question,

you have to kind of sit there and and

read it and think about it and then type

out an answer or just speak an answer.

There's been a number of times when it

shows me something like this and just at

a glance, I kind of know, "Oh, yeah, I

know exactly which one of those options

I want." Sometimes you don't even have

to really consciously read the question.

You get what it's asking you. So, at

some point it asked me this question,

right? And the answers were like, "Add

Heygen or just use text-to-speech." Like

I as soon as I glanced at it, I knew

exactly what it was asking without

needing to read the question. It was

going, "Do you want to build out the

video avatar or just leave that for

later?" Like I knew and I was like,

"Yeah, go ahead and build that out."

Things like this might seem tiny, but it

makes things go a lot faster. It's less

of a cognitive load, right? You're not

wasting cycles having to read the I mean

in your own brain, so to speak, reading

through everything. You're kind of like

cuz you know what you want. The model

doesn't. And so, any little polish we

can add to make that transfer of

knowledge faster and smoother and takes

less effort, it's a big deal, especially

as more and more work begins to get done

like this. So, love this so far. The X

News AI agent that adds the routine and

then just says, "Yep, I'll go ahead and

I'll start checking it three times a day

for the next 10 days." I didn't ask it

for that. It says here, "Do you want me

to watch X for when they flip it?"

Right? When it goes on. Yes. It's like,

"All right, got it." And it adds the

routine. This is beautiful. This is one

of the things where I like this so much

better than Fable 5, as much as I love

Fable 5, but the way that it talks half

of the time, half the answers, I'm like,

"What are you saying?" I mean, at this

point there's like entire memes by it.

It's got this weird way of speaking, you

know, and yeah, it's kind of funny, but

after a while it kind of gets old. It's

just like just tell me what I need to

know. What are we talking about? So far,

going through these and working with

Grok, it's just so much faster. You just

kind of fly through it. So, definitely

encourage everybody to download this.

It's super simple to set up. When you do

so, create your kind of chief of staff

or whatever name you want to use for it,

then pin it to the top, and then try to

use this as much as possible, and have

it create its own structure underneath

it so that you're sort of talking to

just this thing. And then and maybe

there's other sort of side agents like

this X News one that you have off to the

side that that it's its own thing. I

don't need it going through the chief of

staff, but for a lot of this work, have

one agent in charge of it. And of

course, this is easily cross-device,

right? So, I can go from here to using

it on my phone to wherever. I'm not

relying on the computer being on at any

given moment cuz it has its own

computer. So, between Grok bot and Grok

4.6 coming out and the new models that

are coming out, 4.7 and Grok 5, I mean,

those are not out yet, but so far it

looks like XAI is just in a great great

position to start winning. Check it out

within Grok bot if you have access to

Cursor. You get some double the usage

during launch week or if you're using it

with Grok build, same thing. But, I got

to say this is

this is kind of exciting. It seems like

Google is lagging slightly behind, but

we still have, you know, multiple

frontier models because here comes xAI

to fill in Google's spot. So, congrats

to xAI, congrats to Cursor. I was always

very impressed with Cursor and all the

stuff that they've been able to do. So,

them joining forces with xAI, which has

just incredible amounts of compute and

and talent and capital in order to

really put these two things together. I

feel like they're going to have a lot of

success in the future. But, let me know

what you think. If you've had a chance

to do more testing of Grok 4.6 more than

I have so far, let me know. What do you

think of it? Is it, in your opinion, as

good as Fable 5 and GPT 5.6? Will you be

using Grok bot? Let me know in the

comments. If you made this far, thank

you so much for watching. My name is Wes

Roth. I will see you in the next one.