I have something I'm a bit ashamed to

admit. I'm a pretty big fan of Meta. Not

of things like Facebook, like no, gross.

Not my thing at all. But when it comes

to their actual software contributions

and things they do for the ecosystem, I

found them to be pretty solid, if not

great. Projects like React and React

Native were essential to the growth of

the web and to an extent mobile as well,

and are genuinely incredible things that

they've put out there for free for

people to use however they want. Their

contributions to the AR and VR ecosystem

were all things that I enjoyed heavily

at the time because I'm a big VR nerd.

And when they started getting into AI,

they did that similarly well with their

focus on openweight models with the

llama line. To this day, llama is still

used almost like a generic term for

openweight models. And it's crazy how

far the llama models went, but it's even

crazier how far the entire industry has

moved ahead of meta. They've been trying

to correct course for a while with this

hidden quietly worked on line of models

called Muse. And historically, they've

not been accessible beyond their little

web interface to try it out. But

apparently, they've been trying to make

it ready for code. Not just the model,

but the tools around it as well. And now

they're finally ready to release.

Zuckerberg just announced on Twitter, by

the way, that Muse Code is out in beta

today. This is their clone of Claude

Code and it's powered by Muse Spark 1.2,

which is their new coding focused model.

It's been a while since I've seen a lab

publish benchmark numbers where they're

not in first place in anything, but that

doesn't mean it's a bad model. As you

guys probably saw in my video about Gro

4.5, there is still a lot of value in

wellpriced, fast, useful models. And

with Meta's focus on using every model

for everything lately and collecting all

the data they get as a result, it's

actually going to be pretty interesting

to see how well this performs. So, I'm

going to go through this with y'all

together. I'm going to try it live. I'm

going to read through what they have to

say. I'm going to see what others are

doing with it, get a feel for the price

and the weird quirks as well. And I just

had to sign into my terminal with

Facebook. So, I'm feeling a little

gross. This is going to be a journey and

I hope you enjoy it with me. But first,

a quick break for today's sponsor.

There's a lot of code review bots around

nowadays and they're pretty good at

reading the code, but I've noticed that

running cloud code and codecs on my

machine tends to be better simply

because the models can actually run

things and verify the changes. Wouldn't

it be great if those review bots had the

ability to actually verify changes on a

real computer? Reptile thought so, and

that's why they introduced T-Rex. It's a

new sandbox that can actually run your

code and verify the changes rather than

just reading the syntax and hoping it's

good. If you've ever added an element to

your UI and had it overlap something

that was obvious when you ran it that

the code review bots missed, that's what

this is for. Because the models can

actually click through and test things,

and it can even respond with screenshots

of the things it finds along the way.

Their website lists a bunch of realworld

open source projects that have had bugs

prevented through T-Rex. Even companies

like Work OS are finding real bugs. For

example, this code that seems like a

totally safe filter, but has a really

rough edge case that you'll only find

when you run it on real data. If you've

debug in the last 3 months, you owe it

to your team to check out

soyv.link/gretile.

So, let's start with what Zuckerberg had

to say about their most recent release.

Releasing Muse Code in beta today. It's

a terminal coding agent that takes on

complex software engineering tasks

across large repos, lining changes,

writing code, validating the results.

It's powered by Muse Spark 1.2, which is

a coding focused model update. This is

going to be an interesting one because a

lot of the other legacy companies like

uh I don't know Microsoft and Google

don't necessarily have the best stack

internally, whereas Facebook has done

some incredible things with their

internal tech. They didn't use git

because it was too slow for the scale

they were moving at. So they built their

own custom everything on top of

Mercurial and it's really powerful to

the point where people have been trying

to copy their workflows with things like

stacked PRs and stacked diffs. I know a

lot of people who used to work at Meta

and left and just missed all of that

tooling that made it so much easier to

work in these gigantic projects that are

met scale. Other companies like Google

also have gigantic monor repos, but they

are less focused on fixing them. In

fact, they've went as far as laying off

the teams that built the tooling to make

them usable. Meta's always done the

opposite where they'll rewrite a

language if they have to, like they

rewrote PHP and hack in order to try and

make their code bases work better. Since

Meta's code bases are so huge, it's

often easier to rewrite the language

that's underneath them than it is to

rewrite the codebase itself on a better

language. And that type of thinking is,

at least it was unique to Meta for a

long time. And I think it kind of

positions them well to dive into this

world of using models and tools to work

with your code bases at that scale. And

that's why they've been using models

like Fable and Opus so heavily recently.

So, it's interesting to see what they're

cooking here. They shared some benches

here. And the first notable thing is

that they're not putting Fable or Soul

in these lists. They're only putting

Opus and 56 Terra in, which is a choice.

In Terminal Bench 2.1, they slightly

beat out Terra and are slightly behind

Opus 5. There also a big jump from US

Spark 1.1 to 1.2. 1.1 was barely

available to the public. 1.2 2 is now

actually out and the pricing is really

interesting which we'll show in a little

bit. Deep SWE which is one of the

benchmarks I prefer. It did pretty well

here getting a slightly higher score

than Grock 4.5 which is nuts because I

actually found Grock 45 to be a very

pleasant to use model but it's still

lagging behind Terra and Opus which

means it's far behind models like Soul

and Fable. And then Meta's internal

coding bench, it came in second place in

with Muse Spark being much closer in

third. And if this model is as far off

as other benches showed, I think their

internal bench might not be the best.

They also in their internal bench have

Terra in Gemini 36 Flash suspiciously

close. So, we shall see as we go. Oh

god, they're already getting community

noted. Community note: Meta is still

dead last in AI with this release. Okay,

maybe not. Apple is, but they're behind

just as much as Google Gemini is. That's

not a thing you need to put in a

community note, nerds. Anyways,

according to Zuck, Muse Code runs

specialized background agents that stay

active your whole session. So, they

build up context over time instead of

starting from scratch on every task.

Interesting. When a job is big enough,

it fans out the separate sub agents

working in parallel in isolated work

trees. Your working copies never

touched. In testing, we had it build six

features for a game simultaneously with

no collisions. Huh, very interesting.

Again, big codebase stuff. We pointed

Muse Spark 1.2 at a kernel optimization

task and let it run. a thousand tool

calls over 24 hours on Nvidia Hopper. It

kept finding substantial improvements

well beyond the initial exploration

phase. This is also kind of funny to me

because Meta rented a bunch of TPUs from

Google. They're one of the few companies

to actually buy and rack hardware from

Google and then Google went and rented a

bunch of GPUs from Nvidia, all from XAI

and SpaceX. So Google is reselling their

shitty chips to Meta so they can use the

money to go buy better chips from Nvidia

through XAI. And now apparently

Zuckerberg is still stuck on Nvidia.

Auditable by design, every model call,

tool run, and edit has a local event log

before it executes. If it crashes mid

task, it picks up exactly where it left

off from that log. No lost work and no

reprompting. Pricing. It's easy and low

cost to get started. Install Muse Code

with one line and you can start on our

contributor tier. Contributor tier is an

interesting piece we'll talk about in a

second. Musepark 2 is our next step as

we push towards Front Tier with larger,

more capable models on the way. Install

it, use it, and tell us what you think.

Well, the first thing I think is that

when you Google search Muse Spark, you

get all of these old articles about 1.1.

This old developer page. Okay, this one

has 1.2 on it now, but it was actually

kind of hard and annoying to find this

initially. Here's the part I was looking

for, though. The price. $1.25 per mill

in, 15 cents for cashed 1 mil in, and

425 per mill out. Not great pricing

until you realize I was hiding something

from you. The contributor price tier.

This tier is 10 cents per mill in and 20

cents per mill out with 0.2 cents per

cached input token. That is a 10 to 20x

price gap. They're effectively giving

out the contributor version for free

because they need training data so

goddamn badly. Meta's released Muse

Spark 1.2. It's their third release in

four months and it scores a 54 on the

intelligence index, significantly

improving agentic knowledge war

capabilities over prior releases,

putting meta next to SpaceX in a tie for

third place amongst US labs. Crazy they

have to specify this now because the

Chinese labs have caught up so much,

especially with K3. Muse Spark 1.2 lands

at a 54, up three points from 1.1 and 11

points from Muse Spark 1.0 which came

out in April. That's a pretty big jump.

Like if we're just looking at the

trajectory of what meta right here,

that's a lot of improvement in not a

whole lot of time. It's effectively tied

with GBT 55 and Gro 4.5 narrowly behind

current frontier models like Opus 5,

Fable 5, and 56 Soul as well as Kimmy

K3. It's among the most costefficient

models at its intelligence level. It's

40 cents per intelligence index task at

Meta's unchanged $125 and $425 per

million token pricing. They don't

mention I I was right when I was reading

that earlier. They are not measuring the

prices based on the contributor tier.

They are measuring them based on the

normal pricing that they would charge

actually. So if we go back to the

pricing chart here, it's about a 10 to

20x difference in price, which means

Muspark 1.2 on the contributor tier is

actually the cheapest model currently

here because it would be 2 to 3 cents

per task, which is comparable to V4

flash as well as 56 Luna. Very

interesting. This model, especially if

you're down to like let Meta have your

data, could be a really good value for

the short term. Amnition's absention

rates have increased. The model score

rose from an 18 to a 22 and the

hallucination rate fell 10 points. The

attempt rates dropped from 82 to 67.

This is this bench is measuring how

likely is the model to lie when it

doesn't know something and it's much

less likely now even though for my

experience that is not the case.

Scientific reasoning results remain

largely unchanged. does much better at

GDP val, but I still don't love that

bench. Good analysis. We can work with

that. Muse is trained heavily on data

from people using models like anthropic

models internally, and all of the work

they do. I've heard numbers as high as

50% of staff at Meta have been moved

over to some type of data labeling or

like management tasks in order to help

them train smarter models. And one of

the results here is this wonderful tweet

from Luke. Muse Spark is trained on the

screen recordings of Meta employees,

which makes it best-in-class at applying

for jobs at Anthropic. Absolute banger.

I want to go through other numbers that

matter before showing off the code that

I can write with Muse. I already have it

working in the background and it's

surprisingly fast. We'll talk about that

in a sec, but first I want to look at

the numbers from artificial analysis. As

you can see here, Muse Spark 1.2 falls

right between Kimmy K3 and Gro 4.5. I

will say it's a little bit embarrassing

to release a closedweight model as a lab

that has billions of dollars like Meta

does and have it be behind an openweight

model that came out much before it. That

is a little embarrassing. But also, if

it's really fast and cheap and reliable,

these numbers don't mean everything cuz

I can tell you confidently that Opus 5,

despite being at the front of this list

and tricking me into thinking it was a

great model, ended up all being Copus

because I hate using Opus 5. Now, it

once I started merging the code it

wrote, I realized how bad it was cuz I

had to have Fable and 56 Soul come in

and clean up the mess. Not great. But

again, Kimmy K3 actually felt quite

usable, as did Grock 4.5. But Opus at

the front doesn't feel usable for real

code work at all. But what I really want

to see is output tokens. How did they do

here? Interesting. Okay, so Muse Sparks

runs did about 30k tokens per task

compared to about 36k for fable 5 and

compared to about 17k for 56 soul. So

it's not quite as inefficient as

anthropic models, but it is far from the

efficiency that we see from models like

the GPT line or Gro 4.5. Very

interesting. Definitely puts it in a

weird different spot. Like this model

isn't just another openweight model

tuned or something like this really is a

new model and it should feel quite

different especially when you look at

the speed that this model has too. I

don't think they have recent speed

numbers yet on artificial analysis but

they do have them on open router. Open

router is an easy way to use pretty much

every model and they track the

throughput and speeds that users are

actually seeing when they use the models

through open router and they're seeing

an average of 191 tokens per second

which is absolutely nuts. For reference

56 soul is only at about 30 tokens per

second. So that is a gigantic

improvement like 5x plus absolute best

case with soul is 134 TPS but

realistically speaking the p50s in the

40s to 50s on openai's official infra

and the 50s for Azure at best. To

compare Muse Spark gets 316 TPS at its

best and 162 at its P50. That is pretty

nuts and I feel the result coding with

it already. Man, that cerebrous

deployment of 56 soul can't come fast

enough. It was supposed to already be

out, so I'm not sure what happened

there. No one's given me any info,

sadly. Then we get to cost and Muse

Spark is performing absurdly well here

at 40 cents per task on average through

the artificial analysis intelligence

index. That's still more expensive than

V4 flash is, as well as 56 Luna on max,

but it's a lot cheaper than other

things. Even Mistro medium is more

expensive. And Kimmy K3 is more than

twice the cost for similar work. And I'm

assuming this is the normal cost too,

not that contributor tier, which is

effectively free. Think it's about time

we fire off some prompts. I just opened

Muse Code inside of the T3 codebase.

It's defaulting to one two contributor

on high. And since T3 code is fully open

source, I don't mind if they get my

data. They probably have it anyways. I

asked it to give me an overview of the

architecture of the codebase. And it had

this under 30 seconds. Not bad. Let's

see where this goes. If I ask it for

something a little harder where it has

to investigate more. I asked it if it

can find anything suspicious about the

event sourcing model. Things cause weird

behaviors on different platforms,

unnecessary bandwidth, usage and lag,

etc. Let's see what it does for that.

Another interesting thing I noticed when

I set this up is that it pulled all of

my skills and personal rules from claude

code. Normally things don't use the

cloud directory because it's clouds.

They use the shared aagents standard and

agentmd stuff like that. They seem to be

really focused on quad. And from

everything I know about how meta's been

working lately, they're definitely an

enthropic house. They use a lot of opus

and a lot of fable there. Yeah. 174 TPS.

Okay. Okay.

Now we're talking. For comparison, Grock

4.5, which is still quite fast, is at 50

to 52 TPS. And a model like 56 Soul is

30 to 50. It's crazy that Azure is

performing that much better than Open

Eyes official servers are right now. But

yeah, you're welcome. I had to fight

hard to get Azure to move. It had some

interesting findings here. I'm going to

tell it to summarize them. See our

findings in an easy to digest HTML file.

Let's see if it's able to find my HTML

plan skill and actually take all of this

work and synthesize it. Look at that.

Already found the skill. Now it is

thinking. Hopefully, it will synthesize

this properly and then put things up.

It's interesting to state how much data

this skill is. There's definitely some

slop in here. things like the way that

the full path for files is included in

the UI and stuff like this CLI is a bit

slopp here is the HTML it made for me to

summarize the findings that it had pure

decider plus single serialized worker is

correct cost surfaces at the wire full

snapshots on subscribe unbounded

payloads in a global sequence that

couples every thread so here's what

we'll do for a comparison I'm going to

do a cla run with fable we're going to

paste this I have a plan here you know

what I I want to try DeepSeek 4 model.

Deep seek v4 flash. Let's copy over the

same prompt then.

God, that is fast as too. For

reference, Muse 1.2 got me this feedback

and even had an HTML page generated all

in well under a minute. I've just been

trying to get similar feedback from

Fable and we're over four minutes in and

still don't have any useful anything

here. I also have Deepseek Flash running

in the background as well just as a

comparison because the new Flash release

was really solid too just to get a

comparison across things here. RG wasn't

finding them because I used - G quote

star.ts which doesn't work on Mac OS

with rip grip. This is the Deepseek one

running by the way. God, these are both

so slow and not getting me useful

feedback so far in comparison. It's been

very interesting. My plan is I'm going

to take the findings that Muse had and

throw them at these other two models

after and say, "How do these compare?"

So, my favorite ways to get a feel for

what model strengths and weaknesses are

is to have them all do the same task,

have them write up the results, and then

have them compare each other's results.

Deepseek Flash just finished its

analysis. I'll ask it write up your

findings in HTML. See if it can figure

out my HTML skill. Oh, um, use my HTML

skill. You can read it from my agents

orclaude.

It doesn't have skills built in by

default if I recall. Oh, it did. It

found it. Nice. After waiting way too

long for both Fable and Deepseek to do a

similar investigation, I asked them both

to review their findings and compare it

to what was found with Muse 1.2. And

here we see the way that Fable

categorizes the differences. And

remember, Fable, when doing a similar

task with Opus, thought that Opus had a

much better investigation overall. It

seems like that is not the case with the

Muse investigations. There are a few

places where it said that Muse did

better. For example, the coverage breath

it thinks Muse went a little further on,

but for the most part, it thinks it did

worse and that Fable's plan was much

more accurate and overall properly found

root causes. Meanwhile, Deepseek seems

to think that the Muse plan was

meaningfully better. Very interesting.

Yeah, it's a solid model for the price,

but I don't know if I would trust it for

really heavy endto-end work yet. But

we'll test it a little more. I had to

try one of my favorite tasks where I

take my crappy fish game and have it

remake it both in 2D and in 3D. And the

most impressive thing is how quickly it

did it. It made the 2D version in about

2 and 1/2 minutes and the 3D version in

under five. For reference, Opus 5 took

over an hour for both of these. It

doesn't move properly. Like I can't turn

Kind

of nuts that it made this whole thing

from scratch in four minutes, though.

Like, that part is impressive.

It's just that the rest is broken. Mouse

look does not work. And it made a 2D

version of Fishlop as well in 2 minutes

and 39 seconds. For reference, when I

put these same prompts to remake this

game into other models like Opus, it

took multiple hours to do. And this made

it in under five minutes for both the 2D

and 3D versions.

Is it perfect? No, far from it. But it

does have a couple nicities. Like it it

made solid animations using the sprites

that I had access to from the other

project. It did shadows and like

contrast well. Like it it has a weird

type of taste that I haven't seen models

have. It's obviously still jank as

in a lot of ways, but like the little

shadows are nice. The way things are

moving is nice. It feels fluid. It's

this model feels different. It doesn't

feel like it's just distilled on other

popular things. It does have a vibe to

the results. And I am definitely curious

what's going to happen when they make

bigger and smarter versions. It

allegedly just fixed the 3D mouse broken

stuff in a few seconds.

And it does appear that it has. I can

now look around as expected.

The fish are all swimming backwards

though. That's hilarious. Fish are

swimming backwards. I can't look up or

down. Only left and right. Can you use

computer use to test it yourself? Oh, if

I might codeex computer use skill. Oh

no, that is not the solution. As if it

can do it itself without codecs. Does it

have vision? Oh it does. That's

huge. Okay, now the up and down movement

of the mouse tilts this. That is broken

as And the fish are still moving

backwards. It promised me it fixed that.

Oh, wait. No, it didn't. Yes, please fix

and rebuild. Okay. Nope. Fish moved the

right way now, but I still can't look up

and down.

spaces both feed and go up. Yeah,

there's a lot of little things it got

wrong here. Like it

it doesn't understand

that things can collide with each other,

which is interesting. So, one of the

things they said the model was good at

was like breaking things up and letting

lots of stuff work at the same time.

But, it's it's getting work done and

it's doing it alarmingly fast. Like,

that's one of the nicest things here is

using it to like touch something up.

seems to be a useful method. Dra just

got me one of my favorite things to look

at when new models drop. The comparison

of how it handles design on his

witchai.dev site. Very nice for like

looking at the designs that different

models do and comparing them. And I'm

already seeing some interesting things

here. Like I love how it uses the scroll

areas where this like graphic on the

side scrolls with you until you hit a

certain point.

The sections are tasteful, too. I I

don't hate this. I'm tired of the pills

on everything. It feels a little

templaty, but not bad at all. And this

little tilted reminder recall thing here

is nice, too. This is all without the

design skill as well. If we switch to

the second one, we end up with one of

the cringy code style ones. It's laid

out well. Like, its page layouts are

solid so far, but I don't love the

design and the graph it made there. It

really likes doing these tilted things.

It has one here as well. I actually I

really like this one of them so far.

This design looks decent. And I like the

way it's using the like squirle here. I

don't like that it switches from squirle

to rounded there. This isn't bad though.

This again is reminiscent of Gemini to

an extent. Way too much text. Way too

much text, but otherwise not bad. These

brutalist ones are getting tiring. I say

after making lawn video, which is

absolutely of this style. But yeah, it's

not bad. Like nothing jumps out as like

that is horrible. I don't love having

the bar here when it's already a very

brutalist like lineheavy style and now

the classic like Tailwind template

homepage.

Oh, it does these little floaties. Well,

I I like how it uses those other models

get that stuff really wrong and I fought

it a ton on things like the T3 code

marketing site. That's like what's

interesting about this model just has a

bit of a different flavor. You know

what? Let's give it something harder.

Does it have work tree support first

off? It does. Huge. That means I could

play around with less risk a bit here.

Can I yolo mode from here? Nice. Oh,

that's a bit annoying cuz I had that in

a work tree. D-workree- yolo. Now we're

in a safe work tree where I can ask it

to do stupid things like let's try their

voice. Actually, I would like to

implement muse as a provider inside of

T3 code. I'm not sure what Muse code

supports in terms of integration

methods. I haven't played with it enough

or looked into the SDK. I don't even

know if it's open source. What I really

want is to integrate it through a layer

like ACP similar to how we have

integrated other providers in the past.

But if we need something more custom, I

am down. I would like you to start by

investigating everything you need here.

Both how we can implement within T3

code, but also what is offered by Muse

itself. You'll have to do some digging

to find source code where available,

docs, SDKs, and whatnot. Might be a bit

hard to find because this is all still

so new. Use a lot of sub aents to break

up the work as you go. I actually liked

that voice to text. It showed things as

I spoke, which a lot of other solutions

don't. They just show it when you're

done. So, let's send that over and see

how it goes. That's the first good

terminal voice to text I've seen in any

of these CLIs for being real. Look at

that. is already spawning sub aents and

we can look at them almost the exact

same way we can in claude. I I know I

click baited a bit saying this is a cla

code clone, but like this is such a

cloud code clone. It's cool that it spun

up all of these intelligently. Like it's

it's clear they are trying to make this

model good at this type of like broken

up heavy sub aent breakup work. I just

got hit with a rate limit.

Are you kidding? I put in a credit card

and everything. Are these sub agents

getting hit with them, too? They are.

Great. Why why would you make a sub aent

flow like this if your APIs can't even

handle it meta? Uh how much does it cost

so far is a good question. Interesting.

I actually checked out the dashboard

before. They have instructions on how to

set it up with other providers too like

open code, cloud code, codeex as well as

curl and python directly. Assuming the

Python is showing you how to like set it

up with code. I'm assuming this is like

an OpenAI compatible API. I want to see

how much I have spent. I don't know how

accurate or up to date this is, but

everything I've done so far apparently

is only about 13 cents, including both

of those game rewrites as well as the

codebase audit. So, yeah, absurdly

cheap. It's making progress here. Why is

it talking about anti-gravity?

I just asked why is so much of this

about anti-gravity? I didn't mention

anti-gravity at all. It did a deep dive

and it took under three minutes despite

the fact that it was hitting rate

limits, but like the results are

useless. Contributor has very low rate

limits. Okay, that's the problem. It's

cuz I'm on that contributor tier. Okay,

I shouldn't be so amused by this, but

this is actually objectively hilarious.

I

resplap

heavily in recent search results and I

ran the whole investigation on that

assumption without confirming. I don't

have a clear hit for Muse as a distinct

product. Web searches mostly return Muse

Spark, which is Meta's new coding model.

July 2026. An unreleased Muse Markdown

OS project. Something definitive for a

Muse CLI or SDK. Can you point to me to

what you mean by Muse? A URL, the

company behind it, or what the binary or

SDK is called would be enough. I'll give

you a hint. I'm in it right now. Let's

see if this can figure out what it is.

I'm way too amused. Yes, Muse is Muse

Code itself. The CLI you're in right

now. Powered by Muse Code 0.1.0. Powered

by Muse Spark. My first investigation

mapped it to anti-gravity by mistake.

Completely wrong target. So of course

the Muse code layer here does not have a

lot of the functionality that I would

want to integrate with it. Things like

it being open source. Things like it

having an SDK I can plug into ACP as a

way to communicate with it. It's missing

a lot of the stuff I need. I do have to

write up a plan on how we would

integrate Muse code. You know what? I'm

going to send the smarter model through

this with a smarter tool code. code.

I'll do it on this machine so that I

have the results here. I want to

integrate the new Muse code

from meta into T3 code as a new

provider. Not much documentation exists,

but I do have it installed right now.

Investigate and help me figure out how

we can integrate. So I have Fable both

investigating how it would do it and

separately in parallel I have it

reviewing the plan that was written by

Muse. O the page layout's super broken.

Bit annoying to make those types of

mistakes still. There's a lot of things

here that feel really last generation,

but a lot of things that feel next

generation, too, where it's like it

seems to know how to break things up

with sub agents really well, but it also

seems to struggle meaningfully with like

basic page layout stuff or staying on

task and not hallucinating its way down

an entirely incorrect path. Do we get uh

omniscience scores here yet? Um, oh,

apparently it does well on a

omniscience. That is weird to me cuz we

just watched it hallucinate

aggressively. Allegedly, it is about as

bad of hallucinations as 56 Soul is and

slightly better than Kimmy K3, but

considering how aggressively we just

watched it do that, that was rough. I

had Fable give feedback on Muse's plan

to integrate Muse and it said it's a

solid plan and gave a little bit of

feedback. So, I'm going to do what I

usually do in my real world day-to-day

work when doing things like this and I'm

going to take that response. I'm going

to paste it into the original chat with

the first model that made the plan and

see what it does. Oh, it updated the

plan alarmingly quickly as well. It's

It's very nice how fast this model is.

And since it's fast and token efficient,

it just feels great. Plan updated. Take

another look. It is nice having all of

these fast, cheap models getting smart

again. It's been a while cuz everybody's

been fighting so hard on the frontier

considering that OpenAI just lowered the

price of Luna by 80%. Which is a massive

drop on a model that was already really

cheap. $120 per mill out and 20 cents

per mill in is just unbelievable. It's

crazy that you can get a model with a

million context window with the

capabilities that Luna has for this

price. And I've been using it a lot more

for like random background tasks,

categorization, titles, and stuff like

that. We went from having almost no good

small models drop for months if not like

a year to having Luna come out and then

price drop massively by like 5x.

Deepseek V4 Flash just got a new

snapshot that I was testing earlier that

is really good, really fast, and really

cheap. And now we have this new Spark

1.2 model that's also seemingly really

good for the money, especially if you're

on that contributor tier where they are

reading all of the things that get sent.

After the one passive feedback from

Fable, apparently it thinks that the

plan Muse wrote is ready to go. That's

kind of nuts, man. I love T3

code. It's so nice being able to just

like hop between models, harnesses, and

computers. I'm now spinning up soul on

one of my Linux boxes to compare the

plans that Fable and Muse wrote

separately. See, we're up to 36 cents of

spend now that we've done all of the

planning. Oh, nope. 40 cents. Wow, I'm

going to go broke at this rate. For

reference, just my little tests with

Fable and Opus on this computer for

reading those plans and investigating

them is already $32 almost of spend in

Claude. So that is a 100x gap. And I've

done way less work with these models

than I have with the Muse one. That is

on the they have access to all my data

tier. But to be fair, Fable 5 also does

have them storing the data. They're not

training on it allegedly, but they are

storing it for safety reasons. Opus,

they don't do that with. And if you

spend the much higher 10 to 20 times

more on the standard tier for Spark 1.2,

then you will end up paying still not

anywhere near this much money. But if

it's 10 to 20 times more, which is

roughly what it is, that would be 4 to 8

at most for a bunch of work, that is

decent. Oh boy, it's spinning up those

sub agents now. Probably going to hit

rate limits again. Yeah, I'm hitting

rate limits again. I might have to

switch. I'll stop all of these.

We'll move off the contributor tier.

Let's see how it handles. Uh, continue.

It is nice only having one model and

just picking between the two versions

that have different pricing for the

exact same thing. But when one of them

hits rate limited that hard, it's a

little annoying. I just spend way too

much money in comparison. Even though

all the data I'm about to get here is

technically available for anyone to

train on because it's in my videos. The

review sub agents are finishing up their

work. I I switched to the higher price

tier and I'm still getting rate limited.

Are you kidding? Like I'm just doing the

sub agent stuff. They said that they

support well and I'm hitting rate limits

constantly. This is obnoxious. And it

sucks. It's like the harness actually

feels pretty good. It's like a minimal

polished up quad code. It is making my

laptop a little warmer than I would

like. But uh yeah, this is

I am annoyed by the rate limit thing

more than anything else here. The rest

is not the worst. Like rate limits on

subscription plans make some sense. rate

limits on paid per token tiers where I

am using it the way it's intended is

pretty rough. Oh, cool. Soul is now done

comparing things here. Yeah, this is

what I expected. It's a slaughterhouse.

It also decided to weight the different

categories more and less heavily

depending on how it felt about them. So,

it said for the current repository and

API fidelity that the plan from Muse is

4 out of 10. The plan from Fable is 7

out of 10. For the protocol research,

they tied roughly. For end-to-end and

multi-urface completeness, Muse was

nowhere near there. Fable was a lot

further along at 8 out of 10. Life cycle

permission safety recovery pretty close.

Delivery plan closeish with Fable Little

had at seven versus five. Simplicity and

maintainability both didn't score great.

But overall, Muse's plan got a 4.8 out

of 10 and Fables got a seven. Yeah,

considering the gap in cost, that is

reasonable. But considering the reality

that you have to merge the code when

it's done, this is much less reasonable.

Still a very interesting model, just not

necessarily

one I would trust for making big

sweeping changes to my code. Think now

is a good time to answer the important

question, why would someone use this

model? What makes Muse 1.2 useful enough

that someone should use it? Well, the

obvious reason, like the number one

thing that would make someone want to

use this model is that they work at

Meta, in which case they would probably

use this as their second or third most

used model compared to Opus and Fable

because I know they love those models

there. Main reason to use Muse is that

you work at Meta. But there are other

reasons that I'm seeing a bit of. The

biggest, of course, is that you like to

try new things, especially when those

things are cheap. Because this model is

cheap. Even if you're not willing to

share your data and you're using it on

the paid tier that's a little higher up,

it's still a very, very cheap model.

Although, now that I've switched out of

that way cheaper tier, I just went from

40 to $5.32 for the work I've done in

the past 10 minutes because it's going

so fast. But it did make its integration

for Muse inside of T3 code. So, I will

ask it to spin up dev server and share a

link. If it got this working properly

first try after a little bit of plan

review from Fable, then credit to them.

They made a model that's pretty good. If

this fails out right, then expected

models this cheap can't really do tasks

this exploratory and undefined because

again they didn't put out the SDK I need

to build this. So the model had to

integrate Muse into T3 code with hacks

more than anything. Moment of truth.

Let's see if we are in here. Muse

Spark12. What is this project? Not

looking good so far. [snorts] Didn't

even add it in the provider section in

settings. It just added it in the UI

here and it does not appear to work.

Ah, was a nice attempt. The

autogenerated title was from codeex

because that is the default. If you have

codecs, I just use 56 Luna on low for

it. So, the title gen was nothing to do

with Muse. It just appears to be broken.

Okay. So uh sadly what this means is you

cannot trust it for longer running

things for sure. I was hoping it would

be a little more capable at that but it

is not. So what other reasons would you

use this? You like really fast models

that are somewhat capable. Like this is

not anywhere near as good as a model

like 55. Even though the benches suggest

otherwise, it just doesn't get it as

well. And that's what makes this model

so strange to me is that it understands

breaking up work in sub agents and not

stepping on each other's toes when it

does that type of thing really well. But

it's nowhere near as good at actually

seeing complex work through. It's almost

like it it knows how to act like a

modern smart model, but it doesn't know

what the modern smart models know. I

like the comparisons people are making

with gro code fast because it does feel

similar there where it has a lot of the

the layers that make the models work

well, but it's not good enough to really

be trusted. And I'll be honest, it's

hard for me to justify using models like

this for a lot of my work just because I

would rather wait two to three times

longer and have something I can almost

certainly merge versus trying it five

times with a fast model and still have a

mess inside of it. Honestly, the most

impressive thing here is the CLI. It's a

lot more stable and less annoying than

Cloud Code. It's still not my favorite.

Like Pi still smokes it overall, but

it's solid for a thing that like they

threw together for this release. I am

tired of the labs making new CLIs and

things when they're already in last

place, forcing us to install yet another

thing, but they did provide integrations

for things like open code. So, that was

nice and right directionish. And it

still crushes everything Google is doing

by quite a bit. It is funny to see Meta

quickly leaprog Google but still be far

behind everyone else. It's almost like

these big companies are fighting to be

like fifth place and all the startups

that are really embracing the power of

these new models are excelling in

jumping far ahead. Obviously, Anthropic

and OpenAI are far ahead, but companies

like Moonshot with Kimmy, like Zai with

the GLM series and more are all far

ahead of this in my opinion. Hell, I

would still use Grock 4.5 above this. I

do have one last test I want to give it

though. I want you to audit all of the

poll requests I have open on T3 code.

Figure out which ones are mergeable,

which ones need more work, which ones

should be closed, which ones have been

trumped by other things merging, etc. I

want you to make a priority list for me

of what I should look at first and how

confident you are in me merging it.

Break up this work into lots of sub

aents in order to get through it faster.

Cool. We'll see how it does with that.

Bad gateway when trying to hit the

GitHub API through the CLI. That's fun.

Is GitHub down? No. Is this formatting

things wrong? Fun. It's doing some

sketchy to get in. Interesting. It

went through all my PRs in 4 minutes.

That's like genuinely impressive.

Especially cuz I spent the first two

minutes just trying to get into GitHub

and off it. If this review is of decent

quality, then this might be what I use

the model for. I might just use this as

my go-to. like go review all the PRs I

have open. Oh, that's a nice little

thing. When I have Codeex and Claude

make these PR audit pages, they often

don't link the actual PRs with these

like links here, and this did. That's

really nice. I have been annoyed at how

many models don't make these clickable

links. Even gave little confidence

scores. This is one of the better PR

review pages I've gotten. For reference,

here's one I made with a different

model, and I actually had to tell it,

"Please make sure these are links." I

genuinely prefer the layout of the

version that Muse did here by quite a

bit. I might even use this as like my

go-to template in the future. This is

super readable to me. I like the little

confidence scores. I like that it tells

you if it's clean or dirty merge, what

it thinks the status of things are. This

is good.

This might be one of the use cases I end

up using this for a bunch. So to go back

to use cases quick like fast models is

one. One of the big ones I've now

learned is for random code adjacent

analysis work. Obviously I wouldn't

trust this model to like actually go

merge PRs for me. But as a surface level

like go audit what's going on in this

repo for 20 or 30 cents. That was really

good especially with like no additional

effort to try and make it better. Like

that was great. I'm going to tell it to

go go to go further here. How about you

do a similar audit for all open PRs that

have had updates in the past five days?

If this audit costs less than like a

dollar, it was absolutely worth it. One

of the things that'll be hard for this

model with this though is that it's not

trained on GitHub because again, Meta

mostly uses their internal mercurial

stuff where all the other labs are heavy

on GitHub. So, they've trained the model

to be very good at GitHub. I'm about to

hit so many rate limits. It doesn't seem

to have parallel limits like all the

other harnesses do. It's more than happy

to run seven sub agents in the

background at once. And it gave each of

these a ton of PRs to look at. There's

$841 before I hit send on this. We have

spent another 5 cents since. It's going

to hit rate limits. We'll let that run

in the background. I'll come back to it

if it ends up with good results. But

back to why you would use it. I already

said you like to try new things, but I

really want to emphasize this point

because this model has a different

flavor. it like when we went through the

different designs it made, it did things

meaningfully differently. It still has

the like early Gemini 3 Pro era style

overall to it. Like a lot of these look

like what I saw out of Gemini 3 and 3.1

Pro, but it also just has little

nicities to it that give it a vibe

that's different. It's it's like a

little bit of seasoning that they added

that other stuff doesn't have. This

really is a model for enthusiasts right

now. Like you want to go play with it

because it's fun to play with new

things, not because this model is going

to magically save you a bunch of money

or become your go-to. Like no one should

use this model as their day-to-day

coding model. But what it is is

interesting. The things that Meta chose

to focus on, the things they didn't

focus on, and the capabilities that I'm

seeing here, it's fascinating,

genuinely, and I'm definitely going to

play with this model more. Probably

through a better harness, though,

because you can't really integrate this

one with anything. I don't know why they

closed source the Muse CLI here. Like

Muse code should just be open source.

Meta, you guys know better. You're an

open source company at heart. Just put

out the source. In under five minutes,

it was able to index and review 222 poll

requests. And this cost me a total of 10

cents on the contributor tier. That is

insane. Being able to go through that

much real work, like auditing 222 pull

requests in my codebase, organizing them

by how mergeable it thinks they are, so

I can quickly blast through this and

ship real code. That's insane. That's

actually valuable. And this is why it's

fun to experiment with the models. Like

try the different things you do and take

a look at how much it costs and how fast

it runs. Being able to hit a button and

spend 10 cents and in five minutes you

have a page like this for 200 plus poll

requests on your project. That's good.

That's useful. I'm impressed. I would

use this regularly and I might even set

something up to automatically do this

for me every day. That's cool as

Remember though that was on the

contributor tier. So if you're not

willing to share this data to Meta,

you're going to be spending 20ish times

more. But that goes from 10 cents to $2

for this type of work and this much

work. That's genuinely really

impressive. I I think you should play

with this model if you're interested in

this type of thing. Obviously, you

shouldn't trust it for everything. You

shouldn't just blindly go through and

merge stuff, but as a way to like pull

signals out of noise for really really

cheap. It's a fun way to experiment and

try things. stuff like generating

titles, categorizing PRs, generating

summaries for reports, digging through

logs to find useful stuff. Like this is

solid and doing similar work with Fable

cost tens if not hundreds of dollars and

this was literally 10 cents. So yeah,

not a bad model. I think it's going to

get a lot of because it has all the

weird quirks it has, but when you think

about it for the price, the speed, and

the capability, as well as it weird

things it seems to do not better, but

different from other stuff, it's fun. It

almost is like playing with one of those

like interesting toy programming

languages is how it feels to me. It's

it's different in a way that isn't

necessarily ready for real world work

all the time, but it's cool as and

very fun to play with. I definitely had

ups and downs with this exploration, but

overall I'm coming out somewhat

impressed. I don't think this model is

going to kill Opus or Fable anytime

soon, but it's a model that I could see

myself playing with for lots of weird

things, especially when you consider the

price. If I didn't have all of these

accounts across cloud code and codecs

that I just burn for all sorts of stuff,

I would probably be leaning on models

like this. And even though I have those

other things I can burn, I am still

going to be trying this to try and just

organize my work and life better because

it is so surprisingly cheap. I'm curious

how y'all feel though. Am I too Frontier

Pilled or is this model just not that

impressive? Let me know how y'all feel

and if you'll be using it. And until

next time, peace nerds.