Meta puts its newest assistant model behind a paywall

Welcome to The Signal from Superpower Daily with Maya and Theo

Meta is taking its newest and most advanced digital assistant Muse Wonder and locking it behind paid subscription plans

This puts a massive price tag not just on capabilities

It actually paywalls the absolute strongest safety protections the company's ever built Yeah

it is a profound shift in how frontier technology is distributed We

are basically watching a completely new boundary line being drawn in the AI industry We

are starting right there today with our lead story Meta

officially launched Muse Wonder

This is their absolute newest most capable assistant model

But they did not make it widely available to the public

No they did not

They restricted it entirely to paid subscription plans

The fundamental access model for Meta has completely changed

Right

So let us establish exactly what happened with the rollout

The base Muse app itself remains completely free People can still download it

Exactly You can still use that base application without paying a single cent

But accessing the actual Wonder model inside that app

requires a premium subscription

You have to pay 20 a month for the standard tier

or you can pay 100 a month for the advanced consumer tier

And that structure only covers the consumer side of the equation

The developer side has its own distinct tier system

It is much more expensive

Yeah developers who are using Muse code have a totally different setup to

navigate They get access to the more powerful Muse Wonder

full model but that requires significantly more expensive

monthly

plans Right Those developer tiers start at 50

a month and the top tier costs 200

a month It's just a highly tiered approach to

cognitive capability You are literally paying more money to get

more advanced

sustained reasoning We also have to look closely at the

Meta model API Developers need to access Muse Wonder through this API to build

their own products but Meta has not disclosed the exact pricing for that

programmatic access yet Which is a huge

missing piece

right Oh absolutely That missing API pricing is a crucial piece

of the puzzle The API price ultimately determines how third party businesses will build their entire software ecosystems on top

of this new model Well Alexander Wang

is Meta's AI chief He made some highly specific comments regarding

this upcoming API pricing He explicitly noted that the Wonder model operates

at a frontier level And you know the term frontier level is not just marketing

speak in this context It implies the model is actively pushing the absolute boundaries of

current

artificial intelligence capabilities And it also means it is a physically massive model requiring immense

hardware

to

run Wang specifically highlighted that it is much larger than their previous

offerings Therefore he expects the Meta model API access will naturally cost more than standard

consumer subscriptions Which I mean that makes perfect economic sense from an

infrastructure standpoint Larger models require vastly more computing power to process every

single word Yeah More computing power means exponentially higher electricity and

hardware costs for

Meta They frankly have no choice but to pass those immense operational costs directly onto the

developers But Wang also made a very aggressive competitive claim alongside that reality

check He claims the API cost will remain meaningfully below the leading frontier

models He specifically called out competitors like OpenAI and

Anthropic That is the delicate balancing act Meta is attempting right

now They want to charge a justified premium for their massive frontier

model but they still desperately want to undercut their main rivals to capture market

share So they are actively using price as a competitive

weapon But let us unpack why this model actually matters far beyond the headline pricing

structure There is a major behavioral shift happening in how MuseWonder actually

operates

Yeah this behavioral shift is perhaps the most fascinating technical part of the entire launch

It fundamentally changes how the AI agent handles failure

Because the new model is explicitly trained for resilience

Think about how slightly older traditional models work today

If they hit a roadblock while completing a task

they usually just stop working

Right They throw up their virtual hands They ask the human user for help They stall out completely

Yeah

They require constant

tedious human intervention to keep moving forward

Wang used a very relatable

everyday example to illustrate this difference

He talked about an AI failing to get a dinner reservation at a specific restaurant

It is a classic practical task for a digital assistant

You ask the AI to book a table at a highly popular restaurant for Friday night

And a traditional agent tries the task exactly once

It sees the restaurant as fully

booked

The agent immediately stops

It tells you it failed

and asks what you want it to do next

Which represents the old paradigm of artificial intelligence

It is a very brittle

linear way of operating

Exactly

But Muse Wonder is trained entirely differently

Instead of just stopping and giving up

it is trained to independently try another approach

It exhibits true agentic resilience

Yeah

It hits a wall

It pauses to assess the situation

Then it formulates a completely new plan based on that failure

And then it executes that new plan without ever stopping to ask you for permission first

So if the first restaurant is full

it might automatically look for similar cuisine in the same neighborhood Or

it might check for a reservation 30 minutes later It

tries to actually solve your underlying goal rather than just executing your literal first command Which

is amazing But

that shift requires a massive leap in underlying reasoning capability The

model has to deeply understand the broader unstated intent behind your initial prompt Yes

And it has to maintain perfect context over multiple complex steps It

has to critically evaluate the nature of its own failures in real

time Which brings

us directly back to the issue of cost That kind

of iterative looped problem solving is exceptionally expensive from a computational standpoint Oh it

is highly expensive Every single

time the agent tries a new approach it is

running a fresh inference pass It is

consuming thousands of tokens Right It

is burning expensive server time on Meta's end That architectural

reality is exactly why they have to put this capability behind a strict paywall But there

is another highly controversial layer to this specific paywall It is

not just about paying for computational resilience It is

about actively paying for safety Yeah this

is where the business optics and the broader implications get very complicated MetaClaim's Muse

Wonder features its strongest cyber misuse protections to date They are

extremely proud of these new robust safeguards And cyber misuse protections are absolutely critical right now They'd stop

the model from actively helping bad actors right Like

preventing

the generation of malicious weaponized code Exactly They

block

highly specific instructions for launching digital attacks against infrastructure But here

is the massive catch These new

incredibly robust safeguards are currently limited entirely to the paid

model Right They are essentially

bundled

with the premium commodity And I have to

push back on the optics of this decision If I buy a

new car the manufacturer does not

charge me a monthly subscription fee for the airbags and the seatbelts Safety is considered a

baseline requirement Why is Meta allowed

to paywall digital safety It is a completely

valid structural concern You are essentially watching

the creation of a class divide in digital security If these protections are

truly their strongest yet shouldn't everyone using the

ecosystem have them Shouldn't the free tier

be just as safe from generating cyber misuse material as the 200 tier Well ideally yes It

absolutely should

But

implementing these highly

advanced safety features also requires an immense amount of compute Because it is running

secondary models right Exactly The safety

filters

often

involve running specialized models alongside the main generation model They analyze the output

in real time before it reaches the user That adds significant latency

It adds massive operational

cost So if you are

a free user you get the older

weaker protections If you

can afford

100 a month you get the heavily

guarded safe version That seems incredibly problematic

for the broader digital ecosystem It fundamentally ties access

to both capabilities and protections together It treats rigorous safety

as a luxury feature rather than a fundamental baseline Meta is essentially putting

a premium price tag on their best possible behavior They are actively monetizing

the guardrails It is a bold

aggressive business move But it naturally invites

a massive amount of public scrutiny Let us examine the

biggest unanswered questions here We already mentioned the

exact pricing for developers through the API That remains a huge

blind spot for the entire industry right now We urgently need to

see that specific number Saying the price will

be meaningfully below OpenAI and Anthropic is incredibly vague Right We need hard

numerical

figures to truly understand the impending market impact The other massive caveat

is the long term future of these safety features Will these advanced

cyber protections eventually trickle down to the free tiers over time Meta has not

committed to any sort of public timeline for that The strongest safeguards

are currently locked firmly behind the subscription We simply do

not know when or even if the massive user base on the free Muse app will actually get them So watch how

Meta ultimately prices that API to compete directly with Anthropic and OpenAI Also watch closely

whether intense

regulatory or public pressure ultimately forces META to distribute these new cyber safeguards to the free versions They might simply

not be able to keep safety paywalled forever without facing massive blowback That concludes our

look at META's Muse Wonder moving from

digital agents to the racetrack Oracle Red

Bull Racing has upgraded its AI tool It went

from a simple search engine for Formula One rules into an active strategy agent making race day recommendations This represents

a fascinating high stakes application of AI in an intensely physical real time environment Let us

break down what actually happened on the track Oracle Red

Bull Racing maintains a deep technical partnership with Oracle They built

a highly specialized AI tool together for the pit wall And initially

that

specific tool had a very narrow clearly defined scope

It operated basically

as a highly specialized incredibly fast search

engine It was designed

entirely to help team engineers rapidly parse Formula One rules It was a

sophisticated tool for interpreting dense technical regulations And it also

searched through decades of past steward rulings to find precedents Formula One regulations

are notoriously overwhelmingly complex Oh absolutely They

are constantly

changing and evolving Having a dedicated

AI to instantly retrieve specific rules and past applications was a major competitive advantage But the team

has fundamentally upgraded the system It is no

longer just passively retrieving text It has officially

become an active strategy agent The technological shift

here is profound It moved from

simple data retrieval to proactive tactical recommendation It now recommends

specific high stakes race day actions It actively tells

the championship team what they should do in the chaotic middle of a live race This matters immensely beyond the headline Yeah and we

have

a perfect concrete example of this from the 2025 Italian Grand Prix The Monza example

is absolutely crucial to understanding this shift Max Verstappen was

driving

He was fiercely defending

his track

position against a close rival And in the intense heat of

the moment while battling for the lead Verstappen cut a corner on the

track This is a highly critical situation

in Formula One racing Right Cutting a corner gives

a

driver an unfair illegal advantage It almost always results in

a severe time penalty handed down from the race stewards So the AI agent analyzed

the entire situation instantly It recommended giving the physical

position back to the rival driver immediately It explicitly advised surrendering hard

fought track position It mathematically calculated that

giving up the spot right then was strategically better than risking a formal time penalty later in

the race And the human team actually listened to

the machine They followed this deeply counterintuitive advice They

got on the radio and told Verstappen

to let the other car pass him Which had to be incredibly difficult for everyone

involved Verstappen is fiercely competitive But he did give

the spot back And

he

subsequently overtook his rival anyway He ultimately

won back the spot legally But the

initial critical decision to yield was entirely

driven by the AI's calculation It requires a massive unprecedented amount of trust Executive

Jack Harrington explained exactly

how this system works

on the back end The AI is synthesizing an unimaginably massive amount of

context every single second It is not just looking at the isolated incident

on that one corner It is weighing dozens of dynamic track events Like

it looks at the immediate physical situation the

tire degradation the fuel loads right Yes But

it also weighs

off track developments

simultaneously

And most importantly it weighs the broader championship

point picture all

at once in real time It is making holistic strategic judgment calls It's

no longer just parsing static rule books

It is calculating the dynamic probabilities of

success over the entire racing season I have to highlight the immense psychological

friction happening here though Formula One is a hyper competitive ego

driven sport Without a doubt Every single

driver is biologically

wired to fight for every single inch of the track And every engineer sitting on

that pit wall desperately wants to maintain their track position Think about the pure adrenaline

You are traveling at 200

miles an hour You just fought tooth and

nail to keep the lead Exactly Then a computer voice

tells

your engineer you have to hit the

brakes and surrender I mean how hard is

it for a hyper competitive team

to actually follow a machine's advice in that exact moment It goes against every

single biological and professional instinct they possess Their entire instinct is

to defend their position at all costs But the machine's cold

calculation was to retreat temporarily to avoid a much larger statistically probable loss later It fundamentally requires overriding

raw human emotion with pure algorithmic logic And that

represents a massive

difficult cultural shift for

a legacy race team You are

quite

literally trusting a digital black box with millions of dollars and critical championship points But there is

a major limitation to note about this system right now Harrington firmly insists

the AI is only supporting the human engineers Right It is

definitely not replacing them It operates purely as an advisory tool It makes strong recommendations It does not

actually control the physical car It does not

talk directly to the driver over the team radio Humans still ultimately

make the final call The engineers on

the pit wall have to review the data agree with the

AI and then verbally

relay that command to Verstappen They are deliberately

keeping the human firmly in the loop The AI is

treated as just another intelligent voice on the headset albeit a phenomenally

data rich voice So watch

how this specific agent accounts for even more race context going forward The underlying

models will only get faster Yeah they

will ingest exponentially more live telemetry from the cars They will

become vastly more accurate in predicting the exact behavior of human race stewards And watch

if we eventually reach a point where humans simply stop second guessing the machine Will

there be a time when the pit wall just blindly follows its counter instinctual calls without hesitation

That will be the ultimate test of human trust in AI

When the machine says to pit for new tires

and the veteran human strategist strongly disagrees

who ultimately wins that argument

We will leave the Red Bull pit wall there

Next up on the topic of AI understanding

Andrej Karpathy claims the gap in public perception is widening

He says public understanding of AI capabilities is fracturing into what he calls a sharp funnel

Karpathy is offering a very structural

sobering view of exactly how human society currently perceives this rapidly accelerating technology

Let's look at what he actually posted on X He

argues that the general public's understanding of AI capability is fundamentally fracturing He noted that he previously thought of the AI debate as just two distinct groups speaking past each other Right

the

technological optimists on one side and the skeptics on the other But

he has entirely revised that binary view He

now describes the situation as a sharp funnel It

is a multi tiered deeply unequal breakdown of actual exposure to the technology

This funnel concept is absolutely vital to understanding the current disjointed narrative surrounding artificial intelligence Let us break down the exact numbers in his funnel

At the very top representing the widest part of humanity

are roughly six billion people

According to his estimate

those six billion people have barely encountered large language models at all in their daily lives

They exist almost entirely outside the ecosystem

Then you go down one level

Here we find roughly 1 to 2 billion people

These are the people who casually use free chatbots

They use the standard free tiers

They use them for writing simple

polite emails

Or they use them for basic search queries or brainstorming It is a highly surface

level fleeting interaction with the

technology Then the funnel gets dramatically

narrower The next level down consists of only 20 million people

globally These are the dedicated

professionals They consistently use leading frontier

models And crucially they use them for verifiable digital

tasks Karpathy specifically points to math and programming as the defining tasks

here These are domains where you can instantly prove if the AI is right or

wrong

Exactly The code either compiles successfully or it completely

fails This is a completely different cognitive experience from asking a chatbot to write a creative

poem Right A poem is

objective A Python script is

objectively verifiable

Then we reach the absolute bottom of this funnel

the tiny sharp sliver

Karpathy estimates this elite group is only about 5 000 people

worldwide These are the absolute

insiders These are the top researchers working directly at the major AI

labs These 5 000 people have unrestricted access

to internal unreleased

frontier systems These are massive systems that are absolutely not available to the

public yet And the capabilities at this specific level are vastly different from

anything else He describes autonomous agents collaborating continuously for weeks at

a time They are not just answering a single prompt

and stopping No they are working entirely autonomously on massive

cybersecurity projects They are actively doing scientific research They are

executing complex multi stage software engineering

projects

from scratch Karpathy argues that the nature of the task dictates your entire perception of

the technology This is

his central

crucial thesis If you ask a standard chatbot a short

simple question it reveals almost nothing about the AI's true

underlying capability The massive system barely breaks a sweat generating

an email It is exactly like judging the power of a

billion dollar supercomputer by asking it to calculate 2

plus 2 You are seeing

a result but you are absolutely not seeing its actual power

or architecture But assigning a demanding long running coding project that yields verifiable results shows a

completely different picture It deeply

stresses the system It showcases

sustained multi step reasoning The tiny group of people at the bottom of the funnel see this sustained

reasoning every single day And the billions of people at the top of the funnel have absolutely no

idea it even exists This highlights a massive dangerous gap between consumer reality and

internal AI lab reality We are essentially living in two

completely different technological worlds Average consumers still think AI is

just a slightly better

more conversational search engine While lab insiders firmly believe it is a

nascent autonomous digital workforce But I have to raise a massive caveat

regarding these specific numbers We must push back

on his exact figures Carpathy openly admits this is just

napkin math Yeah these are not rigorous peer

reviewed sociological studies These are simply his rough estimates based on his personal observations working deep

inside the

industry More importantly his grand claims about these internal frontier systems have not been independently verified by anyone

outside those labs We are essentially taking his word for it We cannot independently test these internal systems We cannot actually see these

supposed agents collaborating autonomously

for weeks This raises a huge fundamental

unanswered question Are these internal systems actually as capable as Carpathy's rough

estimates suggest Or is this just a classic case of lab

insider optimism Silicon Valley researchers notoriously overestimate the real world readiness of their

own experimental technology It is incredibly easy to see the brilliant potential of a system in

a highly controlled

perfect lab environment It is exponentially harder to make that same system work reliably in the

messy real world So we absolutely have to treat his sweeping descriptions of that 5 000 person tier with a

very healthy dose of skepticism But watch whether the true bottleneck for AI adoption becomes our fundamental human inability to check

the work these agents produce Carpathy warns that human

review is already falling behind This

is

the critical looming bottleneck Human comprehension is vastly

slower

than machine generation If an autonomous agent writes a million lines of complex code in a single week who

exactly reviews it If human engineers cannot even comprehend the sheer volume

of the output they cannot possibly

deploy it safely The systems might rapidly outpace our biological

capacity for oversight That is the real existential danger of

the sharp funnel We might soon possess tools we do not fully understand

how to manage We will leave the

Carpathy discussion there Moving on to

AI model testing Artificial Analysis has launched

a new benchmark Four top AI models are currently tied to the very top of their new

3D building leaderboard This represents a major necessary step away from traditional

text based code evaluation Let us get into exactly what

happened with this release Artificial

Analysis officially released AA3D It is a completely

new rigorous testing environment It judges models not just on their

raw underlying code generation It judges them entirely on the final visual

interactive output they produce It explicitly asks the models to

build 3 js web scenes They have to build these complex scenes

purely from short written briefs And 3

js is an incredibly complex JavaScript library It is used to create and display highly detailed animated 3D computer

graphics directly in a standard web browser It is absolutely

not a simple task for an AI The models are acting

as fully autonomous agents in this test They are given a brief

and they have

to write the

code render the scene visually

inspect it and fix their own errors They

get up

to 250 autonomous agent turns That is three full hours of compute time just

to finish each individual

run It is an

extensive grueling test of endurance And how do

they actually judge the winner

They

use blind human votes Right a panel of humans

looks at the final scenes The human panel does not know which

specific model built which scene They simply vote on which one looks better and functions better based on the original written brief These subjective votes are then used

to create an Elo leaderboard It is the exact same mathematical ranking system used

to rank professional chess players This matters beyond the headline because of the final results and more importantly the

massive spread in operational costs Let us look closely at

the top of that leaderboard Four different models are currently statistically

tied at the absolute top These

top four models are all built by

either Anthropic or OpenAI Claude's Sunit 5 5 technically led the pack It scored 1194 on the ELO

rating But the statistical confidence intervals overlap entirely There is no

clear definitive statistical winner among those top

four But while the raw performance is essentially

tied the operational costs to achieve that performance vary

wildly This is where the data gets incredibly interesting for

developers Let us compare the actual API

bills GPT 6 Sol is one of those top

four models It is the cheapest top five model on

the board It costs an average of 3 46

per generated scene It is also the fastest by

a wide margin It has a median completion time of

exactly 60 minutes Now compare that highly efficient baseline to

the Anthropic models Claude's Sunit 5 5 cost 4 4 times as much money as GPT 6 Sol to

achieve the exact same

statistical result That is a massive incredibly significant

financial premium for statistically identical performance And Claude Fable

5 1 cost 9 3 times as much The price differences

here are absolutely staggering for developers building businesses This massive price discrepancy brings us to

a crucial underlying concept revealed by

this benchmark It is called the context

tax This is a vital metric to understand 69 of the total reported spending across

these benchmark tests went entirely to input processing Almost 70 of the money spent was literally just paying for

the model to read its

own historical memory Why exactly does this happen It happens because large language models are stateless Agents have to actively

resend their entire accumulated context on every single turn Think about the

mechanics of a

250 turn limit On turn one the

model reads the

50 word brief On turn two it reads the