Anthropic says AI built a prototype operating system

Welcome to The Signal from Superpower Daily with Maya and Theo

We are diving straight in today

Right so imagine putting 1 000 software engineers in a

room You lock the door for eight

days And you ask them to build an entire operating system from

scratch

Exactly Now imagine those engineers are not

human They cost a fraction of a cent per

hour Today we are looking at a major shift in how AI handles demanding technical

work Anthropic just dropped an experimental operating

system And they debuted the Fable and Mythos 5

2 models We are starting with our lead

story today Anthropic has debuted its cheaper Fable 5 2 and Mythos

5 2 AI models They are designed

for

demanding technical work The release includes

a striking demonstration The AI successfully built a prototype operating

system from scratch Let us unpack exactly what happened here because it

is a lot

It really is We have two new models

hitting the market They are Claude Fable 5 2

and

Claude Mythos 5 2 Right It is important

to clarify their architecture immediately I mean they actually share

the exact same underlying model Yes

that is a crucial point The difference is

not in their neural networks The difference is entirely

about access rules

and safeguards That is correct Anthropic has

essentially split their safety framework This is

a very deliberate structural choice Fable retains

strict cybersecurity and biology safeguards So

it is

playing it safe Extremely safe It actively

blocks most offensive cyber tasks It refuses to write exploits It refuses

to analyze certain biological pathogens

And Mythos is the opposite Well Mythos has fewer of these safeguards It is designed

to

be far less restrictive But and this

is the big but

it is strictly gated You cannot just sign up for

it on a website Right That

gating process is fascinating You can only get access

through

verified

organizational programs Anthropic calls them the life sciences and cyber

verification programs You basically have to prove who

you are You have to prove what your

company does You have to pass a real

vetting process This creates a fundamental shift for founders

and operators I mean we are looking at a two tiered

AI economy now

We really are You have a public tier of

highly

restricted models Yeah Then you have a verified tier

of unrestricted models So if your startup

operates in cybersecurity getting verified for Mythos becomes a

massive competitive advantage Absolutely You have tools your unverified competitors

simply cannot

access That brings us to the marquee experiment for this launch It is absolutely wild

It is quite the flex from Anthropic

Anthropic wanted to prove what Fable 5 2 could do So

an engineer started with a completely empty digital folder Just

a blank slate Yeah

Fable

5 2 then acted as a project manager

It coordinated over 1 000 AI

sub

agents 1 000 And it managed them for

eight solid days The scale of that

coordination is staggering We need to look at how that actually works

under the

hood Please do Because it

sounds

like chaos Well you do not

just have 1 000 agents all talking at once You have

a

strict hierarchy Right Fable 5 2

acted as the orchestrator It spun up specialized worker

agents

to write code Okay It spun up reviewer agents

to check that code It spun up testing

agents to compile it They worked in

this continuous automated loop And the output of that eight day loop is

an

experimental prototype operating system Yes I mean it is

not just a theoretical concept It actually boots

up on a physical laptop

That is the crazy part It is not just a simple

text based command line

interface either It includes actual drivers

That detail about

drivers

is crucial Why is that Well writing an operating system

means interacting with bare metal hardware It means telling the processor how

to

talk to the disk drive You are getting

way

down in the weeds It means telling the motherboard how to

read a physical keystroke The AI successfully wrote drivers

for the disk the keyboard and

the

network card Wow It built a functional

file storage system It even has a

graphical user interface Like you have windows you

can drag

around It does It has

a login screen It has a

full installer package But I have to admit I am

highly skeptical here I expected

you might be I really want to push back on this

operating system claim An OS built

in eight

days sounds incredible but we need to define

our terms here Skepticism is completely warranted Operating systems are

notoriously

complex beasts Right So is this like coordinating a 1 000 person

construction crew that never sleeps but they are just building a

house out of prefab parts

That is a good analogy I mean did it just take the Linux kernel and wrap

a new interface around it I really wonder how much human oversight was actually

required

to guide those agents

Yeah Operating systems require deep architectural planning They

require complex memory management

They require process scheduling That is

a very fair critique So did

it invent all that No Anthropic did not invent a new computing paradigm

from

thin air here The AI relied heavily on existing

open source libraries It used

known

architectural patterns

It was assembling known concepts in a novel way It was not creating a fundamentally new type of mathematics or anything

This brings us directly to the most important limitation of this release

We really need to make this clear for builders listening

You cannot download this OS and run your company on it today

Absolutely not The

operating

system was purely an internal experiment It is

just a demo It

is a proof of capability It is not a production ready product It

likely has massive security vulnerabilities And

probably severe performance bottlenecks Oh

definitely Anthropic

explicitly acknowledges this In

fact

they still explicitly recommend their older larger model for most tasks They recommend

Opus

5 5 for routine coding and knowledge tasks So Fable

is not replacing Opus for the everyday developer writing a quick script No it

is not Opus is

much

better at single turn complex reasoning Okay Fable

is

optimized for multi agent orchestration It

is designed to manage those thousand sub agents efficiently

That distinction brings us to why this matters beyond the headline

Right

It all comes down to actual cost and speed

Building complex software requires thousands of iterations

Yes it does

If you are running an agent loop

the token cost can bankrupt a startup very quickly

That is why the pricing model is so important here

Both Fable and Mythos are priced at 8 per million input tokens

and

they cost 40 per million output tokens Which

is pretty reasonable

It is significantly cheaper than previous high end orchestrator

versions It totally changes the unit economics of agentic

software There is also an optional

fast mode I think this is highly relevant for operators building

user facing applications Latency

kills

user retention Exactly Fast mode doubles

the API price but it gives you up to 2 5 times faster generation

speed That speed boost is

massive

Yeah And that is on top of a native 30 speed increase over the previous 5

1 models

Oh wow If you are building a customer

service agent you need responses

in

milliseconds Right So paying

double for fast mode is often a very rational business decision to protect your

user experience The raw engineering performance is what really stands out

to

me

though I mean cost is

great Speed is

great But the AI actually has to write good

code Fair

point Anthropic shared a very detailed internal migration

project Their own engineers used an earlier version of Fable to move a critical production

service This is a great

example They migrated the entire code base from Python to

Rust We need to explain why that is such a significant

achievement Yes Let us break that

down Python is a dynamically typed

language It is very easy for humans to write It is very

forgiving Right A lot of startups build their MVPs in

Python

Exactly

It uses something called a garbage collector to manage computer memory automatically

Which sounds

nice It is

nice But that convenience comes with a massive performance

penalty Python can be incredibly slow at

scale And Rust is

the exact opposite Rust

is

a statically typed language It is

notoriously

difficult

for

junior developers to learn It is brutal It

enforces incredibly strict

memory safety rules right at compile time It does not use

a

garbage collector It forces the programmer to

manage memory explicitly But because of that strictness

it is incredibly fast and secure Migrating a massive code base

from Python to Rust manually is a total nightmare I can imagine Human engineers

have to rethink

the entire logic structure They have to fight the

Rust compiler for weeks to fix memory errors Yeah The AI agents handled

that

translation autonomously That is crazy The result

was a 26

times faster startup time for that internal service 26

times faster Having AI

handle that level

of complex translation is a massive deal for engineering teams It really is It allows

startups to write

their initial prototypes in easy languages like Python Then when they hit scale

they

can use Fable to

automatically translate the bottlenecks into Rust It lowers the barrier to

enterprise grade performance And the capability extends

far beyond just software engineering It extends to physical

hardware

design too Anthropor released benchmark scores that

are frankly astonishing The BenchCAD scores Right

Fable

5 2 was

tested

on BenchCAD 2 1 This is a rigorous industry benchmark for hardware design How does that

benchmark actually work What is it measuring exactly Well it measures

how closely an AI generated hardware design matches a perfect human created reference geometry Okay

It

evaluates circuit board layouts It

evaluates physical tolerances Fable

5 2 scored an 83 5 out of 100 on BenchCAD 2 1 That was at maximum effort

settings To give our listeners some context on that

number Fable 5 1 only

scored 44 2 four on that

exact same test Wow so

nearly doubled it Opus

5

5 scored 65 3 So Fable 5 2 beat them

both by a massive margin It nearly doubled the capability of its direct

predecessor in physical hardware layout We also have a very interesting

cost

benefit analysis regarding software readiness Right Anthropic tested Fable

on the FrontierCode 2 0 benchmark I think this is a

crucial metric for founders to understand Yes FrontierCode is not just

about writing a simple function

Right It is much harder It tests whether the generated code is actually ready to merge into a production repository

without breaking the existing app

Which is the real test It requires deep contextual understanding

of the entire code base Fable 5 2 was tested at

high effort settings on

this

benchmark It scored 45 3 Okay It is incredibly

close to the massive Opus 5

5 model which scored 45 5 So they are

essentially tied on real world coding capability But the business implication here is the cost Fable achieved that identical score for roughly half

the cost per task compared to Opus This positions

Fable beautifully within the modern tech stack Yeah It works

best as an orchestrator or an advisor model You can use it

to delegate simpler

tasks

to

much smaller

cheaper models Like middle management Exactly Or you can have those smaller models escalate up and consult Fable when they encounter a hard problem It is

middle management for AI agents There is another major caveat for buyers that

we

need to emphasize though The primary constraint with this release

is access versus capability We mentioned

the two tiered economy earlier Yes

This is the reality of it Fable

safeguards actively block most cyber security

tasks And Mythos requires verified organizational access You cannot just swipe

a

credit

card and get the unrestricted model You have to pass corporate vetting This is a

massive

friction point for independent researchers Definitely It is a friction point for early stage stealth startups who do

not have a public footprint yet So what should

listeners watch next regarding this development You should watch this space incredibly

closely

over the next few months We need to see if verified organization gating becomes the industry

standard for foundation models Anthropic is

pushing this model hard If OpenAI and Google follow suit it creates

a permanent regulatory moat That

is a great point It is a new way

to distribute highly

capable less restricted models It effectively solidifies that two

tiered

AI

economy for builders You will have the public tier for consumers and the verified tier

for serious enterprise development That wraps up our look

at Anthropic's

new models Moving on Anthropic just showed us that massive models can build entire operating systems by

acting as an orchestrator But running 1

000 massive

agents is expensive Very expensive That brings

us to our next story Next up

Meta has released Muse Ember It is a low cost AI model specifically designed to handle

the

repetitive tasks inside AI

agents

The model processes

text images audio and video aiming to take routine sub jobs

off the plates of larger models Let us look at

what actually happened here with

Meta's strategy Meta launched Muse Ember globally The

specific

model identifier for developers is Muse Ember 1 0 Catchy It is available right

now via their Muse code platform and the Meta model API The technical specs

are genuinely

impressive for what is considered a small specialized

model They are It features a 1 million

token context window That is a massive amount of short term memory We should

define what a

context window really means for our non technical operators

listening Yeah please do Think of the context window as the model's workbench It is how much information it can hold

in

its active memory at one single time before it starts forgetting things

1 million tokens is roughly equivalent

to 3 000 pages of dense text That is a lot of reading

You can upload entire code repositories You can

upload dozens of financial reports The

model can analyze all of it simultaneously

It also handles

multimodal inputs natively I think

this is a crucial distinction It absolutely

is You can feed it screenshots You can feed it

complex financial charts You can feed it voice memos It

processes

them directly

Native processing is vastly superior to the old pipeline method How so Previously if you wanted an AI to read a

chart you

had to pass it through a separate vision model first Oh

right That vision model would write a text description

of the chart Then the main text

model would read that description That sounds

like a game of telephone It

is exactly a game of telephone It loses massive

amounts of nuance Ember

understands the pixels and the audio waves

directly It reduces errors significantly The pricing is where this

gets wildly aggressive Meta is

clearly trying to undercut the market

They are coming

in hot It is 0 10 per million

in input tokens

It is 50 cents per million output tokens which

is cheap But the most shocking number is the cash pricing It is an

incredibly cheap four tenths of a cent per

million

for cashed input That specific number

matters immensely beyond the headline Cash pricing is revolutionizing agentic workflows We

really need

to explain how this actually works under the hood Please do because four tenths of a

cent

sounds like a typo How are they offering compute that cheaply

It

involves something called

the

key value cache in transformer architecture Okay

the KV cache Right When an AI reads a long document it performs complex mathematical

calculations to understand how every

word relates to every other word That takes massive compute power But if you are using an agent loop your agent might need to

re

read that exact same document 50 times

to answer different questions Right like an agent

reading

a legal contract over and over to extract different clauses Exactly In the past you paid

the full computation price

every single time it

read the document Which gets expensive fast Cash pricing changes that The

server saves the mathematical representation of

the document in its memory When you ask the second question it skips the reading

phase entirely It just pulls the pre computed math from

the cache Wow

Meta is passing that massive computational saving directly to the developer So if I am

building an app that constantly re reads the same 100 page

legal document my costs just drop to basically zero after

the first read Precisely And that changes the entire economics of building

AI startups It makes long context agent loops

financially

viable for the first time This plays directly into

Meta's broader architectural proposal

They are

pushing a very clear division of labor

for

developers Their larger more

expensive model

is

called Muse Wonder Wonder handles

the heavy complex

planning

Ember executes the high volume structured tasks

The grunt work

Exactly We are talking about basic tool calling

routing user requests running unit tests and reading long data

files It is a classic corporate structure applied to neural networks It is exactly like hiring a highly paid strategist for the vision

and deploying an army of

highly efficient interns for the paperwork That is a perfect analogy You

do not

want your expensive strategist filling out routine expense reports You do not want a massive reasoning model like Opus or Wonder

just parsing a simple PDF

to find a date You hand that to Ember The

benchmark

nuance really supports that

strategist and intern analogy On Automation

Bench AA which specifically tests real world business workflows Ember scored 37 8

Let us put that in context It

beat GPT 6 Luna which only

scored 20 2 It also beat

Claude Haiku 5 5 which scored 13 9 That

is not

a small victory Ember nearly tripled Haiku's score

on business workflows It did Artificial Analysis dug deep into this specific

comparison They found that Ember matches Haiku's overall intelligence score across broad metrics

but it does so at a third of the estimated per task cost

That massive saving is

largely driven by that 60 cheaper cached

input rate That KV cache again Exactly When

agents re read the same documents cache pricing

changes everything about the competitive landscape Haiku used

to be the default fast model Ember is aggressively

targeting that exact

market share There is a very important limitation to

note however Always a

catch It is not a

clean sweep for coding tasks Ember is an intern not a senior software engineer Ember did beat Haiku on

Terminal Bench 4 0 The score was 40 9 to 32 8 But it

lost significantly to Haiku

on Frontier Code 2 0 Ember scored 27 2 Haiku scored

34 9 That distinction

between those two benchmarks is crucial for technical founders Let us explain why Terminal

Bench 4 0

tests whether a model can

run

local scripts and execute basic unit tests

It is a contained environment Frontier Code 2 0 as we discussed earlier

tests if code is actually ready to merge into

a massive complex repository It requires understanding deep dependencies across hundreds of files

So Ember is great for running the automated tests on a pull request but it is much

less

reliable for actually writing the final code to merge the

pull

request Exactly

You still need your strategist model for the final review

You do We also have a

massive unanswered question around safety metrics Oh Meta's transparency here is severely lacking They claim Ember stayed below their internal high

risk thresholds Okay But they published absolutely no numerical

results for everyday deployment behaviors What kind of behaviors are we missing data on The

behaviors that actually matter to a founder putting this in

front of a customer We

do

not know how well Ember resists hidden instructions in

a prompt injection attack We do not have numbers on how often it seeks human input before

taking

a high impact action like deleting a file or sending

an email Meta just said it is safe without showing the work That is a significant

risk if you are trusting Ember to act

autonomously on your behalf What should listeners watch next regarding this release You

definitely

want to watch whether developers shift entirely toward this multi

model architecture This

specific planner and doer

setup optimizes API bills perfectly It really does If it works consistently people might stop

relying on massive monolithic models from open AI for

every single agentic step If developers transition to this hybrid

approach en masse it fractures the market share of

the frontier model providers That shift towards specialized agent architectures is going to change how developers build backends

But the

front end prototyping

experience is also transforming dramatically It is In other news

Google is bringing Maps app building directly into AI

Studio This allows developers

to generate location based apps with Gemini This is pretty cool The workspace handles all the

API setup and credentialing while offering an interactive free prototyping sandbox

What

actually happened here is a huge workflow upgrade for builders In the past integrating

Google Maps into a new app was

a

tedious process Developers had to leave their coding environment

They had to navigate the complex

Google Cloud console They

had to provision a project generate secure

API keys

and manually inject them into their code base Now developers can

just

prompt Gemini right inside Google AI

Studio Yeah They just type in plain English

asking it to build a specific Maps powered app The workspace automatically connects

to an existing Google Cloud

project in the background Which is so much faster Or even better for new founders you can start

a brand new project without even entering a credit card It manages all

the complex API key generation and secure injection for you The result of that prompt is not just a text output

of

JavaScript code that you have

to copy and paste Right That is

the best part It is a live fully interactive map It renders and displays

directly

in the preview area of the AI Studio workspace You can immediately click drag and interact with

the geospatial app you just asked it to build This matters

tremendously because it dramatically reduces the friction of prototyping Getting from an idea to

a visual prototype used

to

take days of scaffolding Now it takes seconds Exactly It also integrates four major capability areas all

at once which used to require separate

API integrations We should break down those four capabilities They

represent a massive surface

area for new startup

ideas Let us do it First you get Maps which now includes

highly

detailed

3D city views and custom styling capabilities