Loading page…
Loading page…
The Signal / Superpower Daily
OpenAI has new models and a lower price for Astra. Today’s workbench also follows Gemini’s shift from answering questions to handling errands.
Superpower Daily: The Signal
Episode guide
OpenAI has new models and a lower price for Astra. Today’s workbench also follows Gemini’s shift from answering questions to handling errands.
Full transcript
Select any speaker or timestamp to continue listening from that point.
Welcome to The Signal from Superpower Daily with Maya and Theo
Right out of the gate today
we're looking at OpenAI's release of the new GPT 6 2
models Yeah they are cutting asterisk prices and promising much better token efficiency but it comes with some pretty complex safety
trade offs
It does
And really today's underlying theme is delegation and its
hidden
costs Right We are looking at what actually happens when we hand over the keys
to AI whether that is for our API billing
our grocery shopping
or even our physics PhDs
So let's start with the lead story today
OpenAI has released its GPT 6 2 Astra and
Sol models and this rollout spans across
their API codecs and chat
GPT work Yeah and the baseline numbers
here are significant Astra's input and output prices
are down 20 While the Sol model maintains
its previous price But it promises near parity with Astra on some
specific
coding tasks Exactly The positioning of these two models tells
us a lot They are deliberately splitting
the
use cases Right Astra is basically framed as
the heavy
lifter It is You bring Astra in for
the hardest projects The examples they provided
really highlight this Like
the
screenshot one right Yeah exactly They showed Astra looking at a screenshot of a
single presentation slide and it doesn't just use the image as
a
loose reference It perfectly matches
the layout It identifies the exact hex codes for
the colors calculates the precise
font sizing and then writes the code to build a fully functional template
from scratch So Astra is the
lead architect Yes and it's the primary tool for modifying massive existing software
codebases But then Sol is this highly efficient
option Very
efficient And the performance gap is closing in some specific
areas
Right like that benchmark
test
Yeah Cognition's Frontier Code 2 0 benchmark It tests if an AI can make changes to real messy software projects that are actually good enough to merge into the
final product And Sol matched Astra
on this To within half
a point And it achieved that at one
fifth the cost Both were running
at max reasoning
by the way
That is wild Achieving near parity at an 80 discount completely changes the financial math
for enterprise
deployments It does You can't afford to run the heaviest models for thousands
of automated agents Which is why the GitHub Copilot integration
is so interesting It's actively rolling
out right now And the early telemetry from
GitHub
is fascinating Yeah Sol is catching twice as many real review issues as the previous GPT
6 1 Sol model While Astra's using fewer tokens
for agentic coding which is
a crucial detail I really want to focus on
that efficiency claim
Let's do it Because it's not just a
rate card discount They are saying these models
complete demanding tasks using fewer output
tokens
overall Right And a token is basically a fragment of a word You're paying for a
token
allowance Exactly There's a hard cap on how many tokens the model can process
for you So if a model used to need 2 000 tokens to solve a problem and now it
only needs 1 000
The math changes entirely You've effectively doubled what you can
do within your limit It's
like a multiplier effect Imagine you have a strict data
plan on your phone
That's
a good analogy Right And suddenly your phone learns how to seamlessly compress high
def videos by 50 You pay
the same monthly fee but you watch
twice
as many videos Exactly The industry focuses so much on the
raw cost per token but the real metric is the total cost
of the completed task Yeah If the model is smarter and reaches the answer with less rambling
the overall cost drops Your
daily productivity ceiling rises But I do have to offer some pushback on OpenAI's framing of
the Sol model
here
The parity claim Yeah They are leaning heavily on Sol matching Astra on that
Frontier
Code 2 0 benchmark Right But we have to remember that is
one highly specific coding test
It evaluates a pull request
A
developer submitting code for review Exactly
It's a review and submission loop It is not a test of general logical reasoning or
advanced math or creative problem solving
It's a highly mechanical constrained workflow Claiming parity based on that is like saying a sprinter is just as good
an athlete as an Olympic decathlete Just because they
tied
in the
100
meter
dash Exactly It's
misleading It is
a
vital
distinction Sol is wildly efficient but it's not a blanket replacement for Astra's broader
reasoning
Right We also need to critically examine this new ultra fast mode
they introduced The 4 5x
standard
speed Yeah The headline feature sounds great but generating text 4 5x
faster does not automatically mean your software project finishes 4 5x faster
No because a massive portion of software development is not typing
It's the system compiling the code
Or running automated tests
Or the AI waiting for feedback from the virtual environment
Speeding up the AI's
speed Typing doesn't speed up the server's compile
time And there is a severe trade off built into this setting It's not a
free upgrade Not
at all OpenAI clearly states that generating tokens at this rate eats up your allowance significantly
faster It's a direct punitive
trade off If you sprint at max speed you burn
calories faster You could hit your daily limit
before lunch Plus it's
heavily gated A standard user doesn't
get
this Right you need the 200 a month pro plan just
for Sol Or the 500 plan for both And you have to actively select it You have to consciously trade your longevity for burst speed
Yeah
But honestly
we have to address the most alarming aspect of this release
The safety regression
Yes
The internal system card shows both GPT 6 2 Astra and Sol hit the critical cybersecurity threshold
Which is a specific metric in their internal framework
It means they've moved from just explaining a vulnerability in theory
to practical application
They are demonstrably better at developing working exploits from known vulnerabilities
They aren't just pointing out a weak lock
They are handing you a custom milled lockpick
They
are crossing a hard line from educational to weaponized And
because of that OpenAI
had to tighten their default restrictions They
put heavier guardrails on cybersecurity assistance Which
creates a massive dilemma A
huge one Because by tightening restrictions they
are inevitably blocking the white hat hackers Right
If you're a legitimate researcher trying to patch a hospital database you
need to know how an attacker works But
if the AI refuses to help because of a safety filter the
tool is useless for defense You're
blinding the defenders to
thwart the attackers And OpenAI's alignment team explicitly acknowledges this problem It's
halting authorized vulnerability testing in corporate environments And
the paper notes these
models showed weaker initial resistance to jailbreaks too So
it's a constant game of whack a
mole You make the model smarter for complex
software and you inherently make it better at writing malicious
code So they put a heavier muzzle on
it Which frustrates the legitimate developers paying premium
prices So the most important thing to watch next is how developers actually navigate these tightened guardrails in the wild
And we need to watch GitHub copilots gradual rollout Will these models face administrative friction
Yeah
CISOs are going to read that critical threshold rating and they might just block the update entirely
We are officially closing the lead story here
And that risk of delegating complex coding ties directly into another kind of delegation
One that's quietly reshaping enterprise software
We're
talking about handing over your wallet Next
up Anthropic has launched a linked billing beta program Right
It lets customers pay for CLAW directly inside third party tools like Linear and
Asana This allows those platforms to charge a customer's existing Anthropic organization without ever handling their actual API
key To grasp why this is a massive
shift we have to look at the old
system
Right Historically you had to generate an API key from Anthropic and paste it into the third
party tool And an API key is essentially a digital blank check tied to your corporate
credit card You are transferring possession of a credential that unconditionally authorizes
paid requests Which puts an immense security burden on that
third party platform They have to store
it encrypt it And if they get breached your
key gets stolen Malicious actors will use it to run automated bot farms
and rack up 50 000 in a single
afternoon
Plus the third party has to build complex engineering just to track your usage and invoice
you But
Anthropic's new linked billing beta removes that credential handoff
entirely It replaces the physical transfer with a one time pre
approved
digital connection Right An admin approves a direct connection just once
From then on Asana
can send requests but they never touch
the billing credential Anthropic bills
the customer directly And the customer can set strict
monthly spending limits I love the structural
analogy for this The old API key system is like walking into
a crowded bar handing the bartender your physical corporate credit card and telling them to keep
a dab open You're just trusting them not to lose it
or
clone it Exactly But the new linked billing system is like setting up a pre approved tab through the bar's mobile app before
you even walk in You authorize a
max spend of 50 The bartender just taps a screen
to log the request They
never touch your card
They hold zero liability This drastically lowers the barrier to
entry for software builders Startups don't have to build complex billing engines or buy
massive cybersecurity insurance And they can choose to split the
bill per request Linear could
cover basic features but pass the cost of heavy compute directly
to the customer It removes the financial risk
for the startup But and this is the crucial limitation listeners
need to understand this
introduces a massive glaring
security blind spot The
boundary of execution Yes Because the AI requests are originating from and running inside the third
party platform's organization the customer's enterprise security settings
do not apply You get
complete spending control but you completely abandon your
data security blanket Let's
make this concrete Say you're
a major bank You have strict rules in
your anthropic dashboard
Whitelisted IP
addresses custom encryption zero
data retention policies When your employee uses linked billing
from inside Asana none of
those
rules apply Right
Asana's policies apply Your legally mandated security
configurations are completely
bypassed Which makes this an absolute non starter for
regulated industries A hospital governed by HIPAA cannot bypass data retention rules just for a smoother
billing experience The compliance department will flag
it instantly So what listeners must watch next is the enterprise
adoption rate Will Fortune 500 companies adopt this or outright block it due to the loss of those
security controls Anthropic solved a massive headache
for startups but introduced a compliance nightmare for
enterprise IT We're talking about delegating our
billing
infrastructure Meanwhile Google is pushing delegation into the
physical world They are asking us to delegate our personal credit cards and
daily errands Google is rolling out free Gemini tools in six countries that allow the AI to actively call businesses analyze
utility bills and
shop online Gemini is moving from simply answering questions to executing multi step actions on
the live internet powered by the Gemini
4 hydrogen model This is
a fundamental evolution We're going from conversational advisor
to active administrator Gemini is currently connected to over
60 external applications VP Josh Woodward provided
some wild examples He had Gemini call two
separate car dealerships It spoke to human receptionists
navigated scheduling systems and cross referenced his calendar to
book an oil change in
under five minutes He also gave it access to his secure utility portal
and email inbox It analyzed a spike in a
389 electricity bill It cross referenced hourly usage data against peak pricing and his heat pump schedule to diagnose exactly why
the bill was high And he told it to shop for
USB C chargers and cereal
It navigated
the retail sites
selected items and handed a populated
part to Instacart To do this Google is using the
Agent Payments Protocol working with
Visa and Mastercard It uses single use digital tokens so the AI never holds the actual 16
digit credit card number And it runs
in isolated virtual machines We need to
unpack why that's
necessary
To
prevent prompt injection right Exactly Prompt injection is when a malicious website hides invisible text telling the
AI to ignore its instructions and do
something malicious like wire money But
by running in
a secure sandboxed virtual machine
Google limits the blast radius The AI can't access the user's broader operating
system or steal persistent credentials The scale
of ambition here is staggering This is the holy grail of agentic AI moving away from passive text generation
to executing external
software states Giving an AI always allow
permissions fundamentally changes the relationship You are granting software a limited power
of attorney over mundane tasks But we need to
heavily pump
the brakes here Woodward's examples are entirely self reported anecdotes They are carefully curated narratives from the
VP of the
product not independent tests It is highly convenient that his interaction with a
car dealership
receptionist
was utterly flawless Exactly The real internet
is
chaotic and messy Furthermore Google admits this is not
a fully autonomous agent It constantly hits
friction on standard websites Often it has
to build a cart stop and return control to the human to physically click
the final buy button
like the Instacart example Which
raises a serious question Are we achieving autonomous agents or are we just turning
AI into an intern An intern who
preps the tedious paperwork but still needs the boss to
sign the bottom line
That analogy is perfect The modern web wasn't
built for AI agents it was built for
human eyeballs and clicks Websites deploy massive security just
to block automated
bots CAPTCHAs behavioral
analysis IP rate limiting So Gemini
fills out the forms hits a
wall of anti bot security and has to tap the user on the
shoulder for a human fingerprint So the critical thing listeners should watch next is the rollout
of the Universal Commerce Protocol We need to see if Google can negotiate back
end programmatic access with major retailers If they can't
convince Walmart or
Target to whitelist
their bots the product fails
The illusion of delegation shatters if you
have to keep solving CAPTCHAs This success depends entirely on the
rest of the internet cooperating Google wants Gemini
to handle our mundane busy work But what happens when that
busy work is
what builds human expertise In other news theoretical physicist Suvrat Raju has officially paused
new PhD admissions to his research group He made
this decision after an AI agent nearly aced
a highly demanding theoretical physics benchmark GPT 6 ASTRA solved 49 out of 50 unpublished
mathematically rigorous problems on the Theophys Vell
test And the nature of this
benchmark is profound These aren't standard high
school kinematics No these are 50 completely unpublished problems that
cannot possibly exist in the AI's core training data It takes a senior human expert about a
full working day to solve just one of these
The agent running Insight Codex with WebSearch and Python tools systematically worked through the
math and solved 49 of them For less than 3
per problem and
compute cost That is staggering And the most wild detail is what happened with the
single problem ASTRA supposedly
failed Oh this is crazy The researchers compared the AI's final output to
their human written answer key
The AI's answer was different But upon closer inspection
the AI was correct ASTRA discovered subtle cascading errors in
the human written solutions The researchers had to rewrite their own grading rubric to give the AI its
near perfect score The automated student caught
the professor's mistake But the broader implications here aren't about AI discovering
new physics It's about the rapid collapse of the
apprenticeship
model Right MIT physicist Daniel Harlow noted that 95 of his daily work is simply deciding what specific question to ask the universe The actual
mechanical calculations are just the final 5 But how does
a junior physicist learn to identify those good questions Historically human beings
learn
by doing that bottom 5 The master painter makes
the apprentice mix paints for years Graduate students spend years
grinding through tedious calculations And through that painful
grinding they build an intrinsic
intuition for how the universe works They learn what a
dead end looks like before they even finish the math If we remove that
mechanical grinding phase the training pipeline breaks
An AI can do
a full day's worth of calculations for 3 in seconds Senior researchers will obviously outsource
that to publish faster But if they do how
do future physicists develop
their intuition We are looking at the
hollowing out of the middle class of scientific training If you're a mid level
manager or a junior developer listening you have to ask
yourself this How much of your
daily mundane grunt work is secretly training your long term intuition Handing entry level
tasks to AI optimizes current senior experts but actively
destroys the training ground for the next generation Sufrat Raju
pausing his PhD admissions is a completely rational defensive response What is
he supposed to train them to do if a 3 API call does it better This introduces
a massive structural crisis The data
notes a sudden 30 surge in paper submissions to the arc of preprint server and high
energy theory Because if
AI is generating the math the volume
of papers scales infinitely If humans
cannot physically read and verify all of them the entire
peer review system fundamentally breaks down The bottleneck
moves from idea generation to verification of truth We can
produce infinite papers but have
a strictly finite number of experts to validate them So what
listeners should watch next is how elite research institutions like MIT and Stanford adapt their
graduate curriculums They urgently need
to figure out how to teach scientific judgment without relying on the calculation heavy grunt work Let's move to
a few quick reads Waymo has officially
suspended driverless passenger freeway rides nationwide after the California DMV labeled the service an imminent public safety hazard This statewide ban
stems from two specific failures Waymo vehicles repeatedly
entered a clearly marked closed construction zone on US 101 And in a
separate incident they interfered with
highway patrol units executing a rolling traffic brake on Highway 280 While surface street
rides continue Waymo has
completely lost the critical high speed peninsula routes Autonomous
systems struggle in construction zones because they lack standard lane lines and involve unpredictable human traffic directors The
algorithms struggle to map shifting environments at 65 miles per hour And the
major caveat here is there's no set reopening date Waymo must
definitively prove they have implemented a functional software fix Listeners should
closely watch for NHTSA's reaction given Waymo
had a previous nationwide recall over very similar construction zone failures Next up
Anthropic is likely to price its upcoming IPO above its originally proposed range of 70 to 78 a share according to
CNBC The institutional
order book closes on November 17 setting the
stage for their NASDAQ debut on November 19 Their financial
metrics are staggering They posted 33 24
billion in top line revenue for the first nine months a 1
200 year over year
increase Driven largely by enterprise adoption of Claude
Code but they also posted a 104 billion net loss for the same
period Have to explain that loss
Yeah it is
skewed by fair value accounting changes on convertible notes When
the company's valuation skyrockets the
paper value of the equity owed to early investors also skyrockets Accounting
rules force them to record that as a massive loss They
didn't actually burn 104 billion operating servers Listeners should watch exactly
how this stock trades on November 19th Pricing above the initial
range is a massive flex of confidence But doing so in
a market that's highly skeptical of AI valuations is risky Institutional confidence doesn't guarantee
retail success Meanwhile Microsoft has launched
a
new suite of specialized AI models directly inside GitHub Copilot They're claiming steep cost
savings by prioritizing narrow task specific performance over broad intelligence Microsoft introduced MAI Code
2 and Code 2 Flash aimed at specific code
repository tasks like syntax formatting They claim the Code 2
model delivers 98 of the quality of OpenAI's GPT 6 1 sole model but at only 40 of the price They are also rolling out specialized models for cybersecurity log analysis
and image editing Narrow the focus drop the compute price The limitation listeners must keep
in mind is
that these claims rely
on selective internal benchmarks And the 40 price drop includes temporary promotional discounts Furthermore there's a
massive catch regarding enterprise customization If a company wants
to
fine tune these models using proprietary data
Microsoft forces them to deploy within the Azure cloud ecosystem A
classic walled garden strategy So watch to see if enterprise
developers actually switch from Cloud
or OpenAI to these native Microsoft models We need to see if the cost savings overcome the resistance
to strict Azure lock in We're moving to our three takeaways from today's deep dive
First the cost of artificial intelligence is dropping incredibly fast but
the
performance metrics are becoming blurry We see OpenAI focusing
on output token efficiency and Microsoft launching
specialized models with steep discounts The industry is pivoting The battle isn't about who has the smartest
model anymore It's a brutal war over who is cheapest per specific task Second trust is becoming the absolute bottleneck for deployment We see this
across every sector Waymo has the tech to
drive on freeways but regulators don't
trust them in construction zones Google's Gemini can
execute web purchases but Internet architecture doesn't trust
bots Anthropic built a seamless billing system but enterprise IT rejects it
over security settings We have the tools but we don't trust them to run
entirely unsupervised Third the
human apprenticeship model is under severe threat The physics
benchmark
proves the grunt work we used to train
junior scientists is now vastly cheaper to run through an AI agent Which leaves a massive gap If machines do the foundational grinding
we do not know how
we will teach deep human intuition
to the next generation For tomorrow the key development to watch is the final closing
of Anthropic's
IPO order books It sets the final stage for their public market debut and
will provide the ultimate test to see if retail investors
still buy into massive AI valuations despite a skeptical macroeconomic backdrop For more on all these stories head over to superpowerdaily com
Thank you for listening and we'll
see you tomorrow
and we'll see you tomorrow.
Original reporting
Read the complete Superpower Daily coverage behind this episode, including reporting context and source links.