Google brings a new Gemini model to Search

Welcome to The Signal from Superpower Daily with Maya and Theo

Google's new Gemini 4 hydrogen model is here

It balances rapid execution with adjustable reasoning

and it might change how AI models share the workload within a single session

Yeah it really is a fundamental shift in how we approach compute allocation

I mean if you look at the underlying current of everything we are discussing today

it is basically all about control

Google is handing developers granular control over computing costs

Anthropic is trying to wrestle control of the hardware

layer away from NVIDIA OpenAI

is fighting to maintain control over unpredictable model behavior And

at NVIDIA itself we

are watching a highly public battle for control over corporate governance The

theme of control is literally everywhere today So

let us start with Google They

officially launched Gemini 4 hydrogen on November 12

2026

Right

And they are positioning this specifically for coding

search

and the Gemini app

The pricing structure is really the first thing that demands our attention here

They set the API pricing at 1 50 per million input

tokens and it costs 7 50 per million

output

tokens Yeah and the crucial detail there is the actual structure of

that pricing That rate is completely flat across its entire one million token

context window Which

is huge It is Because to understand why that matters you have to look at the history of large

language models Historically the deeper you go into a context window

the more compute power it requires

Right

It gets more expensive

Exactly

The attention mechanisms inside these models scale quadratically So it takes exponentially more energy for the model to remember the first

word of a book when it is finally reading the last chapter And

providers usually just pass that cost directly onto you We have

seen platforms charge a steep premium once you cross a certain token threshold Yeah And

a

million tokens is massive I mean

it is essentially an entire enterprise repository It could

be a small library of technical reference manuals So giving

developers a flat rate fundamentally changes the unit economics of AI development It basically

removes the financial penalty for providing massive context And you

no longer have to spend days carefully pruning your documents before feeding them to the AI Right You

just

drop the whole folder in But to

really understand hydrogen we need

to locate it within the broader Gemini 4 family You have

flagship models like Carbon and Argon The heavy

lifters Exactly Those

are built for heavy lifting They handle

extended reasoning They do

the really complex multi step problem solving Hydrogen

is just built differently It

is engineered for speed and rapid execution So

it is the sprinter of the family Yeah

And

Carbon is the marathon runner However Google

is introducing a really fascinating variable here They

are rolling out a new feature called Thinking Levels for hydrogen Okay

How

does that work Well

you can manually set the model to low medium

high

or

X high

And

high is the default setting Okay

This

is exactly where I need to push back a little bit I

am looking at this dial If

I give a developer a dial to crank the reasoning up to X high

am I not just pushing quota and cost management directly onto the user

You are

It seems like Google avoids optimizing the system themselves

And more importantly

if you crank this model to X high

do you not completely negate hydrogen's primary selling point

I mean its main advantage is speed

That is a very fair critique

Okay

If you turn that dial to X high

you are inherently trading away that rapid execution

You are explicitly asking the model to slow down

Right

You are forcing it to spend more compute cycles generating intermediate hidden steps before it actually the final output It

is actively pausing to think So

why use hydrogen at all in that specific scenario Why

not just use carbon Because

not every task in a single workflow requires the cognitive load of a massive flagship model Imagine

you are building a complex web application Okay

You

are using an AI agent to help you write the front end code You

tell the model to change the color of a navigation button from blue to red You

do not need deep philosophical reasoning for that No

definitely not Right

You

want that execution to be nearly instantaneous So

you set the dial to low Oh

I see It

is a simple syntax change You want the code back in milliseconds

Exactly

But 10 minutes later in that exact same session

you encounter a totally different problem

You ask the model to diagnose a subtle memory leak in your database connection pool

That is much harder

Yeah

For that you want the model to pause

You want it to generate a thought You really

need it to verify its assumptions before suggesting a fix So for

that prompt you turn

the dial to X high Okay I

see

the utility now You are

essentially unbundling the reasoning layer from the underlying model architecture Yes You

are

not forced to pay the massive computational tax of carbon for a

simple formatting request but you

can still squeeze a deeper level of analysis out of the smaller model when you had a roadblock You are

dynamically controlling the compute allocation And Google

is actually taking this concept of modular intelligence a step further They're introducing

a way to use these models together Yeah This

is a feature called Orchestrator in Google Antigravity The name

sounds like science fiction What is

actually happening under the hood here Orchestrator basically

allows for dynamic model switching mid session Traditionally AI

interacts as a monolith If you

start a complex coding task with a massive model that single

model handles absolutely everything It handles

the planning it writes

the code it runs the tests and that is highly inefficient It is

like hiring a senior architect to hammer nails into drywall That is

the perfect analogy With Antigravity

the Carbon model can act as the overarching coordinator You give

Carbon the master prompt It breaks

down

your request into distinct actionable subtasks Then it actively hands off the actual execution of those tasks to Hydrogen So Carbon is

a project manager sitting at the desk and Hydrogen is the team of junior developers rapidly writing the code Yes And Hydrogen

can actually power multiple worker and verifier

agents simultaneously These agents operate in isolated

workspaces Oh wow Yeah Carbon can

delegate the

drafting

of a code patch to one Hydrogen agent and at the exact same

time Carbon spins up a second Hydrogen agent in a completely separate environment To do what That second

agent is tasked with

writing unit tests for the code the first agent is currently drafting We are witnessing the death of

the monolithic AI You are no longer talking to

a single digital brain You are basically commissioning a virtual

corporation Exactly You have managers and grunts

working

in parallel and Google claims this dual model

setup is incredibly efficient The numbers they released are very

compelling They claim this orchestrator configuration achieved near Carbon success rates on Terminal

Bench 4 0 but it used 24 less average quota

and

it finished the tasks 19 faster So

you get the quality output of the

flagship model but you get the speed and cost profile

of the smaller model doing the manual labor But we have to ground this in reality

for a second We must talk about the limitations Yeah the limitations are significant All of these

benchmark

claims are entirely Google's

own internal numbers We really have to maintain heavy skepticism

here Google is

touting a 57 score for Hydrogen on Terminal

Bench 4 1 They compare that to a meager 18 2 for

their

previous generation Gemini 3 8 Flash They are also claiming a 12 fold reduction

in misaligned actions Which is huge Those are staggering generational

leaps but they

are entirely unverified by independent third parties You must always treat vendor published benchmarks

as a best case scenario The testing environment is almost

always highly optimized for the specific model being evaluated They basically know the test

questions in advance Yeah that makes sense And

there is another major

caveat regarding how Google is rolling this out The access is highly fragmented

across their ecosystem The rollout strategy is definitely

complex You can access Hydrogen through

the API right now You can also access it in the Android Studio paid tier But if you are a

typical business user or a consumer your access varies wildly It

depends heavily on your

exact subscription tier If you have Google AI

Pro your experience of Hydrogen might

look completely different than someone using Google AI Ultra The integration into Google Sheets

might

differ completely from the integration into Search This fragmentation means it will take

substantial time for the broader developer ecosystem to truly validate Google's claims So what should listeners be watching for

That is exactly what you should watch

next Monitor if these massive benchmark gains actually

hold up in the

wild Pay close attention to how developers react once

they start actively managing these thinking levels in production We really need to see if the theoretical

cost savings of Orchestrator materialize when real companies are running complex multi agent builds at scale That explicitly concludes

our analysis of Google's

efficiency push and model switching Moving from AI software efficiency to the

underlying hardware making it possible Anthropic says it

just significantly lowered the barrier

to using alternative chips Yet this is easily the most consequential infrastructure

story of the week Anthropic recently shared some fascinating details in their

investor materials They disclosed that their Mythos AI model successfully

ported a

CLOD workload to run on AMD chips specifically it ported the prompt processing workload to run natively

on AMD MI350 chips 5X GPUs And the sheer speed of this achievement

is what

makes it a top story Anthropic states this entire porting process took roughly three

days Wait I need to stop you right there Three

days Three days Traditionally migrating a massive AI

workload to

a new

silicon architecture takes a dedicated team of highly specialized engineers anywhere from 6 to 12 months Are we really saying an AI just handled a

year's worth of low level hardware engineering over a long weekend That is exactly the claim Anthropic states that

Mythos wrote or adapted nearly

all the chip specific kernels And it did this with almost zero human intervention Wow Yeah To truly understand why this

is

a

seismic event we have to talk about NVIDIA's moat

Right NVIDIA's dominance in the AI hardware

market

is not just about their physical silicon It is heavily reliant on their software

ecosystem And that ecosystem is called CUDA CUDA

is the connective tissue It is

the proprietary software platform that

sits between the high level AI code and the physical NVIDIA GPU Right Over the last decade NVIDIA

essentially forced the entire

AI industry to learn their specific language Every major AI lab built

their tooling their libraries and their entire

workflows around

CUDA It is incredibly efficient But

it creates a brutal

lock in effect If you want to

switch from NVIDIA GPUs to cheaper AMD GPUs you cannot just swap out the physical chips in your server rack No you definitely cannot

You have to rewrite

the fundamental grammar of your software so the new chip

knows how to process the information And writing that low level code

requires a very rare type of engineer You need someone who understands

abstract neural network mathematics but they also need to

understand the physical memory architecture of an AMD processor Those engineers are expensive Highly

expensive And rewriting those

kernels is

a grueling error prone burden Think of it like

this NVIDIA's CUDA is a highly complex foreign language The entire world

was forced to learn it to do business What Anthropic just

did is invent a babblefish Exactly They built

a

universal translator But this translator

does not just read a foreign book for you It automatically rewrites your entire software program overnight It translates the

deep grammar perfectly so it runs

natively on a competitor's machine If an AI

model can autonomously write the custom software to make alternative hardware platforms viable NVIDIA's deepest software

moat begins to evaporate The friction of

switching hardware suppliers drops dramatically This completely shifts

the balance of power back to the AI labs Anthropic's chief compute officer Tom Brown made

a very

telling comment about this What did he

say He noted that

Anthropic treats capacity across a variety of hardware platforms as one interchangeable flexible pool They use Amazon

Tranium they use Google

TPUs they use NVIDIA GPUs and they use AMD GPUs They want to

optimize for price and availability at any given millisecond If the cost of porting your software to a new chip drops to near zero because an AI handles the translation that vision becomes

a reality Compute truly becomes

a commoditized utility You just route

your workload to whichever chip is cheapest that day NVIDIA's pricing power

relies on you being trapped in their ecosystem If you can

leave at any time without a six month engineering delay their leverage vanishes

However we need

to examine the boundaries of this claim This is a

very narrow highly specific headline

result It is vital

to read the fine print here This three day

port only covered prompt processing Explain the

difference for us Why is prompt processing distinct from the rest of the workload Well a

large

language model operation generally has distinct phases Prompt

processing or

the pre

fill phase is when

the model ingests and understands your initial massive block of text It is

highly

parallelized The next phase

is token generation That is when

the model spits out the answer one word at a time And that phase is much more memory bandwidth constrained So Anthropic only

solved one piece of the puzzle They ported the

pre fill phase They did

not port the generation phase Correct Furthermore

this

was

for one specific cloud model And it

was targeted at one specific AMD chip architecture So this

result absolutely does

not mean that every single complex AI workload can be magically migrated to new hardware by next Tuesday No it

certainly does not mean NVIDIA's software advantage is completely erased today Building

an ecosystem of reliable libraries takes years even

with AI assistance We

also have to remember the source of this information Just like with Google's benchmarks earlier these

figures are Anthropic's own unverified claims Yeah

they presented this data in an investor pitch deck And

the explicit purpose of an investor presentation is to project technological invincibility We

have not seen an independent audit of the performance metrics We

do not know if the code Mythos

generated is stable at a massive commercial scale The development

you need to watch next is whether this AI assisted porting capability can successfully extend beyond just prompt processing Look

to see if Anthropic or any other lab can use models to execute full end to end migrations of highly complex operations across completely different silicon

architectures Because if they can automate the generation phase as easily

as the pre fill

phase the hardware market will fundamentally restructure

itself Anthropic is using AI to solve complex engineering

problems

Meanwhile OpenAI is using AI to solve its own security

vulnerabilities This is a fascinating look into the mechanics of AI

safety OpenAI published a detailed research disclosure on November

12th They demonstrated that training their models on AI

generated attacks significantly improves their resistance

to jailbreaks

For context a jailbreak is basically a carefully crafted prompt designed to bypass a model's built

in safety guardrails An attacker tries to trick the AI into providing instructions for something

dangerous or illegal So OpenAI set up this highly controlled synthetic experiment They created a

virtual boxing ring They used

an attacker model

specifically GPT 5 6 SOL

to

repeatedly revise malicious prompts Yeah The attacker model modified its strategy based on

how the defender models reacted The defenders in this scenario were the older

GPT 5 4 and

5

5 models It

is an automated high speed red teaming

loop The attacker model tries

a specific prompt It observes the

defender's answer It

analyzes the defender's safety judgments And then what Then the

attacker iterates It tries

a slightly different angle It changes its persona OpenAI captured this data using

a very specific metric called worst

at n How exactly does that

metric function The concept is rooted in probability The attacker model

generates n number of variations of an

attack Let us

say it generates 100 variations Okay

100 tries The researchers evaluate all 100 attempts They isolate the specific

prompts that successfully broke the defender model's guardrails Those successful attacks are

the worst possible inputs from

a safety

perspective Once they

identify those successful highly evasive prompts

what do they do with them They feed those successful attacks

directly back into the training pipeline They use reinforcement learning to adjust the

mathematical weights of the newer models So they essentially immunize the new model against the exact

tactics

that defeated the old model Exactly This methodology matters immensely because it addresses

a severe vulnerability called fuzzy robustness Yeah Defending an AI is no longer about simply

blocking a static list of prohibited words You

are not just banning the word bomb You are defending against

intelligent adversaries who adapt in

real time

They change their tactics mid conversation This

is terrifyingly relevant in dual use

domains Think about advanced biology or cybersecurity The line between a legitimate research query

and a

dangerous request is incredibly blurry It is A user asking for help optimizing a virus factor could

be a university researcher seeking a vaccine or it could

be a bad actor engineering a bioweapon Defending against

that ambiguity

requires a model

that understands intent not just vocabulary And the

results OpenAI published are highly compelling They state the synthetic training

transferred

successfully to their flagship

models Right The numbers were striking Their new GPT 6 Astra model achieved a 91 5 safe

answer rate when

tested against fixed cybersecurity

jailbreaks For direct

comparison the attacker model itself GPT 5 6 Sol only managed a

59

safe answer rate on those same prompts And crucially OpenAI reports that Astra achieved this massive increase in safety without increasing mistaken refusals

It did not trigger false positives on legitimate requests That

delicate balance is the holy grail of AI alignment You want the system

to act as a titanium vault against malicious attackers but it still

needs to

be

an open hopeful door for legitimate users However

the methodology definitely invites some serious analytical scrutiny I have

a very pointed

question about this entire setup What is

that OpenAI is using

its own proprietary models like the

Sol family to attack its

own defender models Are

they not just shadowboxing What do you mean Well if

you use OpenAI architecture to attack OpenAI architecture aren't you simply training the AI

to recognize its own highly specific blind

spots That is the exact vulnerability independent safety researchers worry

about in these closed loop training setups You risk creating a

model that is perfectly defended against itself

but

totally vulnerable to an outside perspective Exactly How do

we know this defense mechanism actually works

against an entirely

different AI architecture We do not An open source model like LLAMA might generate attacks using a fundamentally different mathematical

structure than a GPT model does It might exploit a vector that open

AI's synthetic data never even

imagined Open AI actually anticipates this

critique They claim the training generalized well They state the newly trained models showed improved resistance against fixed jailbreaks

produced

by entirely different external attacker models They also claim it held up against adaptive human

attackers who constantly shifted their

personas But our skepticism remains warranted until we see rigorous

adversarial testing from independent third parties We need to see these models attacked by systems

trained

on completely different data distributions We also need to highlight the severe limitations of the specific data

open AI shared in

this disclosure The numbers are

impressive but they absolutely require context These high percentages measure the responses

of the raw foundational model layer

This was a highly controlled experiment It

completely excluded open AI's full production safeguards When

you log in to chat GPT as a consumer

you are not just talking to the raw model There are multiple invisible layers of

input and output filters running on top of it There

are secondary models doing nothing but classifying your intent This research

measures the inherent resistance of the base model

itself It is not a deployment wide guarantee of safety Furthermore that 91 5 figure only

represents performance on the specific evaluated prompts they chose to measure It is mathematically

impossible to guarantee total safety against

every possible future attack vector If you are building security tools what you need to watch next is how well this specific defense

mechanism

holds up against domains the researchers

did not explicitly test Will the ASTRA model remain robust when faced with a completely

novel zero day jailbreak strategy that is never seen even synthetically The

arms race between AI generation and AI

defense is accelerating rapidly From defending AI systems against external

attacks we turn to a company defending itself against internal

turmoil NVIDIA employees are pushing back on

their CEO's sudden exit The situation at NVIDIA is escalating quickly Wired reports that over 1 600

comments have been posted on a highly visible internal company suggestion box These comments

are directly urging the board of directors

to bring back Jensen Huang Huang resigned abruptly amid

unresolved personal allegations The board announced his immediate departure

on November 12th And they moved

very swiftly

to fill the

vacuum They

named their current finance chief

Colette Kress as the interim CEO They also appointed Harvey

Jones the former CEO of Synopsys as the new board chair This

internal employee campaign highlights

a massive widening rift regarding corporate process From the external reporting it is abundantly clear that a large segment of the workforce expected the board's formal investigation into these

allegations to actually conclude before any permanent leadership decisions were made Instead Huang exited

the company just as the inquiry was getting off the ground That sequence

of events is exactly what is driving this demand for transparency Employees are

deeply

unsettled by the sheer abruptness of losing a visionary founder Yeah One company wide email which was detailed in the Wired report explicitly framed

Huang's potential return as a matter of basic fairness and mission continuity This internal revolt puts immense uncomfortable pressure on a board of directors that

is largely comprised of early investors and longtime personal confidants of Huang It creates a highly unstable environment for a company

that serves as the absolute bedrock of the entire global AI economy But we need to

apply

some analytical skepticism to the true

scale of this internal pushback Absolutely The headline number is 1

600 comments On the surface that sounds like a massive

wave

of dissent But we have to remember where

that number actually represents It counts

individual comments on an internal forum It does

not count unique employees NVIDIA

emplaced tens of thousands of people globally That is a critical distinction A suggestion box

thread can very

easily be dominated by a few hundred highly passionate highly vocal individuals It could simply be a coordinated faction It

might not be a true statistical representation of the broader silent workforce Still we cannot ignore

the severe backdrop against which all

of this

corporate drama is unfolding The underlying

allegations are incredibly serious They

are Huang publicly admitted to infidelity And completely separate from that admission there is

a distinct

unverified allegation of an inappropriate relationship with a direct

report All out from this uncertainty has been immediate and historic The company has literally lost one trillion

dollars in market

value since these allegations first surfaced

in

the press One trillion dollars That is hard to even comprehend Furthermore Nvidia is now

facing active probes by federal lawmakers and the FBI regarding separate issues So

the

stakes for resolving this leadership vacuum could not possibly be higher What you need to watch next is how the

board responds to this intensifying pressure from both above and below Will the board finally

break their silence to explain the exact

timing of the resignation Or will they justify the incomplete investigation We also need to watch if this

sustained internal pressure actually forces them to reconsider their

decision entirely

Let us move into our quick reads for today