OpenAI pauses new $200 Pro sign-ups

Welcome to The Signal from Superpower Daily with Maya and Theo Right now OpenAI is halting new signups for its 200 Pro tier because well demand for the GPT 6 Astra model is basically breaking their infrastructure Yeah it's a huge capacity crunch Today we're doing a deep dive into a stack of telemetry reports research papers and industry updates to figure out where AI is hitting its physical limits and you know its psychological limits too We have a really fascinating lineup today We're looking at everything from actual server meltdowns in California to and this is wild AI generated employment contracts in the Central African Republic It's a

massive spectrum Okay let's unpack this OpenAI situation first Imagine running a restaurant that's so popular you actually have to lock the doors on people who are waving 200 in your face Right You do this just so your kitchen doesn't catch fire I mean that is exactly what is happening at OpenAI right now They've temporarily stopped new signups and upgrades for the 200 a month ChatGPT Pro plan And just to be clear that is their flagship consumer offering We are talking about the absolute peak of commercially available AI access right now Exactly The core issue here is just unusually strong demand for their newest model which

is GPT 6 Astra The demand has put severe I mean really severe pressure on their infrastructure Yeah they simply lack the compute capacity to let more people into that specific tier right now It's a hardware problem Which is crazy to think about It is Infrastructure is kind of the invisible ceiling of this whole AI revolution We often think of AI as this magical software that just scales infinitely in the cloud Right like it's just code Exactly but it doesn't scale infinitely It lives on physical servers inside physical data centers and it is totally constrained by physical power grids I do want to clarify something for

anyone currently paying for this tier though Existing subscribers are keeping their access Yeah that's an important caveat OpenAI is not kicking anyone out And the lower priced plans remain totally open You can still sign up for the 100 Pro tier or the Plus plan or the Go plan And that targeted closure is actually the most revealing part of this whole story How so Well it's a profound lesson in infrastructure economics for builders and operators We're basically watching a fundamental clash between two different business models here You have fixed price subscriptions right colliding with highly variable compute usage I mean I get the logic of wanting

to protect the servers but isn't this a massive red flag for investors Oh absolutely Like if your most expensive product is the one that's breaking your company how does OpenAI ever scale to enterprise levels It's a completely valid concern Yeah Because you know when you sell a standard software subscription your profit margins are usually incredible Right The cost to serve a second customer is almost zero Just duplicate the code But generative AI breaks that economic rule entirely The revenues fix at 200 The cost to serve that customer however depends entirely on their behavior And the people paying 200 a month are not just asking for

like a recipe for banana bread Far from it These are power users They are pushing the absolute limits of the system They're running really complex agentive workflows Wait let's pause there for a second For listeners who might just use AI for say drafting emails or summarizing meeting notes what does an agentic workflow actually look like in practice That's a great question Are we talking about AI talking to other AI Yes essentially Think of a traditional AI request as a single question and a single answer You ask a question you get a response and the compute stops Okay An agentic workflow though is an AI that

has been given a goal in the autonomy to figure out the steps to achieve it Give me a concrete example Sure Imagine a financial analyst using the GPT 6 Astra model They don't just ask for a summary of a company They give the AI a massive messy data set of 10 years of global supply chain data Wow Okay And they tell the model to identify vulnerabilities write a Python script to visualize those vulnerabilities execute the script debug any errors it encounters along the way and then output a final presentation So the AI is looping It's constantly checking its own work Continually It might make 50

or 100 internal queries before it ever shows the user a single final result That's intense Right And every single one of those internal steps consumes compute power It eats up GPU memory It hogs bandwidth Which totally explains why OpenAI targeted this specific tier I read that Thibault Socio who's a product leader at OpenAI specifically identified this 200 tier as the biggest strain on their systems Yeah Socio and his engineering team rely heavily on telemetry And telemetry is just the automated collection of data regarding the performance of a system right Exactly They're essentially looking at a massive dashboard of server health So they can see exactly

which users are causing the temperature in the data center to spike Figuratively and literally Yeah They looked at their telemetry and saw the bottleneck Closing the heaviest here was the absolute smallest step OpenAI could take to preserve service for everyone else Like surgically removing the heaviest intake valve Precisely They chose service reliability over short term premium revenue I suppose it is better to have a waiting list than a broken product that constantly times out on your most valuable users Oh broken product destroys trust If they let the system degrade to the point where those 200 users can't complete their agentic workflows those users will cancel

anyway Right And they will take their complaints straight to social media Exactly But there is a massive limitation here The opacity of this whole situation is honestly frustrating Yeah We really don't know the scale or duration of this pause OpenAI has not disclosed how many daily sign up attempts they are getting We have no idea how long this gate will remain closed But the opacity is highly intentional isn't it Oh for sure It gives their engineering teams breathing room They need time to provision more servers or optimize the Astra model itself They basically need to find a way to make it run cheaper and faster

Without a ticking clock forcing their hand But this creates incredibly high stakes for current users Which actually brings up a fascinating psychological angle here The lock in effect Yeah this is huge Because if you are a current subscriber to the 200 pro plan and you decide hey I want to save some money this month you cannot cancel Well I mean you can cancel but you cannot simply change your mind and buy the plan again next week You are locked out until the pause lifts Which means people will hold on to their subscriptions out of pure fear Scarcity makes the resource instantly more valuable People who

might have paused their usage will keep paying 200 just so they don't lose their spot in line It artificially suppresses user churn It is an unintended consequence of capacity limits but it certainly helps their revenue retention in the short term So what you need to watch next is how OpenAI resolves this Watch whether they can eventually reopen this tier without immediately recreating the exact same capacity pressure They really have to fundamentally expand their physical infrastructure or they have to drastically reduce the compute cost of Astra before they can open those doors again Exactly Okay that wraps up the OpenAI capacity issue It totally makes sense

that OpenAI is restricting access to managed server capacity but server overload is not the only reason these labs are pulling the plug on accounts right now Next up Anthropic says it removed Claude accounts tied to nine influence operations They found actors from Russia Iran Turkey and other regions using their Claude model to generate deceptive political posts And this represents a massive shift We're moving from managing hardware to managing global security here Anthropic removed accounts linked to operations spanning six continents The most surprising part of this report for me is how these actors were actually using the model We generally picture a troll farm using AI

to churn out thousands of fake tweets or Facebook posts Right but that is the old playbook AI is no longer just a machine for drafting fake social media content It has evolved AI is now acting as back office software for entire influence operations It is functioning as the organizational infrastructure Exactly Look closely at the Central African Republic campaign detailed in the Anthropic report This was a Russian state aligned operation They were not just asking the Claude model to write propaganda articles They were using it like a human resources department which is wild They really were They used Claude to create staffing rubrics They used it

to draft formal employment contracts for local journalists They used it to build repeatable systematic workflows for their entire operation Let's break that down for a second because it's nuts A staffing rubric is a scoring guide used to evaluate job candidates Right You were telling me a Russian influence operation asked an AI to figure out how to hire the best possible propagandist in the Central African Republic That is precisely what happened They likely fed the model their operational goals and they asked it to design a workflow to identify recruit and manage local talent This feels like industrial scale espionage planning but does this not create a

massive digital footprint It does And that is actually a fatal vulnerability for the operators It allows for upstream detection by the AI company Because Anthropic can see the prompts They can see someone asking to build a staffing rubric for a deceptive news site way before any fake news is actually published They can spot the machinery being built inside the model They see the organizational scaffolding being erected long before the deceptive content ever hits a social network I have to push back here though Are these AI campaigns actually going viral I mean we hear all this panic about AI generated disinformation but is it actually working

That is a crucial caveat in the report Anthropic noted that most of the identified content got very little authentic engagement So no one is actually reading this stuff on Twitter The AI is incredibly good at generating volume but volume does not equal virality A million fake tweets shouting into the void do not change an election if real humans ignore them So where did the real reach happen It happened offline or through established channels The real reach only happened when traditional state media distributed the content In the Central African Republic case the content was distributed daily through a local radio station called Radio Lengo Songo It

was amplified on Telegram It was carried by existing local outlets So the AI just fed the traditional old school propaganda machine It didn't replace it Exactly Which brings us to a major limitation for companies like Anthropic the boundary of visibility Right because once that content leaves Claude and goes to a radio station in Bangui Anthropic completely loses sight of it They can cut off the supply of text at the model level They can ban the account But they cannot recall the content once it is printed broadcast or posted on an encrypted app like Telegram There was a specific constraint mentioned in the report that stood

out to me too Claude actually refused to name real people as militants when the human operator asked it to Yeah that is the alignment training working as intended The model has guardrails It refused to generate targeted potentially dangerous accusations against specific living individuals But the human operator just adapted immediately They treated the AI guardrail like a speed bump They just switched to anonymous sourcing It's so typical They asked the model to write the exact same inflammatory content but attributing the claims to unnamed vague sources instead of specific people It really highlights the persistence of human operators They will always try to engineer around the safety

protocols They will And what you should watch next is how AI providers handle this cat and mouse game Watch if they start disclosing how often these early upstream detections actually prevent distribution Right We need to know if spotting the machinery being built actually stops the factory from running Because right now it feels like labs are just mapping out these operations after they have already happened That is forensics We need prevention It is a critical distinction for the future of information integrity Meanwhile a University of Georgia study finds AI interview scoring encouraged exaggeration Oh this connects perfectly to what we just discussed Yeah We're just looking

at humans trying to bypass AI guardrails in influence operations Now we're looking at human candidates trying to game AI recruiters to get a job If you have ever tried to game an automated video interview by smiling a bit too much or throwing in random corporate buzzwords this next study is going to validate all your suspicions The researchers at the University of Georgia studied hundreds of asynchronous video interviews Let's define that for a second An asynchronous video interview is when you log into a portal a prompt appears on screen and your webcam records your answer Right There is no human on the other end It is

just you talking to a blinking light It is an incredibly sterile environment And the study found something pretty disturbing Candidates embellished their qualifications significantly more when they believed an AI was evaluating that recording They lied Or at least they stretched the truth way further than they normally would They presented an idealized highly exaggerated version of their skills And crucially the AI system did not penalize this exaggeration It either rewarded the bragging or just ignored the deception entirely Yep Why does this matter beyond the headline though It really comes down to basic human psychology in high stakes situations Uncertainty drives strategic behavior These candidates felt that

the AI evaluation environment was entirely unpredictable They felt like they were throwing their resume into a black box Exactly They didn't know what the machine actually wanted Did it want them to use specific keywords Did it care about their eye contact Because they did not know the rules they felt they had to overcompensate just to survive the initial screening But this backfired spectacularly when humans were involved It did The researchers contrasted the AI evaluation with human reviewers When actual hiring managers watched these same asynchronous videos they generally detected the deceptive embellishment I mean humans can spot when someone is faking enthusiasm or overstating their role

in a project And the human reviewers penalized it They rated the authentic honest candidates much higher than the ones who were clearly exaggerating So the AI rewarded the deception while the humans punished it The irony here is just incredible The AI system was likely designed as a black box specifically to prevent candidates from gaming the test That is the standard logic in tech design If you hide the criteria people cannot cheat But this study proves the exact opposite That opacity is what encouraged the deceptive behavior in the first place Which leads us to the transparency test This is the most revealing part of the entire

experiment The researchers changed the variables right They took a group of candidates and told them exactly what the AI was measuring Yes they provided the rubric They told the candidates the AI was analyzing facial expressions looking for specific keywords and measuring teamwork indicators and work style And the exaggeration dropped immediately When the candidates knew the rules of the game authenticity returned It returned to levels comparable to when candidates believed a human was reviewing them It's wild By hiding the criteria you do not prevent gaming You simply force candidates to guess how to game the system Transparency actually fostered honesty We do have to note a

significant limitation here though This study only tested one specific AI rating system in an asynchronous setup Right It doesn't mean every single hiring algorithm on the market behaves exactly this way Some might be much better at detecting deception Some candidates might react differently But the psychological principle remains incredibly solid What you should watch next is how employers balance this tension Will they start explaining their automated tools to candidates to encourage authenticity Or will they keep them opaque for perceived security and intellectual property reasons If I'm a hiring manager I want the authentic candidate I would publish the rubric tomorrow It is a difficult balance between

fairness algorithmic effectiveness and corporate secrecy In other news NASA and IBM release an open source lunar AI model for scientific mapping We are shifting from AI evaluating human candidates in an office to AI evaluating the surface of the moon NASA and IBM released something called the NASA IBM Lunar Foundation model It's designed to map lunar craters identify volcanic features and even estimate where polar ice might remain stable And this matters far beyond the headline because of the architecture of what they actually released They didn't just throw a bare set of model ways onto a server and wish scientists good luck They released a massive reusable

research stack Exactly It includes the underlying code the validation benchmarks and full integration with the TerraTorch toolkit We need to define some terms here for a second First what exactly is a foundation model in this context We hear that term used for CHAT GPT which predicts text How does a foundation model look at rocks Well think of a foundation model like giving the AI a master's degree in planetary physics before you even ask it to look at a specific map So it already knows the basic rules of the subject Exactly Instead of teaching a new algorithm from scratch what a crater looks like this model

already understands the general topography of the moon You just have to fine tune it to look for a specific type of crater That makes sense And what is TerraTorch TerraTorch is a specialized software library designed for geospatial AI It makes it much easier for researchers to plug satellite imagery directly into machine learning workflows The data density they use to train this model is staggering by the way Oh it's massive They trained it on roughly 2 million image tiles This includes data from 17 years of continuous observations by the Lunar Reconnaissance Orbiter Plus they integrated gravity data from the DRAIL mission and topographic data from Japan's

CELINE mission Yeah So what does this all actually mean for the next moon mission Like how does this help the Artemis program get humans back to the lunar surface It vastly accelerates mission planning If you want to land humans on the moon you need to find resources Specifically you need to find water ice hidden in permanently shadowed craters at the lunar poles Right because water ice can be converted into drinking water oxygen and rocket fuel And the performance leap of this new model proves the concept works beautifully for finding that The researchers reported up to a 22 lower error rate on estimating polar ice stability

That was compared to a strong baseline model called SWINV2B right Yes SWINV2B is a highly respected vision transformer model It's essentially a state of the art image recognition system The fact that this new lunar foundation model beat it by 22 is a massive jump in accuracy It also identified lunar craters 19 better And it achieved all that while using only half the training data that the baseline model required Which proves that foundation models can vastly reduce the amount of label data needed for highly specialized non text domains But there is a major constraint in space observation that this model still struggles with Orbital lighting Lighting

in space is harsh and completely unforgiving The sun hits the moon at extreme angles especially near the poles Changing shadows can easily hide smaller craters They can completely alter the visual appearance of the terrain The model has a hard time distinguishing real physical surface changes from mere differences in the lighting conditions It might think a crater disappeared when really the sun just moved Resolving that lighting ambiguity is the next great hurdle for orbital computer vision What you should watch next is how the broader planetary science community adapts this open source tool Keep an eye out to see if independent researchers use it to answer entirely

new questions about the moon's geological history that NASA hasn't even thought to ask yet By open sourcing the model they have effectively crowdsourced the scientific discovery process Moving into our Quick Reads Let's do it First up in Quick Reads OpenAI adds GPT 6 Astra and GPT 5 6 controls to chat GPT voice This is a highly requested user interface update It replaces the old instant medium and high intelligence settings that voice users were forced to rely on Now eligible users can specifically select which underlying model they want to use They can also manually adjust the reasoning effort for spoken requests It finally brings the voice

interface into parity with the text interface You actually have control over the engine under the hood But context is crucial here Access to these models via voice is strictly plan dependent Plus users are capped at three hours of daily use Right and the 100 pro plan gets 15 hours of daily voice access to Astra Only the 200 pro plan gets unlimited usage The most important limitation to note is a change in how they handle those limits It no longer silently downgrades you to a mini model when you hit your daily time cap It gives you a hard stop and a warning This is much better

for transparency You always know exactly what level of compute you are getting Next GSA sets an October shift to zero fee usage based OpenAI pricing The General Services Administration is fundamentally changing how the federal government buys AI software The new OneGov agreement drops the 15 monthly platform fee for government buyers using ChatGPT Enterprise Instead of a flat fee the new deal offers a 50 discount on eligible token usage The context here is broad expansion Eligibility for this pricing structure now includes state local and tribal governments But it is vital to understand that a zero platform fee does not mean free AI Agencies must still fund

their own token consumption out of their own local budgets And there is a significant limitation regarding security The Advanced Daybreak Red Cyber model is completely excluded from this 50 discount It remains at standard commercial pricing OpenAI is protecting the margins on their most specialized high value security tools Definitely Finally Pocket fm says AI powers 93 of its catalog as its run rate reaches 500 million The scale of automation happening in audio entertainment right now is just wild Yeah AI now produces 99 of their new content It cuts their production costs by 80 times Producing 100 hours of audio used to take them a full year

Now it takes one single day But the human element remains vital Human creators still supply the core story ideas and the narrative arcs The AI is simply acting as a high speed execution engine And the platform is seeing massive results from this hybrid approach They are reaching 76 revenue retention They also recently launched Pocket Saga That is a U S video app that is entirely AI produced with zero live action production costs And that app is already at a 15 million run rate The limitation however is financial transparency The company claims to be profitable But they refuse to disclose their actual profit margins or their

cash flow numbers So we simply don't know how much of those massive 80 times production savings are actually reaching the bottom line or if they are just burning cash on marketing We are moving to three takeaways from today I'm ready First high end AI demand is currently outpacing physical compute capacity We are seeing this clearly with open AI's infrastructure economics High end models require massive variable compute Companies are being forced to throttle their premium revenue tiers just to protect service stability for existing users Second opaque AI systems encourage humans to behave strategically and deceptively The University of Georgia study proved this When human candidates do

not understand how an AI evaluator works they exaggerate their skills to survive the uncertainty But when transparency is introduced authenticity is restored Third the most lucrative application of generative AI right now is acting as a high speed production engine for human generated ideas Pocket FM's business model proves this By pairing human creativity with AI execution they drastically lower the cost of entertainment at scale It proves the immediate creative return on investment for generative AI We have talked a lot today about AI acting as an HR department for influence operations and AI evaluating human job candidates It leaves you wondering how long until an AI candidate

applies for a job that is evaluated by an AI recruiter with zero human involvement Keep an eye out for that As for what to watch tomorrow keep your eyes open for any new signals from OpenAI regarding their infrastructure expansion timelines We are all waiting to see when that 200 pro tier might finally reopen You find more details on all these stories at superpoweredaily com Thank you for joining us We'll see you tomorrow