OpenAI says new GPT-6 models cost half as much

Welcome to the Signal from Superpower Daily with Maya and Theo OpenAI has just fundamentally reset the cost of intelligence with two new models GPT 6 Sol and Luna They are slashing API prices in half which you know it entirely changes the math for anyone building software today That really is the core reality of what happened this morning OpenAI has officially released GPT 6 Sol and GPT 6 Luna and the company is offering these two new models at half the API price of their GPT 5 6 predecessors They are aggressively targeting operational scale over maximum capability We really need to unpack what a 50 price cut

actually means when you are dealing with machine intelligence at this scale I mean it is massive Right Because if you are a founder or a builder or an operator right now this is not just a modest discount This is a foundational shift in what your business can actually afford to do Absolutely Let us start with the exact numbers because they established a completely new price floor for the industry With GPT 6 Sol model the cost is 2 per million input tokens and it is 10 per million output tokens Which is a significant drop on its own Sol is positioned as the highly capable everyday reasoning

engine but the Luna model is where the underlying economics of the Internet completely change Oh totally Luna is priced at just 0 10 for input and 0 50 for output per million tokens 0 10 I mean let me just pause there because we throw around the word tokens all the time and it can start to sound like arcade money Yeah it really can For you listening let us actually visualize what a million tokens represents In the English language a million tokens is roughly 750 000 words That is a lot of text It is The entire text of the Harry Potter series is about a million

words You could feed this model a massive towering stack of heavy novels or thousands of customer service transcripts or even a decade's worth of corporate legal briefs and it would cost you a single dime to process the input It completely changes what you can afford to automate It really does It shifts the barrier from financial to purely operational But to understand the impact we have to look at the rollout strategy because access right now is strictly segmented It's not for everyone yet It is not just universally available to everyone with a login If you are a paid user on the ChatGPT work tier or the

Codex programming tier you have access to both Sol and Luna right now But if you are on the free or go tiers you only get access to Luna And there is a massive caveat there You only get it through the desktop application That is a very specific choice Yeah And if you are just navigating to the basic web chat interface that hundreds of millions of people use every day you do not have access to either model yet Well that segmentation is highly deliberate You have to look at the underlying strategy OpenAI is expanding downward from their high end GPT 6 Astra model Which is their

powerhouse Right For the past four years the entire artificial intelligence race has been about maximum capability building the absolute smartest most complex most computationally heavy model possible Yeah Everyone wanted to win the benchmark for raw intelligence Exactly Astra fills that role But Sol and Luna are entirely about mass operational scale So here's how I think about it It is exactly like tiered staffing for a major corporation Oh that is a good way to frame it Right You do not hire a senior executive which is Astra in this scenario paying them a massive salary just to have them sit in the mail room and sort incoming

envelopes by zip code No you definitely do not It is a massive waste of resources Precisely If you are running a logistics company you need a few brilliant strategists at the top But what you really need to function is reliable affordable high volume labor to handle repetitive daily tasks And Luna makes it economically viable to automate millions of those micro tasks We're talking about things like formatting raw data feeds or tagging thousands of daily support tickets by priority Or doing basic sentiment analysis on social media mentions Yes exactly These are tasks where you could never justify the computational cost of firing up the flagship Astra

model The math would bankrupt you before you got through your morning data sweep Well to understand how they achieve this price we need to look at the underlying mechanism The training methods for Sol and Luna are actually deeply connected to Astra How so OpenAI is essentially taking the architectural lessons and the high quality synthetic data they generated while building their frontier model and applying them to smaller footprints They are making them faster and cheaper Oh I see This is a process often referred to in the industry as distillation You use the massive brilliant model to teach the smaller faster models how to behave efficiently They

are optimizing for workloads at entirely different scales But that brings us right to the performance claims which is where I have to push back and inject a little bit of skepticism here Fair enough Because OpenAI is not just saying these models are cheaper If they just said hey this is a dumbed down model that costs less that would be one thing But they are making a very specific very aggressive quality case They are They claim both Sol and Luna make half as many factual errors as their GPT 5 6 predecessors Yes the factual accuracy claim is central to their pitch And they are also pointing

to a specific internal benchmark Right the automation bench thing Exactly OpenAI reports that Luna beat GPT 5 6 by 5 4 on automation bench while costing 58 less per task Okay those numbers look absolutely incredible on a glossy presentation slide More accurate faster and vastly cheaper It is the holy grail of software It is the dream scenario But let us be real for a second These are vendor supplied benchmarks This is the equivalent of taking the salesperson's word that the car gets 100 miles to the gallon when driven on a perfectly flat track with a tailwind That is the dry uncomfortable reality of artificial intelligence

benchmarks right now The benchmark environment is sterile It is not the real world No it is not A lower API rate only saves your engineering team money if the model actually resolves the complex task on the first try in the real world Exactly If I am an operator and I swap out my current model for Luna because it is 10 cents but then Luna hallucinates a bad response Or it drops the context halfway through a long document What happens You have to start over Right My system has to catch that error re prompt the model and try again If you have to build in endless

token heavy retries just to get a usable output those cheap 10 cent tokens start adding up very fast They absolutely do And that does not even factor in the most expensive cost of all which is human correction If the model fails entirely and a human employee has to step in to fix the mess you have not saved any money at all You have actually lost time What you are describing is often called the hallucination tax The true cost of a model is never just the API rate printed on the pricing page It is the API rate multiplied by the reliability rate plus the engineering overhead

required to keep it on the rails That makes perfect sense If Luna truly has half the factual errors of GPT 5 6 as they claim that radically changes the math favorably It means the hallucination tax is cut in half But enterprise teams cannot take that on faith They have to validate that in their own messy chaotic data environments So for anyone listening who is actively building on these tools or managing a budget for software development here's what you need to watch next What is that Watch whether enterprise teams actually see these benchmark gains in their real world workflows Do the cost savings hold up when

the model is dealing with misspelled customer emails Or highly specific internal company jargon Or poorly formatted spreadsheets instead of a pristine benchmark test Exactly Also keep a very close eye on when OpenAI finally drops these models into the basic chat interface That will be the true signal that they trust the scale and stability of Sol and Luna for the general public We're going to explicitly close the book on this lead story right here But the implications of Tencent Intelligence are going to echo through every other piece of news we cover today Next up Meta is aggressively defending the origins of its new desktop agent This

is a fascinating story It really is The company maintains its muse Personal AI was built entirely from scratch despite officially acknowledging it was heavily inspired by the open source project OpenClaw The details of how this entire controversy came to light are absolutely incredible It was not a corporate leak Users out in the wild essentially caught Meta red handed Inspin and developers started noticing that Meta's new muse agent was using matching workspace filenames But the filenames were just the beginning The real smoking gun that set the internet on fire was a very specific configuration file called soll md Yes the SOL file Users digging into the

local files on their machines found that the soll md file in Meta's was the one found in the open source OpenClaw repository Line for line in some places Exactly Line for line For those who might not spend their days digging through GitHub repositories or poking around hidden software files let us break down what sol md actually is It is essentially the artificial brain's instruction manual That is a great way to put it It is a plain text file that dictates the agent's entire personality and operational parameters It defines its communication style its core values and its ethical boundaries It even dictates how it should react

to frustration and its specific areas of expertise In technical terms it is the foundational prompt architecture The raw language model itself is somewhat of a blank slate Right It does not know who it is supposed to be yet Exactly The sol file is what tells that raw model how to behave as a highly specific personalized product It acts as the persistent character sheet So when users pointed out these undeniable textual similarities the response from Meta's leadership was quite revealing because they did not deny it No they leaned right into it Nat Friedman who leads the Meta Superintelligence Labs actually admitted the inspiration directly in a

public response He did He stated unequivocally that they built Muse from scratch from the ground up But he openly acknowledged it was heavily inspired as a product experience by OpenClaw And he went further than just acknowledging inspiration He explained that Meta employees had been actively using OpenClaw internally to manage their own workflows Right In fact Friedman said that after he personally used OpenClaw back in January he was so impressed that he went out and bought hundreds of Mac minis for his engineering team Just to run this open source tool Yeah The explicit goal was to take that open source concept study it and build a

safer easier to use version one that could scale not just to thousands of developers but to billions of everyday consumers He even praised the original developer He said that the OpenClaw creator Peter Steinberger had gotten those workspace elements and that specific personality file exactly right Meta did not dispute the copy text at all No They essentially argued that they adopted it as a product convention almost like adopting the layout of a steering wheel in a car But I mean this raises an incredibly obvious and frankly skeptical question I have to ask it Go ahead If you copy the exact configuration files that tell the artificial

intelligence who to be how to act and what its personality is isn't that just copying the product It definitely looks that way to a lot of people Right How can you say it is built from scratch when the soul of the machine is copy pasted Well that is the crux of the debate And this is where we have to fundamentally distinguish between prompt engineering and systems engineering Okay Unpack that for us The soul md file while deeply important for the vibe of the product is essentially just an elaborate text prompt Meta's defense is that they did not copy the underlying back end code the actual

complex software engineering that makes the agent function securely and reliably on a desktop operating system So let me see if I have this right They copied the personality the script the actor is reading from but they built their own stage their own theater and their own nervous system to support the actor That is a very accurate way to look at it Meta's core argument is that the architecture surrounding the model is the actual product they built Because Muse does a lot more than just chat Right Muse launched recently to handle deep complex personal tasks directly on your computer We're not just talking about answering trivia

questions We are talking about an agent that can open your email client read your messages and draft replies It can navigate travel websites to book flights and even execute purchases And to do that the agent needs profound system level access to your digital life It cannot just live in an isolated browser window like a standard chatbot It needs to see what you see Exactly And because of that terrifyingly deep level of access Meta built a dedicated secure virtual machine This is the engineering they are proud of How's that VM work Muse runs entirely inside this isolated sandbox It keeps the core agent and your raw

personal data separated from the rest of your computer's operating system So that if the agent behaves unpredictably it cannot easily corrupt your hard drive They also built a secondary layer of security called a Sentinel agent right How does that actually work in practice The Sentinel agent is fascinating It acts as a distinct entirely separate security guard It does not do any tasks It just watches Yes Instead it constantly reviews every single action the primary Muse agent tries to take that requires internet access or system changes If Muse decides it needs to send an email or transfer a file the Sentinel agent audits that request against

a strict set of safety rules Oh wow And it explicitly pops up a window to request your human permission when necessary That dual agent architecture one agent acting and a second independent agent auditing is what Meta claims is the real proprietary engineering here So they're saying the secure framework is the innovation not the text file that tells the bot to be polite Precisely That makes sense from an engineering perspective But there's a vital limitation here regarding privacy that we cannot gloss over Meta notes that users have granular control Right You can manage the settings You can choose exactly which apps to connect to Muse and

you can adjust access permissions at any time Importantly you can also completely opt out of having your personal interactions and your local data used to train Meta's future AI models However the fundamental reality of a personal action taking agent is that it requires vast invasive access to be useful at all Yeah it has to look at your stuff You can opt out of the corporate training data collection yes but to function locally on your machine Muse still has to have its hands deep in your personal files your private communication tools Right The privacy risk is inherent to the product category itself If it cannot see

your data it cannot help you Which means the thing to watch next is precisely that dual agent security architecture We need to watch whether Meta's secure virtual machine and that auditing Sentinel agent actually prevent the kind of agent hijacking that has plagued earlier open source tools We really do Can a malicious website trick Muse into handing over your files We will see if Meta's massive engineering budget holds up better than the open source product they drew their initial inspiration from Meanwhile in physical automation Toyota has revealed exactly how its factory workers are training the next generation of machines This is wild The automaker is having

its human employees physically teach a two handed humanoid robot named Elay This story is incredible because it completely flips our usual understanding of how automation works Usually when we think of robotics we picture a highly trained engineer sitting at a keyboard in an office writing thousands of lines of code to tell a robot exactly where to move its arm on an X Y and Z axis That is the standard model yeah But Toyota is currently gearing up to spend 6 42 billion annually starting in 2028 on a massive sweeping factory upgrade And they are throwing out the keyboard The scale of this investment is staggering

The target they have set is to deploy 400 000 robots 400 000 That is huge It is That deployment breaks down to 150 000 machines directly inside Toyota's core automotive plants and another 250 000 spread across their vast network of group companies that manufacture the underlying components and raw materials And the way they are training the humanoid robots the ones called ELY is where this gets almost science fiction Oh absolutely They are using their 18 000 veteran human workers These are people who have been on the assembly line for decades These workers are using special custom built finger shaped jigs to physically grab the robot's

hands and guide it through the precise movements required to build a car The mechanism here is known as kinesthetic teaching Instead of writing code to approximate a movement the human operator physically backdrives the robot's motors So they are literally moving the robot like a puppet Yes The robot's internal sensors record the exact force the torque the spatial positioning and the subtle angle adjustments of the human's movement They are capturing the invisible expertise of a master mechanic That is brilliant This approach is deeply rooted in the philosophy of the Toyota production system specifically the concept of kaizen or continuous improvement Executive Vice President Hiroki Nakajima has

championed this specific idea of coexistence between human and machine They are emphatically not talking about a sudden sweeping replacement of the workforce No They are keeping the human veteran entirely in the loop The philosophy is that the human worker establishes the absolute baseline of quality through their physical movement Right Because they know exactly how a door hinge should feel when it slots into place Exactly They teach that feeling to the robot and only then is that movement automated It is a stunning visual to imagine It really is You have these highly experienced mechanics standing on the factory floor in their uniforms holding metal jigs physically

teaching a wheeled metal robot how to use its mechanical hands to manipulate tools It is the physical manifestation of machine learning If you think about large language models like we discussed with OpenAI they scrape text from the internet to learn how to write In this case Toyota is essentially scraping human motor skills muscle memory and tactile intuition straight from their best employees The data source is human movement I love the idea of capturing physical intuition But we need to inject a very necessary very grounded reality check into this narrative right now I do Because the framing here is a little bit deceptive Yes The limitation

here is hidden inside the headline numbers 400 000 is the total number of factory robots planned across the entire Toyota group over the coming years Right The headlines make it sound like a standing army of 400 000 highly dexterous two handed humanoid robots is about to march onto the factory floors tomorrow It conjures up images of C3PO building a Camry That is the PR spin But Toyota has completely refused to disclose how many of those planned machines will actually be complex humanoids like Elay Ah So there is the catch Given the massive cost and fragility of humanoid robotics today the vast majority of that 6

billion annual budget will almost certainly go toward traditional single task industrial robotics So just the standard stuff We are talking about the massive bolted down robotic arms that do heavy spot welding high speed painting and move heavy chassis parts exactly as they have done for decades So what we are really looking at is a huge traditional industrial automation push combined with a fascinating but likely much smaller pilot program for these new humanoids But from a corporate messaging standpoint they are blurring the math between the two to sound incredibly futuristic Exactly What you need to watch next is corporate transparency Watch for Toyota to finally disclose

the actual ratio of humanoid robots to traditional industrial robotic arms in this massive capital expenditure Yeah They are buying 399 000 traditional arms and 1 000 humanoids Only when they release that breakdown will we know if humanoids are just an expensive factory experiment or if they are becoming a true manufacturing standard In agricultural news we are moving from the controlled environment of the factory floor out to the messy unpredictable reality of global farming This is a major shift in focus It is Google and the Gates Foundation are vastly expanding their AI ambitions for the global South The partners are directing over 100 million toward AI

tools specifically aimed at 200 million smallholder farmers The scale of this initiative has grown significantly since it was first piloted Their target reach has quadrupled Wow They are jumping from a goal of 50 million farmers to a massive target of 200 million farmers across sub Saharan Africa and South Asia The statistics behind smallholder farms which are typically defined as farms smaller than 2 hectares are absolutely staggering Most people in the West do not realize this but these small family run plots produce nearly 35 of the entire world's food supply That is incredible They are spread across more than 500 million individual farms globally Giving these

farmers access to highly localized farm level climate forecasts could literally change the math on global food security and prevent regional famines The initiative is designed to be comprehensive funding several different layers of technology simultaneously On one layer they are integrating AI driven forecasting models into a climate platform called TomorrowNow Okay so that is the weather side Right On another layer they are actively funding local on the ground innovators They are specifically directing capital to organizations like Wadwani AI and Digital Green in India empowering them to build the front end tools That makes sense And on a third layer they are backing deep agricultural research through

CGIAR to develop new crop varieties that can withstand the increasing threats of drought extreme heat and novel diseases But what really caught my eye the detail that shows they actually understand the problem is how they are handling language Oh this part is brilliant Because historically Silicon Valley just builds a tool in English slaps a Google Translate widget on top of it and calls it global But that does not work for complex agricultural advice It is a brilliant architectural choice The partnership is actively funding the creation of open source speech and text datasets for more than 40 distinct African languages Which means they are building native

agricultural AI If a farmer in rural Kenya asks a complex question about soil acidity in their local dialect the artificial intelligence model actually processes and understands the nuance of that specific language natively Right It does not have to translate the local dialect into English Try to understand the English generate an English answer and then translate it back That translation layer is where all the crucial local context gets lost By building native models it allows the technology to work seamlessly through established local institutions and trusted regional networks It also fundamentally addresses the issue of data sovereignty By funding open source datasets they are establishing data governance

that keeps the intellectual property the training data and the actual technological capability rooted in the regions where the tools are actually being used That is so important It is a bottom up approach to data infrastructure rather than extracting all the local knowledge back to a server in California to enrich an American tech giant The caveat to all of this incredible ambition is the brutal reality of time This is a multi year promise backed by 100 million of intent But right now it is not yet a proven impact on actual farms Very true The native language tools have to be painstakingly built they have to be

deployed to rural areas with intermittent connectivity and most importantly they have to be trusted by farmers whose entire livelihoods depend on the accuracy of the advice Which leads exactly to what we should watch next in this space Watch to see if these high tech weather forecasts and native language models actually translate into timely physical decisions in the dirt Yeah It's one thing to deliver an alert Can a farmer get an alert on their phone trust that the alert is accurate and secure the resources to change their planting schedule in time to actually save a crop And just as importantly from an institutional level watch if

the local organizations receiving this funding can sustain the complex data governance and technical infrastructure over the next decade long after the initial press releases fade We are moving into our Quick Read section now We have three sharp updates across the industry First SpaceX AI says their new Grok bot absorbed a 175 surge in customer support tickets without requiring a single new support hire I have to admit whenever a company reports massive deflection metrics like this I immediately get skeptical I wonder if the bot is actually solving complex customer problems or if it is just firing off generic automated FAQ replies that frustrate customers so badly

they eventually just give up and close the ticket That is the usual playbook but SpaceX AI claims the underlying mechanism here is much more than a simple reply generator They describe Grok bot as a comprehensive operating layer for their business How does it work When a ticket comes in it does not just read the text It conducts a full pre investigation It instantly checks for known engineering issues in their linear project management software It scans for back end server errors using Datadog Wow So it is actually investigating Yes If a user reports a UI glitch it can actually reproduce the bug and record a video

of the failure for the engineering team Most impressively it resolves refund requests autonomously 99 of the time based on strictly defined financial instructions And what does that cost them They report this entire automated process costs just 0 20 to 0 30 per ticket The limitation of course is that these are entirely company reported cherry picked metrics Achieving a 99 success rate on refunds is relatively easy when the policy instructions are mathematically black and white If the system sees a specific error code it issues the refund But watch how this level of autonomous resolution holds up when the bot encounters harder multi step edge cases What

happens when a user has a complex hardware failure combined with a billing dispute that requires actual human judgment and empathy Next up Google is providing 100 000 AI Skills Certificate Scholarships across more than 80 countries They are doing this in partnership through the ITU's AI Skills Coalition Delivering educational training on a global scale is incredibly difficult to get right It usually ends up being too generic to be useful The delivery method here is the critical component Google is not just pushing a uniform top down global curriculum from Mountain View The coalition works directly through local governments and the ITU UNICEF Giga Network This means local

ministries of education and labor can choose from a massive catalog of more than 200 different courses and align the training specifically with their own local workforce priorities and economic needs That localization is incredibly smart It tailors the tool to the environment But we have to look at the massive underlying gap in demand here The ITU themselves estimate that roughly 1 2 billion people worldwide currently need basic digital skills training to participate in the modern economy It is a massive deficit Against a deficit of over a billion people 100 000 scholarships is practically a drop in the ocean It is a pilot program not a global

solution The measurement data is what you really need to watch next Do these individual learners actually complete the rigorous training and use it to meaningfully improve real world public services and gain new employment Or does the program just boost their self reported confidence with a few new digital tools Finally renowned security researcher Patrick Wardle has found a zero day flaw in Meta's macOS Muse Assistant that can successfully capture a user's sensitive authentication token An exposed authentication token for an AI agent that explicitly reads your private email manages your schedule and accesses your file system sounds absolutely disastrous For anyone not in cybersecurity an authentication token

is basically the digital VIP pass that proves you are who you say you are letting the software bypass the password screen If someone steals that token they become you Exactly Wardle dug into the architecture and found that an attacker can exploit an undocumented setting within the Muse app This specific setting controls where Muse routes your local voice prompts to be transcribed into text So what happens if they manipulate it By manipulating that setting and pointing the routing to a malicious external server the attacker can silently intercept your raw audio prompts and crucially steal the authentication token attached to the request But before everyone panics and

uninstalls the software we need to calm the alarm slightly This is not a remote break in vulnerability Right An attacker sitting in another country cannot just hack into your clean secure computer over the open internet using this specific flaw Precisely To execute this redirect the attacker must already have malware running on your Mac Or they must have successfully phished you into downloading a malicious third party application So they need access first Yes They need an existing active local foothold on your machine to change that setting in the first place But the danger is that because Muse requires such deep pervasive access to your WhatsApp messages

your cameras and your local files just to function that relatively minor local foothold suddenly escalates into an incredibly dangerous system wide breach Watch how quickly Meta isolates these sensitive routing settings from other unprivileged local apps on the machine to permanently close this loophole We are moving to three takeaways from today First the absolute cost of baseline machine intelligence is collapsing in real time Between OpenAI launching the Luna model at just 0 10 per million tokens and SpaceX AI resolving complex customer support tickets autonomously for a quarter the barrier to mass enterprise automation is now purely an operational challenge not a financial one Second the concept

of the human in the loop is shifting dynamically from sitting behind a software keyboard to standing on the physical factory floor Toyota's veteran workers physically kinesthetically teaching humanoid robots prove that the next massive frontier of training data is not scraped internet text but physical expert human movement Third as highly capable AI agents like Meta's Muse gain the necessary ability to act across our intimate personal ecosystems the traditional digital security perimeter is dissolving It is shifting directly onto the agent itself requiring entirely new architectures like dual agent auditing to keep our data safe Watch tomorrow how rival model builders companies like Anthropic and Google react over

the next 24 hours to OpenAI effectively establishing a new rock bottom price floor for developer APIs As we look at intelligence becoming practically free it leaves you with a fascinating question When raw compute and reasoning cost pennies what becomes the new scarce resource in business You can find links to all these stories and deeper analysis at superpowerdaily com Thank you for listening and we'll see you tomorrow