OpenAI adds agents that work between chats

Welcome to The Signal from Superpower Daily with Maya and Theo Yeah great to be here So imagine closing your laptop on a Friday and over the weekend a digital version of you just keeps working Right It reviews your code Yeah It attends to bug reports It even drafts your budget That is no longer sci fi No it is not Because as of opening eyes Dev Day 2026 it is a literal product The announcements out of San Francisco fundamentally change everything I mean they change how we interact with these systems entirely Which is why we are making it our lead story today We really need to

spend time unpacking everything that happened on that stage at Fort Mason Because we are moving from asking isolated questions to delegating persistent responsibilities It is a massive shift Oh absolutely Software is no longer just a tool It is becoming a co worker So let's start with the centerpiece of the entire event Right The new product called Dots Exactly Dots These are personal agents that simply do not sleep They remain active between your conversations Yeah They have their own dedicated cloud computers They even have their own web browsers Which is wild It is a structural evolution in consumer AI Think about the past four years We

have been completely trapped in a prompt and wait loop You ask a question the AI types an answer and you're done Right The interaction ends But with Dots you don't just ask for a recipe or a quick code snippet You actually hand off an ongoing project Okay So give me a practical example of that What does that handoff look like for someone working a normal desk job Well let's say you manage a software product You can spin up a doc You assign it to monitor your team's bug reports in the background Okay While you are sitting in meetings that dot is actively reading incoming tickets

It is cross referencing those tickets with your existing code base Wow And it is even preparing a preliminary budget cycle It estimates the engineering time required to fix those bugs That is incredible And OpeningEye says these agents can work across more than 4 000 connected applications Right The level of access is just unprecedented The agent can investigate dependencies across your entire tech stack It works through the necessary changes and then it just returns a pull request or a summary for human review Yeah And for anyone outside of software engineering a pull request is basically a proposed change A human manager has to approve it or

deny it Exactly It feels exactly like managing a junior employee It really does The underlying theme here is continuity OpenAI really wants the agent to carry context across your devices The work has to continue after you close your laptop The AI is no longer just a static encyclopedia waiting for your query Right It is an active participant in your workflow now But practically speaking I have a question about this Where does all this ongoing work actually live That is the big question Because if I close my chat window I lose my visual connection to the AI And that brings us to their second major announcement

OpenAI launched ChatGPT Space Yes Space is designed to solve that exact friction It is a shared collaborative workspace And it basically replaces the old library feature for eligible users It gives humans and agents a shared digital home This home holds living pages It holds your files And these pages can contain writing right Writing research data visualizations and executable code All in one place So it's kind of like a Google Doc But the AI actually has a cursor too That is exactly it Teammates can edit these pages together You can leave comments for each other Yeah But you can also mention ChatGPT directly Or you can

tag a specific dot to request changes right inside the document The most fascinating part to me is that these pages remain current I mean think about how we use AI right now It is very static Right You ask for a summary of a weekly metrics report A week later you have to ask for a totally new summary Yeah And then you have to copy and paste that new summary into a document for your boss It is just this highly manual copy paste routine Well that routine is completely dead in space Because you can just ask a dot to monitor an approved data source Like a

live sales database for example Okay The dot updates the page automatically as the data changes So space becomes this living destination for your ongoing work It is no longer just a storage folder for finished chat outputs But we really have to talk about the infrastructure required to pull this off Yes the compute Because that level of constant background computation requires immense processing power I mean if an AI is constantly checking a database and updating a document it is consuming server resources every single second It is incredibly resource intensive Which brings us to the new model announcement OpenAI released GPT 6 1 Sol Right Sol This

model is specifically optimized for complex coding It handles all that professional background work The capabilities they showed on stage were definitely impressive OpenAI reported strong benchmark improvements on difficult coding tasks But the real story here is not the benchmarks The real story is the economics of it all Right The pricing structure of this specific model is the only thing that makes these always on agents viable in the real world So I want to break down those prices because this really gets to the core of how AI businesses actually survive Let's do it The standard API rate for Sol is 2 per million input tokens It

is 10 per million output tokens And just for context a token is roughly three quarters of a typical word Right Thank you But the most important number OpenAI announced is the cached input price Yes this is crucial Cached input is remarkably low That 0 10 figure is literally the quiet infrastructure revolution of DubDay It really is It represents a 95 discount compared to standard input pricing And we really need to explain why this matters so much Please do Agentic loops require an AI to constantly re read its own context OK wait Let's explain what an agentic loop is For someone listening who maybe isn't building

these tools every day Imagine you give an AI a complex 10 step task To complete step 2 the AI has to remember exactly what it did in step 1 To complete step 3 it has to review steps 1 and 2 Oh I see It is constantly talking to itself It is continuously evaluating its own progress against your original prompt That process is an agentic loop So it essentially has to keep re reading the entire instruction manual every single time it makes a move Precisely And without context caching the AI has to re read the entire history of the project from scratch For every single step

That sounds exhausting I think of it like reading a 500 page book If you have to start at page 1 every single time you want to read a new chapter it takes forever And it uses immense cognitive energy Right Context caching is like just leaving the book open on your desk The AI just glances down at the page And before this update developers were paying full price every single time the AI re read that 500 page book Exactly If an agent loops 100 times to solve a coding problem you were paying for millions of input tokens just to keep the AI's memory fresh Yes And

cheap cash reads make this entire process financially sustainable Because at a 95 discount a developer can actually afford to let an agent think through a complex problem You can do it without going bankrupt Exactly But ongoing background work is just one piece of the puzzle You set a dot to work on Friday and you don't really care if it takes two days to finish Right It's just happening in the background But sometimes you need an answer immediately Which is why OpenAI demonstrated a new premium speed tier called Ultrafast Ultrafast Yeah this trades a significantly higher price for a lot less waiting OpenAI demonstrated this using

the existing Astra model They have the model build and launch a small virtual rocket scene right on stage And the Ultrafast version completed the task visibly earlier than the standard version It did OpenAI advertises up to eight times faster token generation in Codex which is their coding specific environment Eight times faster Yeah that reaches 300 tokens per second And it is up to six times faster in the general API But I do want to push back on what faster actually means here Oh absolutely Because we must be precise Faster token generation does not mean an entire coding job finishes eight times faster That is a

crucial distinction to make The AI is essentially just typing faster Right But it still has to perform external actions It still has to browse the web for documentation It still has to compile the code It still has to run automated tests to see if the code actually works And those external steps take physical time You cannot speed up the time it takes a server to compile a heavy application It does not matter how fast the AI writes the script The token speed only accelerates the thinking and the writing phases The execution phase is still strictly bound by the laws of traditional computing Right And to

access this accelerated thinking phase OpenAI introduced a massive new subscription tier Pro 500 Yes Pro 500 As the name suggests it costs 500 per month This plan offers 25 times the usage of the standard plus plan And it includes full access to that ultra fast mode So let's shift to how developers are actually meant to use all this power Because DevDay focused heavily on developer workflows They really did OpenAI rolled out a feature called Codex Cloud This creates reusable developer environments Previously if you wanted an AI to help you build something complex you had to spend hours giving it context You had to explain your

file structure You had to list your dependencies You had to explain your custom internal libraries Which is incredibly tedious Very With Codex Cloud you configure an environment once You include your repositories You include your approved credentials Then you can start or continue work from literally any device Each task gets its own dedicated pre configured workspace Exactly You just don't have to rebuild your setup for a single short task anymore It is ready to go instantly They also heavily expanded the Codex command line interface Developers now have these granular controls for parallel work You can inspect and switch between multiple running tasks You can fork work

to branch it with its existing context They even added cloud based vulnerability scanning It is a system called Codex Security Cloud That security scanning uses what they call cyber capable Daybreak Blue models Right It basically offers continuous checks on the code the agent is writing And it actually proposes fixes for vulnerabilities But here's where we really need to inject a heavy dose of reality into this I agree completely Because a verified proposed patch is not an automatically deployed fix A human must still review it And this brings me to my major caveat regarding this entire Dev Day presentation You are looking at the friction The

difference between the polished stage presentation and the messy reality of daily use I absolutely am Because the main limitation we have to watch here is access and reliability But let's look at GPT 6 1 Sol OpenAI explicitly stated it is not in regular consumer chat at launch Right It is only available in the API chat GPT work and Codex A flashy model announcement on a stage does not mean it is everywhere on your phone immediately And the exact same rollout friction applies to DOTS Yeah The initial release only covers pro and business premium users in eligible markets But for enterprise education and healthcare sectors access

is strictly a beta IT administrators must actively go in and enable it It is off by default for their biggest clients Furthermore live stage demos are not daily reality We saw genuine really uncomfortable friction during the keynote We did A product team member named Holly experienced a noticeable voice exchange delay during her demo She was just waiting awkwardly for the AI to respond Yeah that was tough to watch And later another presenter named Romain experienced a complete command line voice failure He actually had to stop talking and switch to typing just to get the demo to work Those unscripted moments are vital reminders especially for

anyone trying to build a real business on this technology Right The cost per token means absolutely nothing if the cost per accepted result remains high Say more about that Break that down If an agent hallucinates a block of code you still pay for the tokens it generated Yeah Then you have to pay a human engineer for the time they spend reviewing the code and finding the error Finally you pay for the additional prompts and tokens required to fix that error That adds up fast Endless human correction completely destroys the entire economic advantage of cheap compute If you have to babysit the dot constantly it is

not a co worker It is a burden That is the exact metric every CTO needs to track this year Cost per accepted result Finally we need to watch how these tools actually reach end users inside large companies Because OpenAI announced massive changes to their product distribution model too They launched the OpenAI marketplace Right The marketplace This completely changes business to business AI distribution Enterprise customers can now spend their existing OpenAI financial commitments on third party products The keynote named 32 launch partners These include major established tools companies like CodeRabbit and Notion Yeah So if an enterprise already has a massive million dollar contract with OpenAI

they can allocate some of that existing budget directly to these partner apps It brilliantly lowers the barrier to entry Typically if a team wants to try a new AI tool they have to clear a whole separate vendor procurement process Which is a nightmare They have to get legal approval They have to get finance to sign off on a brand new budget line Yeah But now a customer who already pays for eligible OpenAI usage can simply bring their budget along The procurement hurdle is completely bypassed That is a masterclass in locking in enterprise clients Seriously So we are officially closing the book on the Dev Day

lead story That covers the major keynote announcements Yeah but those agentic workflows are really the foundation of everything else we covered today Right Because OpenAI's pricing structure isn't the only thing taking AI out of the chat box Next up we are looking at how OpenAI actually charges its power users OpenAI will reopen 200 pro signups on September 30th with completely revised usage terms This is a really fascinating shift in subscription economics OpenAI had previously paused new signups for this specific pro tier simply because they couldn't handle the raw compute demand Yeah Now they are bringing it back But the underlying math of what you are

actually buying is entirely different Thibaut Sotiat announced the changes The monthly price stays exactly the same It is still 200 Okay But the weekly allocation is now valued at half the API spend of the old plan Now on the surface if you are a user that sounds like a massive reduction in value It sounds terrible You are getting half the dollar equivalent for the exact same price But we really have to look at the structural change behind that number Right Because OpenAI is completely removing the rigid five hour usage limit Under the old plan you were forced to pace yourself You had to carefully watch

the clock to avoid hitting a timeout wall every five hours Exactly If you were in the middle of a massive coding sprint the AI would just cut you off But the new plan gives you a weekly allocation to spend however you want This marks a structural shift from time based access to a bucket of compute A bucket of compute Yeah It gives operators immense flexibility You can now burn your entire weekly allocation over a single intense weekend coding sprint Which a lot of developers do Right You are no longer tethered to a rolling hourly limit that interrupts your flow state Now Santras argues that subscribers

will actually get more work done than they could a month ago He claims a lot of new highly efficient compute is coming online Yeah But I have some serious pushback here on behalf of developers Because the math is incredibly opaque You are struggling to quantify the actual utility of this new compute bucket If I'm a freelance developer and I use this tool to pay my rent saying a plan nets out at half the dollar value in API spending tells me absolutely nothing about my practical output Right It doesn't tell me how many distinct tasks I can complete It doesn't provide a clear dollar denominated allowance

I can actually track in a dashboard No it doesn't It offers no baseline estimate for different kinds of workloads I mean I can't build a business model around an abstraction It is exceptionally difficult for a small business to budget when the primary resource metric is hidden behind a vague weekly bucket So I have to question if this is simply a downgrade disguised as user flexibility OpenAI clearly knows this change will be jarring for their power users Oh absolutely They are allowing existing subscribers to keep a 20x multiplier on their compute for a while That temporary grace period exists purely to soften the blow of a

massive capability reduction It also signals OpenAI's long term corporate strategy They really want to narrow the pricing gap between flat rate consumer subscriptions and raw API pay as you go costs Yeah Historically power users on subscriptions were wildly unprofitable because they consumed far more compute than their flat fee actually covered So OpenAI is trying to align consumer usage more closely with the actual back end server costs Which makes perfect business sense for their profit margins But it leaves prospective subscribers guessing whether the 200 is actually worth it It does However there is one major piece to this puzzle that we don't have yet We need

to watch what happens tomorrow Right OpenAI promised to release new non usage based features specifically for this tier on September 30th And these features will supposedly not consume your weekly compute allocation The problem is they have not named these features yet We have no idea what they are Right We cannot accurately judge the true value of the revised 200 plan until we see what these additions actually are Exactly If they are just minor UI tweaks there will be a revolt If they offer unlimited access to a specific lower tier model well that changes the math entirely Those hidden features will absolutely be the deciding factor

Especially for independent developers wondering if they should just downgrade to the Standard Plus plan Meanwhile we are shifting our focus entirely While software developers are figuring out their compute budgets researchers are applying these exact same models to human biology Microsoft introduces Quine an AI biology system tested on tumor cell shifts This story is incredible It takes AI out of the text box and puts it directly into the laboratory Microsoft Research collaborated with the Broad Institute to tackle a very specific problem They focused on pancreatic ductal adenocarcinoma This is an incredibly complex aggressive and highly lethal form of cancer The researchers wanted to know if tumor

cells could be intentionally shifted between different biological states So they used an AI system named Quine to predict how thousands of different chemical compounds might actually affect these cancer cells And the speed of this process is what makes it a massive breakthrough It was so fast Over a single weekend Quine ranked thousands of compounds It evaluated their molecular properties and narrowed the massive search space to a handful of high probability candidates for physical testing Physical lab tests then confirmed Quine's predictions The selected compounds successfully shifted tumor cells from a state known as classical to a state known as basal We really need to explain why

this matters Quine is acting as a multimodal world model for biology When we say multimodal in the context of chat GPT we just mean it can read text and look at photos But when we say multimodal in biology it is entirely different It is processing massive disparate scientific data sets Right It's not just summarizing Wikipedia articles about oncology No not at all It actually analyzes raw genetic sequences It maps three dimensional protein structures It evaluates high resolution cellular images It combines all of this raw data with existing scientific literature and direct researcher input Most fascinatingly Quine revealed something the human researchers completely missed This blew

my mind The original scientific framework assumed a simple two state system The researchers believed the cells were either in the classical state or the basal state Like a light switch Exactly It was a binary toggle But during the AI guided experiments the cells moved toward a distinct third state Wow Quine had actually predicted this complex response The AI spotted biological complexities that humans had not anticipated in their initial framing of the experiment Because several of the compounds produced strong effects through biological mechanisms the researchers had not even considered Right This is the holy grail of AI and science It isn't just automating what we already

know It is creating entirely new leads for follow up research It proves AI can model biological reality with shocking nuance This is incredible research But I must stress a massive caveat here for anyone listening who has a personal connection to this disease Yes please I want to use an analogy Quine is not a doctor prescribing a cure Quine is a librarian A highly efficient biologically literate librarian Yes exactly It is a librarian pointing a scientist to the exact right haystack to find a hidden needle Right This is an experimental research workflow It is designed to radically narrow down the number of lab tests humans have

to run It is absolutely not clinical evidence of a cancer treatment The experiments do not show that any of these compounds can treat cancer in a living human patient They only show that the compounds alter cell states in a petri dish Moving from a petri dish to a safe effective human trial takes years and it involves entirely different testing protocols Exactly We are looking at early stage discovery not a finished drug Listeners interested in this space should definitely watch for Microsoft Discovery's broader rollout Because access to Quine is currently very limited It is strictly restricted to Quine fellows and select academic collaborations You cannot just

log in and start testing compounds yourself Microsoft does plan to expand access as the system matures But the next major test to watch is the integration of new RNA data into the model That will be huge We need to see if adding even more complex biological data and rigorous confidence estimates can help Quine untangle deeper cellular mysteries without hallucinating false leads Scientists will still need to meticulously review and physically validate every single suggestion that AI makes in a real lab setting So in other news we are moving from a research assistant in a Microsoft lab to a digital assistant living right on your local machine

Manus 2 0 adds editable games message triggered tasks and remote computer use The Manus release is wild It combines hands on creative generation with deep dangerous system automations Yeah dangerous is the right word It shifts the agent far beyond a simple one shot generation tool Let's start with what actually launched on the creative side first The desktop app has been completely rebranded as Manus Studio It now includes dedicated game dev and video editor environments Previously if an AI generated a game for you and you didn't like the jump mechanics you had a huge problem Oh yeah You had to write a brand new prompt and

just hope the AI understood what to fix without breaking everything else It rarely did Right Now you can manually revise the AI generated games and videos directly in the interface You don't have to accept the first version Manus also states that their dedicated cloud computers can keep multiplayer game servers running for these generated games But I have to point out that hosting a game server doesn't mean the generated game is actually fun to play The fact that they built a whole studio for editing proves that initial AI generation still requires heavy human revision to be usable It does But the more significant update and the

one that really changes how we work lies in their new automation features Manus can now trigger a complex task based on an incoming event For example a specific type of email or a designated Slack message can automatically start an agent workflow And this breaks the traditional prompt to wait loop entirely You no longer have to open the app and stare at a blinking cursor and type an instruction The AI acts proactively based on parameters you set up in advance If a client emails an invoice the agent sees it extracts the data and updates your accounting software before you even check your phone It's incredible And

the third major feature is computer use Manus can now access approved files and applications directly on your local machine Local machine access Yes Furthermore you can direct this agent remotely from your mobile phone This completely decouples where you command the AI from where the AI executes the work You can be sitting at a coffee shop on your phone You could type a prompt Your agent can then wake up your desktop at home open Excel manipulate local spreadsheets save the file and send it directly to your boss It separates the direction of the task from the physical location of the execution It is incredibly powerful But

as you said earlier it also introduces massive operational risk That is my main pushback on this entire concept This system requires intense flawless human oversight It really does If an arbitrary incoming email can trigger an automation that touches your local files security is paramount Yeah What happens if a malicious email triggers the agent to delete a directory Or what happens if it misinterprets a Slack message and sends confidential local data to a public channel The approval permissions become the most critical part of the entire setup The convenience is entirely dependent on the security architecture You cannot just let an agent run wild on your desktop

while you're at the grocery store No The user must meticulously decide which incoming events warrant action The user must strictly isolate which specific local files the agent is allowed to touch Remote convenience depends entirely on establishing bulletproof local access limits first And honestly most users are terrible at setting up granular permissions We will definitely need to watch how everyday users navigate this exact friction They have to balance the intense boring setup of local access permissions against the undeniable convenience of remote execution I fully expect we will see some high profile disastrous mistakes as users learn these boundaries the hard way We are shifting to our

Quick Reads section now Let's cover the broader signals First up in Quick Reads Kennedy calls for wider AI analysis of Americans health data This happened at the Maha Open Data Summit in Washington Health Secretary Robert F Kennedy Jr outlined a broad plan to link massive federal data sets He called for connecting Medicare and Medicaid claims medical records and lifestyle data Right This lifestyle data includes information on exercise habits food consumption and environmental chemical exposures The stated goal is to allow hundreds of outside researchers to use AI to study chronic disease patterns across the entire population Kennedy argues that AI can accelerate this kind of epidemiological

discovery He claims AI can perform complex demographic analyses in seconds that previously took researchers years to complete manually He pointed to current federal work using the CDC Vaccine Safety Data Link as a baseline for comparing large scale health outcomes The major caveat here is the complete lack of policy details Kennedy announced no actual methodology to support his claims about analytical speed He provided no results from the ongoing federal studies he referenced More importantly for privacy advocates he announced no new access pathways or data anonymization rules for these outside researchers There are also deep practical obstacles to this vision State data formats sent to the CDC

are notoriously hard to combine They are Different states simply collect different information using entirely different software architectures Watch to see if this ambition actually translates into a defined access process with concrete HIPAA safeguards or if it remains purely aspirational Next up OpenAI adds app like interfaces to chat GPT plugins and plans event triggers OpenAI is giving developers tools to build rich visual interfaces directly inside the chat GPT window Plugins are getting dedicated sidebar homes They are getting interactive conversational panels They're even getting custom file viewers Plugins are really evolving They used to just be invisible data fetchers operating completely behind the scenes You asked a

question it pulled data and fed it back as plain text Now they are becoming embedded mini apps You interact with them visually directly inside your main chat interface Developers can use a new plugin creator tool to build these They also have access to shared chat GPT sites Okay This allows teammates to use the exact same app experience without having to pool their individual connected accounts or share passwords The caveat to watch here involves how these plugins actually start working OpenAI plans to support MCP events MCP events This stands for Model Context Protocol It is a proposed specification for connected app events This would allow an

external app to trigger a plugin automatically It would not require a human to type a chat prompt at all Right Sound familiar It is the exact same proactive automation we just discussed with Manus Watch to see if third party developers actually adopt this proposed MCP event standard in the wild or if they just stick to their own proprietary triggers Our final quick read today Google steers more search users into AI chat as websites face fewer clicks Google is fundamentally changing the traditional search interface They are making their conversational AI mode vastly easier to trigger Yeah they really are Many users now see a prominent unavoidable

AI mode button directly on the home page They have also added a highly visible ask anything prompt directly beneath their existing AI overviews Right This UI design essentially encourages users to continue asking follow up questions right in the chat interface instead of clicking through to a traditional website Google reports massive consumer adoption of this behavior AI mode now has over a billion monthly active users worldwide Wow Furthermore the average query in AI mode is three times longer than a traditional keyword search People are treating it like a conversation not a directory But this interface shift creates existential dread for digital publishers who rely on web

traffic A recent study by Ahrefs analyzed millions of search behaviors It found that the presence of AI overviews is associated with a 58 drop in click through rates for top ranking organic results That is a staggering decline in traffic News site traffic is falling broadly across the entire publishing industry It is Google pushes back on this narrative They argue they're heavily linking to primary sources within their AI content trays They claim users still want to click through to verify facts and read deep dives The ultimate watch item here is purely behavioral Will users actually click those embedded citation links or is being cited by the

AI fully replacing being visited by the user That distinction could literally bankrupt the modern ad supported web We are moving to three takeaways from today First AI is moving from answering isolated questions to holding persistent jobs We see this with OpenAI's dots living and working continuously in the cloud without human prompting We see it with Manus running local files from your phone based on arbitrary Slack triggers The entire paradigm of computing is shifting from query to agency and delegation Second the underlying economics of AI are shifting dramatically We are moving from strict hourly time limits to flexible compute buckets This structural change in pricing is

driven by massive cost reductions in specific backend infrastructure The 95 discount on cached input costs is the quiet essential engine making continuous agent loops financially possible Third AI is becoming the primary interface between raw data and human discovery Google's AI mode is keeping users on the search page to explore concepts instead of navigating the web Microsoft's Quine is crunching multimodal biology data to directly guide lab experiments AI is no longer just a summarization tool It is the definitive lens through which we process complex information The translation layer between raw data and actionable insight is being completely automated There is one specific thing you need to

watch tomorrow Watch for the exact details of the non usage features dropping on OpenAI's Pro 200 tier Those hidden features will determine the true value of their revised pricing strategy which kicks in on September 30th Head over to superpowerdaily com for the complete timeline and original keynote clips Thank you for listening We'll see you tomorrow