OpenAI
An AI research and product company behind ChatGPT, Codex and the Dots agent.
OpenAI plans in 60 to 90 day cycles, builds for models two to three months ahead, and re-releases features once models catch up (Operator to Atlas to Codex). Codex reached over 5 million weekly users, up 6x since January, spreading from engineers to non-technical staff. Lesson: dogfood heavily, retest model-dependent ideas at each leap, and treat curation as the bottleneck.
- 1Non-technical staff adopt CodexMarketing, comms, finance, legal, despite code
- 2Codex added elsewhereInto ChatGPT desktop and Atlas browser
- 3Nobody leaves Codex appThe signal to generalize Codex
- 4Generalize Codex itselfStart simple, add complexity as needed
- 5Ship capabilities into ChatGPTLifting the floor for ~1.2 billion users
From Tibo Sottiaux and Andrew Ambrosino, OpenAI
What was said about OpenAI
1Merging Codex and ChatGPT into one product
OpenAI is folding chat and agentic work into one product with no mode picking. ChatGPT's work mode runs on Codex underneath with the coding UI removed, and all of Codex's capabilities will eventually ship in ChatGPT, lifting the floor for about 1.2 billion users. The north star is that users type a task and the system picks the right harness and model.
Users love work's capability but sometimes want chat because it is faster and more pleasant, so the two are being merged rather than one replacing the other. Sottiaux: "You kind of need a PhD in model pickers." Altman frames the end state as one direct interface plus one API.
Codex grew from a developer tool into a home base for knowledge work. Codex usage grew 6x since January to over 5 million weekly active users, and marketing, comms, finance and legal staff adopted it even though it was built for developers. None of them would leave it for the apps meant for them.
It drives specialist tools rather than replacing them: it talks to Excel's add-in and built itself an extension to edit in Premiere Pro. The quality bar is that people never hesitate to open it, the way they reflexively open a browser tab, and Andrew Ambrosino's stated aim is the best desktop app ever built.
An imperfect agent toggle got agents to a billion users without breaking developers. OpenAI shipped a rough toggle so an agentic harness could reach over a billion ChatGPT users while leaving developers' workflows intact. The obvious next step after the toggle is no toggle.
Users accept turns in a product when they can see the story behind them. Tara Seshan says a coherent story lets a team change a lot without burning bridges.
2Planning 60 to 90 days out and shipping before it's perfect
Build for where the models will be in two to three months. Building for today's models gets you left behind, and building for a model a year away is equally wrong. Product constructs should get out of the model's way so the product is right when the next model lands.
Tara keeps a personal doc of her failed predictions as a reminder of how poor long-range forecasts are. Product work stays tied to the research agenda, such as which capabilities (coding, writing) the models are being pushed on.
Set planning cadence by how predictable your market is. Stripe's payments market can still be modeled with bull, base and bear cases and planned annually, but OpenAI's business cannot, so it plans in 60 to 90 day horizons.
Andrew Ambrosino describes plans as detailed in the short term and hazy over the long term. He is confident the Codex app would have failed if released in November instead of February, with the models as the only difference, and calls Operator, Atlas and Codex web early versions of ideas that worked only once models caught up.
In an emergent market, a fast test beats a rigorous plan. Tara wrote plans like a "PhD thesis" at Stripe, where polish mattered and every enum name was weighed. At OpenAI she names the one most important thing to test and ships done before perfect when the product is transformative.
The ChatGPT work launch drew confusion comments and the team iterated fast in response. She says nothing compares to watching users try a feature.
Dogfood everything and ship only what people keep using internally. Features ship internally first, and the team checks whether people keep using them, find something delightful, and whether they unlock a new use of the models. There is no fixed number, and weak internal uptake is a signal not to ship.
Three recurring memes keep the loop tight: whether scope is ambitious enough, "Is this maximally accelerated?" (from Nick Turley), and "Are you mainlining it yet?", meaning using the product all day for your own work.
Ship the frontier to enterprises even when they say it's too fast. Most enterprise customers were using ChatGPT only for questions and answers while the agent revolution was happening, so OpenAI moved them to agents even though it broke their processes and expectations for how updates arrive.
Customer success and sales teams say clients cannot absorb the rate of change. The team's rule: "do what your users need, not what your users say they need," or risk being leapfrogged by alternatives.
Let bottoms-up projects grow, then hold a high bar on launch. The decisions API began as a Slack channel of four people hacking on a weekend after realizing the Luna model could use constraint sampling, and it shipped quickly.
Leadership pulls back things that need more baking, such as products held from DevDay to launch in the following weeks. "It feels completely unstructured and then you take it all the way and ship something."
3Dots, the always-on agent
One learning agent running 24/7 replaces hand-tuned loops and model pickers. Dots learns goals and preferences from feedback, and the only thing users configure is which channels they talk through. The harness does not run on the user's machine, so one dot can reach many devices.
Sottiaux wants one intelligence reachable from any screen, joining a meeting to take notes, then picked up by email or text. He calls the laptop a brick that ties you to the technology. The always-on costs, such as mistakes, security risk and a demo-time outage Dots could not fix, got little scrutiny.
Dots launched with one primary agent before letting users build teams. Everyone gets one primary dot that they text and that learns their preferences most deeply, with a second, third and fourth coming later to form a virtual team. Starting with one lets OpenAI see what people do and tune the system.
Sottiaux runs one dot just for monitoring Twitter, a job big enough for a whole dot. His own agent teams expand while he pushes the frontier and shrink when a better model lets one agent do it all.
Dots joins years of long-task and memory research to the Codex harness. Work on long, persistent tasks and on memory each go back more than two years, and Codex Cloud, shipped about a year ago, inspired the rival agent bots that appeared since.
Dots launched with Astra, described as OpenAI's safest, most aligned model, with extra effort on security. Sottiaux writes almost no code by hand and uses Codex for much of the Dots code.
4The plugin ecosystem, sign-in with ChatGPT and revenue sharing
ChatGPT recommends plugins by retention and quality, not clever wording. ChatGPT surfaces plugins in conversations based on retention numbers and quality, so a strong plugin can reach a significant slice of its 1.2 billion users and a weak one stops being recommended.
Getting people to write about you does not help. The test is whether a plugin lets ChatGPT do something new. "Obviously, if your plugin is not quite good, they will stop being recommended."
OpenAI pays partners when subscribers spend usage in their products. When subscribers spend ChatGPT usage inside a popular plugin such as Notion or Figma, or in a product using sign-in with ChatGPT, the partner gets a revenue share. Sign-in with ChatGPT has 16 launch partners.
Sign-in began as a virtual handshake with the creators of Pi and OpenCode, who trusted each other not to do anything shady, and its popularity made it a supported feature. Revenue sharing was not mentioned in the keynote.
Structure capabilities as layers, with computer use as the guaranteed fallback. OpenAI treats the platform as an array of routes: native features, first and third party plugins via shared hooks, then computer use when a plugin or MCP is missing. Computer use may be slow and token-heavy but finishes the job.
Nan Yu: an agent that gets "99% there and then it just like barfs" is worse than one that never starts. A meetings plugin can pull all of a user's meeting data, and computer use covers missing interfaces until the ecosystem catches up.
Prepare for agents taking most actions on the internet. Sottiaux expects most internet actions to be taken by agents and tells builders to imagine everything roughly 10 times better in a year.
Notion's MCP drew so much agent traffic that it strained its systems and forced Notion to work out the economics. The claim comes with no data or timeline.
5Sam Altman's bets: research, compute and cutting good products
OpenAI ran four and a half years without a product by simulating users. With no customer signal, the team used objective measures like a leaderboard during the Dota 2 reinforcement learning work, plus external demos for eminent people researchers wanted to impress. Fake deadlines did not work.
The founding day in January 2016 at Greg Brockman's apartment had about 11 or 12 people and no plan, so they wrote papers and slowly found a rhythm. The work ran from the sentiment neuron paper to GPT-1 and the scaling laws, and Altman mostly remembers failing to raise money.
Back non-consensus bets where a win would be hugely valuable. In 2015 almost nobody besides DeepMind chased AGI while thousands of founders built photo sharing apps, and OpenAI was hammered for it, then again for focusing on large language models. Altman's rule: high-risk bets are fine if success is super valuable.
He avoids founders with "a thin veneer" of difference on the last thousand pitches and backs people who stand by unpopular convictions. He credits YC's ideas of iterative deployment, technical people in charge and betting on young people with making OpenAI possible.
Altman spends most of his time on models and compute, the hardest lever. He believes that if models and abundant compute are right, everything else follows, and calls the compute buildout probably the most expensive infrastructure project in history.
It spans chips, fabs, racks, power, finance, policy and supply chain. "Try to find the high-leverage difficult problem that will continue the exponential."
OpenAI killed Sora and Atlas to concentrate compute on Codex. With limited compute and people, OpenAI shut down Sora, which used a lot of compute and mattered less than Codex, and its Atlas browser, moving resources to Codex and general intelligence for knowledge work and science.
Altman calls it a good product sacrificed for a great one, says this is the hardest lesson for any entrepreneur, and admits he is terrible at it.
Be a platform for 100 million businesses, not a builder of everything. Altman wants one interface and one API with the best AI at every point of the cost and performance curve, without competing with customers or trying to subsume the economy.
He expects the greatest boom in small business formation ever and says the field, OpenAI included, has not talked about it enough.
AI's missing iPhone moment is a product failure, not a technology gap. Altman still clicks around, pastes between apps and handles email and his to-do list by hand despite having Codex. He says the latest models are limited more by context on him than by intelligence.
He wants an agent that reads internal Slack, customer stories and papers and advises him on decisions. He compares today to his Palm Treo around 2003 or 2004, when product ideas were missing more than hardware.
6Safety through deployment and agent monitoring
Deploying imperfect AI early is how OpenAI makes it safe. Altman says putting products into the world and seeing where they break makes them both good and safe, modeled on aviation's accident reporting and public postmortems. ChatGPT reached about a billion users in under four years without the predicted catastrophes.
"You're not going to solve it by disconnecting yourself from reality."
Spend compute on a second system that watches the primary agent. OpenAI spends more and more compute on secondary monitoring that stops high-risk actions or interrupts the agent if it may be prompt injected. Sottiaux says most investment on the API stack goes into safety.
A model, 6.1 Astra, was not released, and OpenAI holds back models until they feel hardened. "We don't gamble there."
Altman names loss of control and concentrated power as the two risks. He rejects trading autonomy for cures and cheap goods as anti-human, and his principle is to keep people deeply in control and empowered.
He says the two risks are a little in tension, and that builders fueled fear by talking of a 25% chance of destroying the world without explaining benefits or mitigation.
7How product works with research and designs agents
Bring research real user sessions and evals, then post-train the fix in. Product people read sample sessions to see what the model failed to do, write evals as much as possible, and when a prompt and skills reach the outcome, ask research to post-train it into the model.
Research trusts you enough to iterate on evals only once you bring good user data. Tara calls writing good evals "the most important thing."
Design agents around human limits and practical questions, not philosophy. People cannot manage 40 agents, so they bundle them under a chief of staff agent. The right design depends on memory, permissions and whose credentials the agent uses.
Memory may need walls per place, like a private Slack channel, and the UI must show which identity and data the agent is using. Nan and Tara name user empathy, systems thinking and relentless iteration as the skills.
Users must be able to predict what a semi-autonomous agent will do. With agents acting on their own, surprising behavior can no longer be excused with a settings menu or popup. Users also need to understand privacy and how their data is used.
Onboarding matters far more because of the capability overhang. Models can do much more than people use, so the limit is how much users can absorb, and the initial experience drives success even though many engineers dislike working on it.
Tara is bullish on voice for 2027 as an intuitive interface: it cut the tech support she does for family and changed how OpenAI's people team onboarded onto a new product.
Stay DMable and ask follow-ups, because agent feedback is subtle. OpenAI's product and engineering team is highly reachable by users, and good feedback on problems like a bad tone often comes only on the second or third follow-up question.
Being a DMable PM gives research the specific use cases it needs.
8Taste, ownership and who OpenAI hires
With implementation cheap, taste and curation are the bottleneck. Unlimited tokens mean everybody at OpenAI builds everything, so there are roughly 90 uncoordinated attempts at any needed feature, and 100% of Codex's code is AI-written.
Taste means judging what to build and picking the medium, such as a document for vague areas and a prototype for testing an interaction. The cost of reviewing the 90 attempts is left unquantified.
Hire for taste, user focus and drive to build, not typing speed. OpenAI hires many founders for this reason and has more than 120 ex-YC founders on staff, while roles between design, engineering and product blur.
Product managers run a "zone defense", spreading out to cover gaps. With execution cheap, Tara says raising others' ambition, asking for a faster or 10x bigger version, is a core part of the PM job.
Give people founder-like autonomy and expect fast fixes. Teams own their areas like founders with little top-down direction, and Sottiaux took production down on his third day and kept his job.
Tara expected a strategy bible and found none: OpenAI's thinking quickly becomes public product. Sottiaux can press the reset button without a chain of approvals.
9Design at OpenAI under Ian Silber
Give most users a simple default and test new things with enthusiasts. Usage ranges from recipe ideas to automating a farm, so ChatGPT keeps one streamlined input for a billion users while early features like the desktop app and Codex go first to people who care.
Ian sees design moving past the blank chat box toward adaptive, proactive interfaces, with writing blocks and tap-through follow-up questions already appearing.
Hire designers for curiosity and prototyping, not AI background. OpenAI looks for curiosity, aptitude for using AI in design, prototyping skill and systems thinking, and builds teams of complementary strengths rather than identical generalists.
The mix covers generalists, visual and brand designers, prototypers and strategic thinkers. "I haven't found that one designer necessarily checks every single one of those boxes."
Throw ideas straight into a coding agent to prototype them. OpenAI designers drop ideas into Codex to think them through, and Ian types late-night ideas into ChatGPT Cowork in the cloud and brings the artifact to the team.
Its research-lab culture lets designers try bad ideas and throw them out, and empowers them to package research concepts into products. Ian contrasts this with Instagram's slower "do the simple thing first" DNA.
10Codex features: shareable sites and /visualize
Codex builds shareable sites from a prompt instead of slide decks. Sites have a database, can auto update, and can be public, team-only or private. Tara built a team game and a backpacking route with elevation and food inputs.
A /visualize command pulls in data such as your ChatGPT usage and arranges it into a visual story.
11How outsiders see OpenAI
Ads in chat may need auctions that price trust, not only bids. Peter Sellis says trust could replace organic value in the auction term, deciding how much advertisers can influence an answer, and he was disappointed that OpenAI's top tier is ad-free, cutting off its most valuable users.
He cites Amazon showing 16 ads before the first search result as eroded trust.
Vertical apps compete with ChatGPT through proprietary data and opinionated workflows. Career app Apt treats ChatGPT as its main rival and beats its slow web searches with a database of over 10 million jobs and assessments from over two million people.
Its founder doubts the labs care enough to build a dedicated career product. Other founders note OpenAI extended a coding surface to other work, the opposite of xAI's Grok Bot, and that observers assumed OpenAI had won in 2023.
OpenAI’s Head of ChatGPT: We’re entering a new era of AI (again) | Tibo Sottiaux
Merging chat and work in ChatGPT, building an open plugin ecosystem with revenue sharing, and spending more compute on safety monitoring. It has over 120 ex-YC founders and holds back models until they feel hardened. Tibo writes almost no code by hand and uses Codex for analysis and for much of the Dots code. Codex Cloud shipped about a year ago and inspired rival agent bots.
OpenAI is folding chat and agentic work into one product with no mode picking. ChatGPT's work mode runs on Codex underneath with the coding UI removed, and all of Codex's capabilities will eventually ship in ChatGPT, lifting the floor for about 1.2 billion users. The north star is that users type a task and the system picks the right harness and model.
Users love work's capability but sometimes want chat because it is faster and more pleasant, so the two are being merged rather than one replacing the other. Sottiaux: "You kind of need a PhD in model pickers." Altman frames the end state as one direct interface plus one API.
Let bottoms-up projects grow, then hold a high bar on launch. The decisions API began as a Slack channel of four people hacking on a weekend after realizing the Luna model could use constraint sampling, and it shipped quickly.
Leadership pulls back things that need more baking, such as products held from DevDay to launch in the following weeks. "It feels completely unstructured and then you take it all the way and ship something."
One learning agent running 24/7 replaces hand-tuned loops and model pickers. Dots learns goals and preferences from feedback, and the only thing users configure is which channels they talk through. The harness does not run on the user's machine, so one dot can reach many devices.
Sottiaux wants one intelligence reachable from any screen, joining a meeting to take notes, then picked up by email or text. He calls the laptop a brick that ties you to the technology. The always-on costs, such as mistakes, security risk and a demo-time outage Dots could not fix, got little scrutiny.
Dots launched with one primary agent before letting users build teams. Everyone gets one primary dot that they text and that learns their preferences most deeply, with a second, third and fourth coming later to form a virtual team. Starting with one lets OpenAI see what people do and tune the system.
Sottiaux runs one dot just for monitoring Twitter, a job big enough for a whole dot. His own agent teams expand while he pushes the frontier and shrink when a better model lets one agent do it all.
Dots joins years of long-task and memory research to the Codex harness. Work on long, persistent tasks and on memory each go back more than two years, and Codex Cloud, shipped about a year ago, inspired the rival agent bots that appeared since.
Dots launched with Astra, described as OpenAI's safest, most aligned model, with extra effort on security. Sottiaux writes almost no code by hand and uses Codex for much of the Dots code.
ChatGPT recommends plugins by retention and quality, not clever wording. ChatGPT surfaces plugins in conversations based on retention numbers and quality, so a strong plugin can reach a significant slice of its 1.2 billion users and a weak one stops being recommended.
Getting people to write about you does not help. The test is whether a plugin lets ChatGPT do something new. "Obviously, if your plugin is not quite good, they will stop being recommended."
OpenAI pays partners when subscribers spend usage in their products. When subscribers spend ChatGPT usage inside a popular plugin such as Notion or Figma, or in a product using sign-in with ChatGPT, the partner gets a revenue share. Sign-in with ChatGPT has 16 launch partners.
Sign-in began as a virtual handshake with the creators of Pi and OpenCode, who trusted each other not to do anything shady, and its popularity made it a supported feature. Revenue sharing was not mentioned in the keynote.
Prepare for agents taking most actions on the internet. Sottiaux expects most internet actions to be taken by agents and tells builders to imagine everything roughly 10 times better in a year.
Notion's MCP drew so much agent traffic that it strained its systems and forced Notion to work out the economics. The claim comes with no data or timeline.
Spend compute on a second system that watches the primary agent. OpenAI spends more and more compute on secondary monitoring that stops high-risk actions or interrupts the agent if it may be prompt injected. Sottiaux says most investment on the API stack goes into safety.
A model, 6.1 Astra, was not released, and OpenAI holds back models until they feel hardened. "We don't gamble there."
Hire for taste, user focus and drive to build, not typing speed. OpenAI hires many founders for this reason and has more than 120 ex-YC founders on staff, while roles between design, engineering and product blur.
Product managers run a "zone defense", spreading out to cover gaps. With execution cheap, Tara says raising others' ambition, asking for a faster or 10x bigger version, is a core part of the PM job.
Give people founder-like autonomy and expect fast fixes. Teams own their areas like founders with little top-down direction, and Sottiaux took production down on his third day and kept his job.
Tara expected a strategy bible and found none: OpenAI's thinking quickly becomes public product. Sottiaux can press the reset button without a chain of approvals.
Stop planning for 2027: how OpenAI builds product 90 days at a time | Tara Sesha and Nan Yu (OpenAI)
Plans in 60 to 90 day horizons because its market can't be modeled annually. It ships frontier features like agents to enterprises fast, even when it breaks their processes, and works with research through user sessions and evals. Claire runs a single Codex agent on her desktop. She also predicts a small Codex-in-a-box hardware device by 2027.
An imperfect agent toggle got agents to a billion users without breaking developers. OpenAI shipped a rough toggle so an agentic harness could reach over a billion ChatGPT users while leaving developers' workflows intact. The obvious next step after the toggle is no toggle.
Users accept turns in a product when they can see the story behind them. Tara Seshan says a coherent story lets a team change a lot without burning bridges.
Build for where the models will be in two to three months. Building for today's models gets you left behind, and building for a model a year away is equally wrong. Product constructs should get out of the model's way so the product is right when the next model lands.
Tara keeps a personal doc of her failed predictions as a reminder of how poor long-range forecasts are. Product work stays tied to the research agenda, such as which capabilities (coding, writing) the models are being pushed on.
Set planning cadence by how predictable your market is. Stripe's payments market can still be modeled with bull, base and bear cases and planned annually, but OpenAI's business cannot, so it plans in 60 to 90 day horizons.
Andrew Ambrosino describes plans as detailed in the short term and hazy over the long term. He is confident the Codex app would have failed if released in November instead of February, with the models as the only difference, and calls Operator, Atlas and Codex web early versions of ideas that worked only once models caught up.
In an emergent market, a fast test beats a rigorous plan. Tara wrote plans like a "PhD thesis" at Stripe, where polish mattered and every enum name was weighed. At OpenAI she names the one most important thing to test and ships done before perfect when the product is transformative.
The ChatGPT work launch drew confusion comments and the team iterated fast in response. She says nothing compares to watching users try a feature.
Dogfood everything and ship only what people keep using internally. Features ship internally first, and the team checks whether people keep using them, find something delightful, and whether they unlock a new use of the models. There is no fixed number, and weak internal uptake is a signal not to ship.
Three recurring memes keep the loop tight: whether scope is ambitious enough, "Is this maximally accelerated?" (from Nick Turley), and "Are you mainlining it yet?", meaning using the product all day for your own work.
Ship the frontier to enterprises even when they say it's too fast. Most enterprise customers were using ChatGPT only for questions and answers while the agent revolution was happening, so OpenAI moved them to agents even though it broke their processes and expectations for how updates arrive.
Customer success and sales teams say clients cannot absorb the rate of change. The team's rule: "do what your users need, not what your users say they need," or risk being leapfrogged by alternatives.
Structure capabilities as layers, with computer use as the guaranteed fallback. OpenAI treats the platform as an array of routes: native features, first and third party plugins via shared hooks, then computer use when a plugin or MCP is missing. Computer use may be slow and token-heavy but finishes the job.
Nan Yu: an agent that gets "99% there and then it just like barfs" is worse than one that never starts. A meetings plugin can pull all of a user's meeting data, and computer use covers missing interfaces until the ecosystem catches up.
Bring research real user sessions and evals, then post-train the fix in. Product people read sample sessions to see what the model failed to do, write evals as much as possible, and when a prompt and skills reach the outcome, ask research to post-train it into the model.
Research trusts you enough to iterate on evals only once you bring good user data. Tara calls writing good evals "the most important thing."
Design agents around human limits and practical questions, not philosophy. People cannot manage 40 agents, so they bundle them under a chief of staff agent. The right design depends on memory, permissions and whose credentials the agent uses.
Memory may need walls per place, like a private Slack channel, and the UI must show which identity and data the agent is using. Nan and Tara name user empathy, systems thinking and relentless iteration as the skills.
Users must be able to predict what a semi-autonomous agent will do. With agents acting on their own, surprising behavior can no longer be excused with a settings menu or popup. Users also need to understand privacy and how their data is used.
Onboarding matters far more because of the capability overhang. Models can do much more than people use, so the limit is how much users can absorb, and the initial experience drives success even though many engineers dislike working on it.
Tara is bullish on voice for 2027 as an intuitive interface: it cut the tech support she does for family and changed how OpenAI's people team onboarded onto a new product.
Stay DMable and ask follow-ups, because agent feedback is subtle. OpenAI's product and engineering team is highly reachable by users, and good feedback on problems like a bad tone often comes only on the second or third follow-up question.
Being a DMable PM gives research the specific use cases it needs.
AI’s third era: the rise of persistent AI coworkers | Tara Seshan (OpenAI’s product lead)
Teams own their areas like founders, ideas reach users faster than anywhere else Tara has worked, and there is no secret strategy vault. Product work stays tied to the research agenda, and memes like 'Is this maximally accelerated?' keep feedback loops tight. Codex gained attention because the market noticed habits the team already had, like tight iteration and mainlining the app. It builds shareable sites from a prompt and has a /visualize command that turns usage data into a story.
Ads in chat may need auctions that price trust, not only bids. Peter Sellis says trust could replace organic value in the auction term, deciding how much advertisers can influence an answer, and he was disappointed that OpenAI's top tier is ad-free, cutting off its most valuable users.
He cites Amazon showing 16 ads before the first search result as eroded trust.
Sam Altman on Building OpenAI & Betting on the Impossible
Went 4.5 years without a product, then bet on ChatGPT, which reached about a billion users in under four years. It now cuts good products to concentrate compute on general intelligence, and aims to be a platform enabling 100 million new businesses. Merged with ChatGPT into one interface, and Altman says it was unfortunately named since it can do any kind of work. It received the compute freed by killing Sora, though Altman still handles email and tasks by hand. Killed because it used a lot of compute and mattered less than Codex, where that compute went. Altman calls it a good product sacrificed for a great one.
Vertical apps compete with ChatGPT through proprietary data and opinionated workflows. Career app Apt treats ChatGPT as its main rival and beats its slow web searches with a database of over 10 million jobs and assessments from over two million people.
Its founder doubts the labs care enough to build a dedicated career product. Other founders note OpenAI extended a coding surface to other work, the opposite of xAI's Grok Bot, and that observers assumed OpenAI had won in 2023.
Chatbots are not the final interface: OpenAI’s Head of Design on what’s next | Ian Silber
Ian's coding agents kept hitting hiccups until Codex dropped a month later and just worked. Designers throw ideas straight into it to prototype, and it tests new ideas early with enthusiasts. Ian runs design there. It hires for curiosity and prototyping rather than AI background, and its research-lab culture lets designers try bad ideas and empowers them to package research concepts into products.
Why OpenAI is merging Codex and ChatGPT and the future of knowledge work | Andrew Ambrosino
Usage has grown 6x since January to over 5 million weekly active users, and 100% of its code is AI-written. Marketing, comms, finance and legal staff adopted it though built for developers, so it is becoming a home base that drives tools like Excel and Premiere Pro. With unlimited tokens everyone builds everything, so there are roughly 90 uncoordinated attempts at any needed feature. Curation, not building, is the bottleneck. An early version of an idea that only worked once models caught up, later reshaped as agent mode, Atlas and Codex's in-app browser.
OpenAI is folding chat and agentic work into one product with no mode picking. ChatGPT's work mode runs on Codex underneath with the coding UI removed, and all of Codex's capabilities will eventually ship in ChatGPT, lifting the floor for about 1.2 billion users. The north star is that users type a task and the system picks the right harness and model.
Users love work's capability but sometimes want chat because it is faster and more pleasant, so the two are being merged rather than one replacing the other. Sottiaux: "You kind of need a PhD in model pickers." Altman frames the end state as one direct interface plus one API.
Build for where the models will be in two to three months. Building for today's models gets you left behind, and building for a model a year away is equally wrong. Product constructs should get out of the model's way so the product is right when the next model lands.
Tara keeps a personal doc of her failed predictions as a reminder of how poor long-range forecasts are. Product work stays tied to the research agenda, such as which capabilities (coding, writing) the models are being pushed on.
In an emergent market, a fast test beats a rigorous plan. Tara wrote plans like a "PhD thesis" at Stripe, where polish mattered and every enum name was weighed. At OpenAI she names the one most important thing to test and ships done before perfect when the product is transformative.
The ChatGPT work launch drew confusion comments and the team iterated fast in response. She says nothing compares to watching users try a feature.
Dogfood everything and ship only what people keep using internally. Features ship internally first, and the team checks whether people keep using them, find something delightful, and whether they unlock a new use of the models. There is no fixed number, and weak internal uptake is a signal not to ship.
Three recurring memes keep the loop tight: whether scope is ambitious enough, "Is this maximally accelerated?" (from Nick Turley), and "Are you mainlining it yet?", meaning using the product all day for your own work.
Hire for taste, user focus and drive to build, not typing speed. OpenAI hires many founders for this reason and has more than 120 ex-YC founders on staff, while roles between design, engineering and product blur.
Product managers run a "zone defense", spreading out to cover gaps. With execution cheap, Tara says raising others' ambition, asking for a faster or 10x bigger version, is a core part of the PM job.
Give people founder-like autonomy and expect fast fixes. Teams own their areas like founders with little top-down direction, and Sottiaux took production down on his third day and kept his job.
Tara expected a strategy bible and found none: OpenAI's thinking quickly becomes public product. Sottiaux can press the reset button without a chain of approvals.
Codex builds shareable sites from a prompt instead of slide decks. Sites have a database, can auto update, and can be public, team-only or private. Tara built a team game and a backpacking route with elevation and food inputs.
A /visualize command pulls in data such as your ChatGPT usage and arranges it into a visual story.
Why most AI products fail: Lessons from 50+ AI deployments at OpenAI, Google & Amazon
OpenAI's support volume spiked after launches like image generation and GPT-5, and its support flow starts with AI suggestions that human agents mark good or bad. The guests work there, and the company ships ChatGPT Pulse for daily updates. Codex uses core evals, A/B tests, customer feedback and hands-on testing of each new model, because a customizable coding agent has no fixed set of workflows to test. Its code review product is validated by A/B tests, and the team watches social media for problems and fixes them quickly.
OpenAI is folding chat and agentic work into one product with no mode picking. ChatGPT's work mode runs on Codex underneath with the coding UI removed, and all of Codex's capabilities will eventually ship in ChatGPT, lifting the floor for about 1.2 billion users. The north star is that users type a task and the system picks the right harness and model.
Users love work's capability but sometimes want chat because it is faster and more pleasant, so the two are being merged rather than one replacing the other. Sottiaux: "You kind of need a PhD in model pickers." Altman frames the end state as one direct interface plus one API.
OpenAI ran four and a half years without a product by simulating users. With no customer signal, the team used objective measures like a leaderboard during the Dota 2 reinforcement learning work, plus external demos for eminent people researchers wanted to impress. Fake deadlines did not work.
The founding day in January 2016 at Greg Brockman's apartment had about 11 or 12 people and no plan, so they wrote papers and slowly found a rhythm. The work ran from the sentiment neuron paper to GPT-1 and the scaling laws, and Altman mostly remembers failing to raise money.
Back non-consensus bets where a win would be hugely valuable. In 2015 almost nobody besides DeepMind chased AGI while thousands of founders built photo sharing apps, and OpenAI was hammered for it, then again for focusing on large language models. Altman's rule: high-risk bets are fine if success is super valuable.
He avoids founders with "a thin veneer" of difference on the last thousand pitches and backs people who stand by unpopular convictions. He credits YC's ideas of iterative deployment, technical people in charge and betting on young people with making OpenAI possible.
Altman spends most of his time on models and compute, the hardest lever. He believes that if models and abundant compute are right, everything else follows, and calls the compute buildout probably the most expensive infrastructure project in history.
It spans chips, fabs, racks, power, finance, policy and supply chain. "Try to find the high-leverage difficult problem that will continue the exponential."
OpenAI killed Sora and Atlas to concentrate compute on Codex. With limited compute and people, OpenAI shut down Sora, which used a lot of compute and mattered less than Codex, and its Atlas browser, moving resources to Codex and general intelligence for knowledge work and science.
Altman calls it a good product sacrificed for a great one, says this is the hardest lesson for any entrepreneur, and admits he is terrible at it.
Be a platform for 100 million businesses, not a builder of everything. Altman wants one interface and one API with the best AI at every point of the cost and performance curve, without competing with customers or trying to subsume the economy.
He expects the greatest boom in small business formation ever and says the field, OpenAI included, has not talked about it enough.
AI's missing iPhone moment is a product failure, not a technology gap. Altman still clicks around, pastes between apps and handles email and his to-do list by hand despite having Codex. He says the latest models are limited more by context on him than by intelligence.
He wants an agent that reads internal Slack, customer stories and papers and advises him on decisions. He compares today to his Palm Treo around 2003 or 2004, when product ideas were missing more than hardware.
Deploying imperfect AI early is how OpenAI makes it safe. Altman says putting products into the world and seeing where they break makes them both good and safe, modeled on aviation's accident reporting and public postmortems. ChatGPT reached about a billion users in under four years without the predicted catastrophes.
"You're not going to solve it by disconnecting yourself from reality."
Altman names loss of control and concentrated power as the two risks. He rejects trading autonomy for cures and cheap goods as anti-human, and his principle is to keep people deeply in control and empowered.
He says the two risks are a little in tension, and that builders fueled fear by talking of a 25% chance of destroying the world without explaining benefits or mitigation.
Inside OpenAI: 2026 is the year of agents, AI’s biggest bottleneck, and why compute isn’t the issue
Usage is up about 20x since August and the models serve many trillions of tokens a week. It grew by starting in the IDE and terminal with a local sandbox, and the model, API and harness are built together. Its bottoms-up culture learns empirically and moves far faster than a startup. Staff train reasoning models all day, so internal dogfooding of async cloud agents differs from the wider market. The Android app took 18 days to reach employees and 28 days to reach the public with two or three engineers, using Codex to port the iOS app. It became the number one app in the App Store in under a month.
Vertical apps compete with ChatGPT through proprietary data and opinionated workflows. Career app Apt treats ChatGPT as its main rival and beats its slow web searches with a database of over 10 million jobs and assessments from over two million people.
Its founder doubts the labs care enough to build a dedicated career product. Other founders note OpenAI extended a coding surface to other work, the opposite of xAI's Grok Bot, and that observers assumed OpenAI had won in 2023.
Inside ChatGPT: The fastest growing product in history | Nick Turley (OpenAI)
Runs consumer, API, and business products at once, which Turley calls objectively crazy. It keeps heavy process like red teaming and system cards for frontier model safety while product teams move fast with a maximally accelerated mindset. Described as the smartest, most useful and fastest model OpenAI has shipped, free for everyone, with its biggest jump in coding and front-end work. It is state of the art on math, reasoning and health, and users will judge it by feel more than benchmarks. Shipped rough, and the team learned a lot only after launch. Real use cases then let them optimize it into what is now data analysis.
Give most users a simple default and test new things with enthusiasts. Usage ranges from recipe ideas to automating a farm, so ChatGPT keeps one streamlined input for a billion users while early features like the desktop app and Codex go first to people who care.
Ian sees design moving past the blank chat box toward adaptive, proactive interfaces, with writing blocks and tap-through follow-up questions already appearing.
Hire designers for curiosity and prototyping, not AI background. OpenAI looks for curiosity, aptitude for using AI in design, prototyping skill and systems thinking, and builds teams of complementary strengths rather than identical generalists.
The mix covers generalists, visual and brand designers, prototypers and strategic thinkers. "I haven't found that one designer necessarily checks every single one of those boxes."
Throw ideas straight into a coding agent to prototype them. OpenAI designers drop ideas into Codex to think them through, and Ian types late-night ideas into ChatGPT Cowork in the cloud and brings the artifact to the team.
Its research-lab culture lets designers try bad ideas and throw them out, and empowers them to package research concepts into products. Ian contrasts this with Instagram's slower "do the simple thing first" DNA.
From ChatGPT to Instagram to Uber: The quiet architect behind the world’s most popular products
Deng led ChatGPT product here and called it his last operating role. Product sat next to post-training, and the company created a model designer role after Joanne Jang wrote a job description, hiring two people.
Vertical apps compete with ChatGPT through proprietary data and opinionated workflows. Career app Apt treats ChatGPT as its main rival and beats its slow web searches with a database of over 10 million jobs and assessments from over two million people.
Its founder doubts the labs care enough to build a dedicated career product. Other founders note OpenAI extended a coding surface to other work, the opposite of xAI's Grok Bot, and that observers assumed OpenAI had won in 2023.
Growth tactics from OpenAI and Stripe’s first marketer | Krithika Shankarraman
Awareness was never the problem for ChatGPT; people did not know what to use it for, so marketing built a use case epiphany. Turning on the ChatGPT Enterprise contact sales form raised lead volume 40x overnight, and a ChatGPT-coded Python script served as the first lead scoring model.
Codex grew from a developer tool into a home base for knowledge work. Codex usage grew 6x since January to over 5 million weekly active users, and marketing, comms, finance and legal staff adopted it even though it was built for developers. None of them would leave it for the apps meant for them.
It drives specialist tools rather than replacing them: it talks to Excel's add-in and built itself an extension to edit in Premiere Pro. The quality bar is that people never hesitate to open it, the way they reflexively open a browser tab, and Andrew Ambrosino's stated aim is the best desktop app ever built.
Set planning cadence by how predictable your market is. Stripe's payments market can still be modeled with bull, base and bear cases and planned annually, but OpenAI's business cannot, so it plans in 60 to 90 day horizons.
Andrew Ambrosino describes plans as detailed in the short term and hazy over the long term. He is confident the Codex app would have failed if released in November instead of February, with the models as the only difference, and calls Operator, Atlas and Codex web early versions of ideas that worked only once models caught up.
With implementation cheap, taste and curation are the bottleneck. Unlimited tokens mean everybody at OpenAI builds everything, so there are roughly 90 uncoordinated attempts at any needed feature, and 100% of Codex's code is AI-written.
Taste means judging what to build and picking the medium, such as a document for vague areas and a prototype for testing an interaction. The cost of reviewing the 90 attempts is left unquantified.
Hire for taste, user focus and drive to build, not typing speed. OpenAI hires many founders for this reason and has more than 120 ex-YC founders on staff, while roles between design, engineering and product blur.
Product managers run a "zone defense", spreading out to cover gaps. With execution cheap, Tara says raising others' ambition, asking for a faster or 10x bigger version, is a core part of the PM job.
OpenAI’s CPO on how AI changes must-have skills, moats, coding, startup playbooks, more | Kevin Weil
Ships a new o-series reasoning model every three to four months, builds products from evals and fine-tuning, and has only about 25 PMs. Customer support for 400 million weekly users runs with 30 to 40 people, and about 3 million developers use the API.
A conversation with OpenAI's CPO Kevin Weil, Anthropic's CPO Mike Krieger, and Sarah Guo
Sells consumer, enterprise and developer products at once. Ran an eval-writing boot camp for every PM, and researchers sometimes had a capability for three months before realizing it mattered. One large customer asked for 60 days' notice before any launch. Scales intelligence at query time, thinking 30 to 60 seconds now and possibly hours or days later. It reasons well but is slow and not multimodal, and Weil says it is at the GPT-1 phase of this scaling. Krieger calls it an early UI for async work. Weil used it as a live translator in business talks in Korea and Japan with people who shared no language with him. Teens on TikTok pour their hearts out to it. Co-designed, co-researched and co-fine-tuned with research teams, with designers embedded early. Finishing it was a normal product job once the capability existed.
Inside OpenAI | Logan Kilpatrick (head of developer relations)
Moves fast by hiring high-agency people, with half-year and quarter goals instead of OKRs, and tests each decision against the mission and reliability. Headcount was about 750 to 780 at the end of last year, with research kept deliberately small because researchers share a fixed GPU pool.
90 minutes of unfiltered product advice from Snap and Discord’s product chief | Peter Sellis
It is rebuilding a solved ads stack, and the open question is how a modified auction prices trust in chat. Sellis was disappointed it made the top tier ad-free, which cuts off its most valuable users early.
How we built Grok Bot in a month | Roman Ugarte (SpaceXAI)
It took the opposite path from Grok Bot, extending a coding surface to other work, and is now merging into one product.
Why companies are becoming a series of loops | Anish Acharya (a16z)
OpenAI's go-to-market team reportedly uses it more than its engineers do. Anish built a 20-slide Mother's Day deck for his wife with it, drawing on texts and photos.
Building the AI Career Mentor With a Freemium Bet
It is the main rival Apt must beat. It lacks tailored assessments and runs slow web searches for jobs, and the labs are unlikely to care enough to build a dedicated career product.
Why AI is going vertical (again) | Dianne Penn (Anthropic)
In 2023 observers assumed OpenAI had won. GPT-4 was used for coding then, but as one use case among many.
How to Build a $2M/Month App (Copy Me)
GPT Image 2 keeps text consistent, which helps with text-wall ads. It is used first to build a scene before turning it into video.
Tony Fadell: How to build real taste (and why AI makes it matter more)
Tony says it began as a viral technology demo and only now builds product teams and product marketing. He compares it to Netscape, where people bought the tool and then asked what to use it for.
Why Most Apps Are Failing in 2026: 7 Mistakes Killing Your App Success
Some of Jaime's developers prefer it because it is slower but more detailed on the code itself.
AI era skills: Why cultivating agency matters more than job titles | Max Schoening (Notion)
PMs and designers at Notion use it in the terminal, and it lets you publish the work of you plus the AI for bragging rights.
Head of Claude Code: What happens after coding is solved | Boris Cherny
Cherny has barely used it and found it looked a lot like Claude Code, which he found flattering. The team focuses on user feedback rather than rivals.
The rise of the professional vibe coder (a new AI-era job)
Lazar exports a build to GitHub and imports it into Codex for diagnosis only, then applies the fix in Lovable. He does not let it change code because he doesn't know its agent well enough to steer it.
How Speak Built a $1B AI Language Learning App
Speak put GPT-4 into production three or four months before ChatGPT launched, testing it with 30 power users.
How a Meta PM ships products without ever writing code | Zevi Arnovitz
Pictured as the silent, hoodie-wearing best coder in the company. He uses it for the worst bugs and branch reviews; it takes much longer but returns fixed. His starting point: a CTO project gave him a safe place to plan and learn, and its prompt became the backbone of his /commands. Plain ChatGPT was a people-pleaser that agreed Bun was the same as Zustand, and its memory mixed up his running, work and studies.
he had nothing.. until he built a stupid simple $30k/mo app with AI
He built on GPT-3 (Da Vinci 2) before ChatGPT existed, and commercial use required a Google form with a wait of a couple of weeks. ChatGPT also helps estimate feature complexity and cost for outsourced builds.
The new AI growth playbook for 2026 | How Lovable hit $200M ARR in one year
Lost market share when Gemini 3 launched, showing that product-market fit has to be recaptured every few months. It has had to reach adjacent users because of its size, unlike most AI companies.
Why AI is disrupting traditional product management | Tomer Cohen (LinkedIn CPO)
LinkedIn's knowledge agents run on ChatGPT Enterprise alongside Copilot Enterprise, with heavy customization on top that produced the biggest gain.
$1M to $10M: The enterprise sales playbook with Jen Abel
Customers brought their own use cases, so it connected to nothing, which made enterprise entry easy. Jen heard it spoke to CTOs well before launch to explain where it was heading.
How OpenArt is Using AI to Reinvent Content Creation
Playing with DALL-E 2 in August 2022 sparked the OpenArt idea. OpenArt integrated it alongside other models, becoming an aggregator. Shows how fast models are closing in on storytelling, since it synthesizes audio and video well enough that a feed of 10-second AI clips can pass for real. It raises the question of what app companies must offer beyond the model.
Google Part III: The AI Company. Google is amazingly well-positioned... will they win in AI?
Founded partly to pull talent from Google, it bet on pre-training transformers and took $1 billion from Microsoft in 2018. ChatGPT launched in November 2022 and hit 30 million users by year end, 100 million in two months.
Why AI evals are the hottest new skill for product builders | Hamel Husain & Shreya Shankar
Its chief product officer calls evals the most important new skill for product builders. It mines Twitter and Reddit complaints and tied them to products, and its Statsig acquisition shows evals and A/B testing are the same data science.
He saved OpenAI, invented the “Like” button, and built Google Maps: Bret Taylor (Sierra)
Taylor's kids use ChatGPT to quiz themselves before tests, and he sees ChatGPT as closer to Google search than to a phone. He is excited about agents like Codex.
The AI-native startup: 5 products, 7-figure revenue, 100% AI-written code. | Dan Shipper (Every)
Every uses it for self-contained, one-off features alongside Claude Code.
AI prompt engineering in 2025: What works and what doesn’t | Sander Schulhoff
It sponsored HackAPrompt from the start and has cited the dataset in five recent publications. ChatGPT answered a bomb question that was translated to Spanish and Base64 encoded a month before the interview. In a medical coding project, pasting coded examples with reasoning lifted its accuracy by about 70% from near zero. Run on thousands of inputs, it wrote out reasoning 99 times in 100, so a phrase was added to force it.
How Tolan Built a Viral AI Companion
Model price cuts, sometimes around 80% at a dev day, lift margins for apps like Tolan. Mehta uses its o3 model for research, writing and creative work.
How Notion’s Community Helped it Win in AI
Notion's founders got early access to its models. A GPT release created a huge inflection point that convinced them AI would change the industry.
How AllTrails Crossed the Chasm to 1M+ Subscribers
AllTrails was a launch partner for OpenAI's GPT store a year earlier, but it has not replaced Google as a traffic source. AllTrails also uses a GPT review summary on trail pages.
Behind the product: Replit | Amjad Masad (co-founder and CEO)
Its models are used alongside Claude Sonnet in Replit's multi-model agent setup. Some users also get unstuck by asking ChatGPT.
Becoming an AI PM | Aman Khan (Arize AI, ex-Spotify, Apple, Cruise)
Trying the voice experience in ChatGPT can give a PM a light-bulb moment that it feels early but might work for their team's customer support.
How LinkedIn became interesting: The inside story | Tomer Cohen (CPO at LinkedIn)
LinkedIn built prompt engineering know-how ahead of the market and shared its learnings with OpenAI and Microsoft.
The future of AI in software development | Inbal Shani (CPO of GitHub)
GitHub engineers worked with designers and OpenAI GPT in 2020 to build Copilot. OpenAI models were brought in only after GitHub decided which coding work to automate.
Brian Chesky’s new playbook
Brian advised Sam Altman that OpenAI also needed product, design, marketing, leadership and sales beyond research, and Altman took some of the advice and discarded the rest.
What AI means for your product strategy | Paul Adams (CPO of Intercom)
Intercom adapted its models for customer support, and Adams reads its blog and papers to keep up.
NVIDIA CEO Jensen Huang
Huang knew its founders and hand delivered the first DGX to OpenAI, wanting them to have the computer Nvidia was building.
How to become a category pirate | Christopher Lochhead (Author of Play Bigger, Niche Down, more)
It is cited as today's example of winning by creating category language, since few people said large language model not long ago.
Nvidia Part III: The Dawn of the AI Era (2022-2023)
Founded to break the Google and Facebook AI duopoly, it became a capped-profit entity in March 2019 to pay for compute. ChatGPT launched November 30, 2022 and reached 100 million users faster than any app, and the company is rumored to be past a $1 billion run rate.
Mastering paid growth | Jonathan Becker (Thrive Digital)
Used with Midjourney for quick creative mock-ups that can be re-prompted live on client calls from feedback.
The role of AI in new product development | Ryan J. Salva (VP of Product at GitHub)
Its heavy cloning of GitHub repos looked like a denial-of-service attack until traced, which led GitHub to share a packaged snapshot of public code. Its researchers experimented with models and tossed them back to GitHub to test.
More in AI assistants
AI assistantsChatGPT
OpenAI's AI assistant app used for chat, plugins and agent work.
51 sources · 102 tips
AI assistantsAnthropic
An AI company that builds the Claude models and products such as Claude chat and Cowork.
25 sources · 71 tips
AI assistantsClaude
Anthropic's AI assistant and model family, used for chat, coding and agentic work.
19 sources · 35 tips
AI assistantsElevenLabs
A voice AI company with text-to-speech, dubbing, speech-to-text and music models, plus consumer and creator apps.
6 sources · 37 tips
AI assistantsGemini
Google's AI assistant app, available as a chatbot and moving toward a personal agent.
11 sources · 13 tips
AI assistantsNotebookLM
A Google Labs app that answers questions about your own sources and turns them into AI-hosted audio shows.
2 sources · 18 tips
AI assistantsTolan
A voice-first AI companion app with an animated alien character that remembers your conversations.
2 sources · 17 tipsGenie
A ChatGPT-powered chat app that writes marketing copy, taglines, tweets and app store descriptions from short prompts.
3 sources · 13 tips