ChatGPT

Tiers and credits cover AI's usage-driven compute costs. ChatGPT pairs a $20 a month plan with a $200 Pro plan and sells credits, and its free-tier lockouts trained users to expect usage limits. Tolan copied the split with Light, Plus and Unlimited, and hundreds or thousands pay $200 a month for Unlimited.

Phil Carter's pattern: non-AI features and a taste of AI in a basic plan, full AI usage in a pro or max plan, plus paid credits beyond the cap. $20 to $200 a month plans would have been rejected three years ago.

Lenny's Podcast▶ 27:46

Compete on what a chat box can't do, not on persona. Tolan focuses on emotional connection, and many of its users also use ChatGPT for work emails. Purpose says a celebrity name plus content is no different from asking ChatGPT to act like that person, so it adds validated frameworks.

About 100 million people use ChatGPT for life advice every month. Raj Singh calls a vanilla chat box a commodity.

Lenny's Podcast▶ 19:58

Users should type a task and never pick a mode or model. The north star is that ChatGPT's billion users type a task and the system picks the right harness and model. Chat, work mode and Codex differ mostly in UI: work mode runs on Codex underneath with the coding UI removed, and all of Codex's capabilities will eventually ship in ChatGPT.

Feedback showed people liked work's capability but sometimes wanted the speed of chat, so the chat and work toggle is being merged. Tibo Sottiaux says even he tires of model pickers: "you kind of need a PhD in model pickers." Ian Silber's design goal is one input that knows whether to answer fast, chat or go off and do work, with advanced options still exposed for those who want them.

Lenny's Podcast▶ 6:28

Plugins earn recommendations through retention and quality, not wording. ChatGPT recommends plugins inside conversations based on retention numbers and plugin quality. A strong plugin can reach a significant slice of about 1.2 billion users, and a weak one stops being recommended.

Getting people to write about you does not help; the test is whether the plugin lets ChatGPT do something new. "Obviously, if your plugin is not quite good, they will stop being recommended."

Lenny's Podcast▶ 21:25

Partners get paid when subscribers spend usage in their products. When subscribers spend their ChatGPT usage through a plugin like Notion or Figma, or a product using sign-in with ChatGPT, the partner gets a revenue share. Sign-in with ChatGPT already has 16 partners.

It began as a virtual handshake with the creators of Pi and OpenCode, who agreed to use the sign-in and trusted each other not to do anything shady. Revenue sharing was not mentioned in the keynote.

Lenny's Podcast▶ 1:27

The agent mode toggle was a stepping stone users could follow. OpenAI shipped an admittedly imperfect agent mode toggle so over a billion ChatGPT users could get an agentic harness without disrupting developer workflows. The toggle is expected to go away within a year.

Tara Seshan's view is that users accept turns in a product when they can follow a coherent story and see what is happening behind the scenes, so a team can deprecate a lot without burning bridges. The obvious next step after the toggle is no toggle.

Lenny's Podcast▶ 1:03:36

Ship done before perfect when the product is transformative. The ChatGPT Work launch drew confusion comments, and the team iterated quickly in response. Tara Seshan, who came from Stripe where every enum name was weighed, says nothing compares to watching real users try a feature.

Her fix for the urge to over-plan is to name the one most important thing to test and run the fastest possible test. Enterprises say they cannot absorb the pace, but the team builds what users need, not what they say they need.

Lenny's Podcast▶ 56:34

Build for where models will be in two to three months. Tara Seshan says building for today's models fails and building for a model a year away fails too, and both are equally wrong. OpenAI plans in 60 to 90 day horizons because its market cannot be modeled like Stripe's payments market.

Andrew Ambrosino is confident the Codex app would have failed if released in November instead of February, and the only difference was the models. The two-to-three-month rule rests on the speakers' experience, not data.

Lenny's Podcast

Capabilities stack in layers, with computer use as the backstop. ChatGPT treats native features, first and third party plugins and computer use as an array of tools exposed through shared hooks. If a plugin or MCP is missing, computer use still finishes the task.

Nan Yu says an agent that gets 99% of the way and then fails is worse than one that never starts. Computer use may be slow and cost many tokens, but "it will 100% do the job for you."

Lenny's Podcast▶ 24:30

Context-aware elements help users express intent without prompt engineering. ChatGPT ships tap-through follow-up questions for open-ended asks and writing blocks that hold an email or letter you edit directly. Writing blocks made drafting much better than plain chat text.

Image generation gets its own tools, and the interface should change based on who the user is and what they are doing. Nan Yu expects products to drive themselves to give a gentle on-ramp to users stuck at an empty input box.

Lenny's Podcast▶ 59:00

Users should type a task and never pick a mode or model. The north star is that ChatGPT's billion users type a task and the system picks the right harness and model. Chat, work mode and Codex differ mostly in UI: work mode runs on Codex underneath with the coding UI removed, and all of Codex's capabilities will eventually ship in ChatGPT.

Feedback showed people liked work's capability but sometimes wanted the speed of chat, so the chat and work toggle is being merged. Tibo Sottiaux says even he tires of model pickers: "you kind of need a PhD in model pickers." Ian Silber's design goal is one input that knows whether to answer fast, chat or go off and do work, with advanced options still exposed for those who want them.

Lenny's Podcast▶ 58:00

Ship done before perfect when the product is transformative. The ChatGPT Work launch drew confusion comments, and the team iterated quickly in response. Tara Seshan, who came from Stripe where every enum name was weighed, says nothing compares to watching real users try a feature.

Her fix for the urge to over-plan is to name the one most important thing to test and run the fastest possible test. Enterprises say they cannot absorb the pace, but the team builds what users need, not what they say they need.

Lenny's Podcast▶ 36:15

Build for where models will be in two to three months. Tara Seshan says building for today's models fails and building for a model a year away fails too, and both are equally wrong. OpenAI plans in 60 to 90 day horizons because its market cannot be modeled like Stripe's payments market.

Andrew Ambrosino is confident the Codex app would have failed if released in November instead of February, and the only difference was the models. The two-to-three-month rule rests on the speakers' experience, not data.

Lenny's Podcast

Knowledge-work agents must show process and citations, not just output. Code can be checked with tests, but a final deck number cannot be trusted alone. Building ChatGPT Work taught the team to expose in-progress work, citations and reasoning.

"I can't simply look at the deck in the end and see the numbers."

Lenny's Podcast▶ 23:05

OpenAI went all in on ChatGPT when insiders wanted to pivot. About two months after launch, growth looked unstable and many at OpenAI wanted to focus on five or six other things. Peter Thiel told Sam Altman it was an obvious mistake to do anything else, because the power was a Google-style text box that does the right thing for any input.

Altman went hard on ChatGPT even though it had no feeds, network effects or lock-in. It reached about a billion users in under four years. Thiel's words: "It's an obvious mistake to do anything but this."

Lenny's Podcast▶ 57:00

Users should type a task and never pick a mode or model. The north star is that ChatGPT's billion users type a task and the system picks the right harness and model. Chat, work mode and Codex differ mostly in UI: work mode runs on Codex underneath with the coding UI removed, and all of Codex's capabilities will eventually ship in ChatGPT.

Feedback showed people liked work's capability but sometimes wanted the speed of chat, so the chat and work toggle is being merged. Tibo Sottiaux says even he tires of model pickers: "you kind of need a PhD in model pickers." Ian Silber's design goal is one input that knows whether to answer fast, chat or go off and do work, with advanced options still exposed for those who want them.

Lenny's Podcast

Altman argues deploying imperfect AI early is how it gets safe. Sam Altman says putting products into the world, getting feedback from real customers and seeing where they break makes them both good and safe, as accident reporting and postmortems did for aviation.

OpenAI ships imperfect models, studies failures and shares what it learns. "You're not going to solve it by disconnecting yourself from reality."

Lenny's Podcast▶ 48:09

OpenAI chose to be a platform, not to build every product. Altman wants one direct interface to AI and one API to build anything on top, with the best AI at every point of the cost and performance curve. He says OpenAI should not compete with its customers or try to subsume the economy.

The aim is 100 million new businesses and 8 billion people using it. ChatGPT, Codex and the API are being merged into one interface plus one API.

Lenny's Podcast▶ 20:44

Users should type a task and never pick a mode or model. The north star is that ChatGPT's billion users type a task and the system picks the right harness and model. Chat, work mode and Codex differ mostly in UI: work mode runs on Codex underneath with the coding UI removed, and all of Codex's capabilities will eventually ship in ChatGPT.

Feedback showed people liked work's capability but sometimes wanted the speed of chat, so the chat and work toggle is being merged. Tibo Sottiaux says even he tires of model pickers: "you kind of need a PhD in model pickers." Ian Silber's design goal is one input that knows whether to answer fast, chat or go off and do work, with advanced options still exposed for those who want them.

Lenny's Podcast▶ 29:24

Re-release ideas that failed only because the model wasn't ready. Operator, the agent in Atlas and the in-app browser in Codex are the same idea re-released with better intelligence, with very different outcomes. Old code is treated as an artifact to test against future models.

Ian Silber gives the same advice: early voice mode and computer use barely worked when he joined, and now OpenAI uses both to build software. "What was totally broken yesterday now actually just kind of works."

Lenny's Podcast▶ 25:42

Obsess only over durable features and move fast on the rest. For durable surfaces like the ChatGPT composer, the team tries 100 things, throws out 99 and ships one. Other work goes from idea to ship in about 4 hours and is built in public.

Heavy design effort goes where the technology is unlikely to change next month. Bigger bets like the 'personal life' vision get whiteboarding, Figma, prototypes and dogfooding first.

Lenny's Podcast▶ 16:25

Context-aware elements help users express intent without prompt engineering. ChatGPT ships tap-through follow-up questions for open-ended asks and writing blocks that hold an email or letter you edit directly. Writing blocks made drafting much better than plain chat text.

Image generation gets its own tools, and the interface should change based on who the user is and what they are doing. Nan Yu expects products to drive themselves to give a gentle on-ramp to users stuck at an empty input box.

Lenny's Podcast▶ 41:25

Give most users a simple default and test advanced features with power users. Usage runs from recipe ideas to automating a farm, and from free users to people paying hundreds of dollars. Most get a streamlined experience, while new things like the desktop app and Codex are tried first with people who care.

"We can experiment early with the people that really care."

Lenny's Podcast▶ 38:21

Proactive help should appear in context, not as notifications. If agents help thousands of times a day, push notifications become unbearable, so OpenAI favors mixed-initiative UX like autocomplete, where being wrong costs little. ChatGPT Pulse already gives daily updates on things users may care about.

The Atlas browser exists partly to control what the AI sees and give a place for contextual actions. Next steps include suggestions from calendar and Slack.

Lenny's Podcast▶ 29:24

Compete on what a chat box can't do, not on persona. Tolan focuses on emotional connection, and many of its users also use ChatGPT for work emails. Purpose says a celebrity name plus content is no different from asking ChatGPT to act like that person, so it adds validated frameworks.

About 100 million people use ChatGPT for life advice every month. Raj Singh calls a vanilla chat box a commodity.

Lenny's Podcast▶ 1:26:47

ChatGPT's $20 a month reset what consumers pay for apps. Before ChatGPT, consumers rarely paid more than about $60 a year for an app. Its $20 a month plan made premium pricing credible for apps that truly solve a problem.

Fewer payers at higher prices means less acquisition spend and infrastructure. Waking Up charges $20 a month and $150 a year and earns over $1M a month on about 25K downloads a month.

Lenny's Podcast▶ 21:51

Where users go despite the wrong surface shows what to generalize. Marketing, comms, finance and legal staff used the Codex app even though it showed them code and asked for approvals. OpenAI added Codex to the ChatGPT desktop app and the Atlas browser for them, but nobody left the Codex app, so the team generalized Codex itself.

Internal product-market fit on engineering and research work came before launch, then spread to non-technical teams. Codex has grown 6x since January to over 5 million weekly active users. A merge like this can still draw a 2,000-message internal Slack thread calling it stupid.

Lenny's Podcast▶ 28:57

Build for where models will be in two to three months. Tara Seshan says building for today's models fails and building for a model a year away fails too, and both are equally wrong. OpenAI plans in 60 to 90 day horizons because its market cannot be modeled like Stripe's payments market.

Andrew Ambrosino is confident the Codex app would have failed if released in November instead of February, and the only difference was the models. The two-to-three-month rule rests on the speakers' experience, not data.

Lenny's Podcast

Re-release ideas that failed only because the model wasn't ready. Operator, the agent in Atlas and the in-app browser in Codex are the same idea re-released with better intelligence, with very different outcomes. Old code is treated as an artifact to test against future models.

Ian Silber gives the same advice: early voice mode and computer use barely worked when he joined, and now OpenAI uses both to build software. "What was totally broken yesterday now actually just kind of works."

Lenny's Podcast▶ 29:29

Tiers and credits cover AI's usage-driven compute costs. ChatGPT pairs a $20 a month plan with a $200 Pro plan and sells credits, and its free-tier lockouts trained users to expect usage limits. Tolan copied the split with Light, Plus and Unlimited, and hundreds or thousands pay $200 a month for Unlimited.

Phil Carter's pattern: non-AI features and a taste of AI in a basic plan, full AI usage in a pro or max plan, plus paid credits beyond the cap. $20 to $200 a month plans would have been rejected three years ago.

Subversive▶ 47:04

Proactive help should appear in context, not as notifications. If agents help thousands of times a day, push notifications become unbearable, so OpenAI favors mixed-initiative UX like autocomplete, where being wrong costs little. ChatGPT Pulse already gives daily updates on things users may care about.

The Atlas browser exists partly to control what the AI sees and give a place for contextual actions. Next steps include suggestions from calendar and Slack.

Lenny's Podcast

Regenerations are implicit failure signals that should become evals. A user regenerating a ChatGPT answer is an implicit signal of a failure pattern. Evals guard known must-not-fail behavior, while production monitoring shows which traces to inspect at high volume.

When monitoring finds a failure mode you care about, it becomes a new eval. Codex pairs evals with A/B tests, user reactions and social media feedback.

Lenny's Podcast▶ 27:30

Most free users never see the paid features worth buying. Most free users never tried thinking pro or deep research. Jeff Morris suggests offering one deep research a month so free users see what the paid tier does.

Flighty uses a similar taste: the first flight is free with pro features, then premium locks. Ben Thompson argues ChatGPT should have leaned on ads from the start.

David Senra▶ 1:05:00

Chat is the front door, with specialist interfaces behind it. Alexander Embiricos sees chat as the right entry point when people don't yet know what to use a tool for, like messaging a teammate in Slack, while experts can open a GUI to work deep in the code.

Someone who uses ChatGPT outside work will simply ask it for help at work without learning every connector. He frames this as a one-year view, not a plan.

Superwall

Proactive help should appear in context, not as notifications. If agents help thousands of times a day, push notifications become unbearable, so OpenAI favors mixed-initiative UX like autocomplete, where being wrong costs little. ChatGPT Pulse already gives daily updates on things users may care about.

The Atlas browser exists partly to control what the AI sees and give a place for contextual actions. Next steps include suggestions from calendar and Slack.

Superwall▶ 49:16

Re-sign up as a new user and watch D7 retention. Alexander Embiricos keeps several ChatGPT Pro accounts at about $200 a month each to feel the sign-up experience from scratch. Early retention counters the pull to build only for power users.

"the feeling of being a user and the early retention stats are still like super important for us"

Superwall▶ 15:02

Validate ChatGPT's App Store keyword ideas against real search data. Steve P. Young found ChatGPT's keyword suggestions had literally no App Store search volume. Check each against an ASO tool's traffic score.

App Masters▶ 6:04