On Friday, November 17, 2023, around 12:30 PM San Francisco time, Sam Altman joined a Google Meet the board had requested. The call ran twenty-three minutes. When it ended, Altman was no longer CEO of OpenAI.
The board's statement was clipped: the directors had concluded Altman "had not been consistently candid in his communications." They didn't say what about. Inside the company, no one knew. Outside, nobody knew either.
What followed was a weekend that put on public display, for the first time, the contradiction OpenAI is built on. Monday the 20th, Microsoft — the main investor — announced Altman and Greg Brockman would join to lead a new internal lab. Tuesday the 21st, a letter signed by 702 of OpenAI's 770 employees threatened resignation if Altman didn't return. Wednesday the 22nd in the morning, the same board that had fired him on Friday agreed to bring him back. Of the four directors who voted to remove him, three stepped down over the following months.
Five days. That's the distance between "CEO removed for lack of candor" and "CEO reinstated with a board reset." That's the actual tempo of the company that made AI mainstream.
From 2015 to now, in paragraphs
OpenAI was founded in December 2015. Five co-founders signed the original letter: Sam Altman, Ilya Sutskever, Greg Brockman, Wojciech Zaremba, and John Schulman. Elon Musk and Reid Hoffman brought capital and sat on the initial board. The structure was a non-profit with a stated mission: build AI that benefits all of humanity.
In 2018, Musk left the board after an internal dispute. That same year Sutskever and Schulman started work on what would become GPT-2, and the team began to suspect that scaling the model was more powerful than inventing new architectures. Scaling, though, costs money that a non-profit can't raise.
The answer came in 2019 with a hybrid structure no one had tried before: the non-profit on top, a capped-profit subsidiary underneath, a ceiling on investor returns, and a clause saying the non-profit board can shut down the for-profit arm if the mission requires it. That same year Microsoft made its first large investment, a billion dollars. The investment kept growing: by 2023 it was reportedly ten billion, and analysts today put it above thirteen.
What that money bought: the ability to train GPT-3 in 2020, then ChatGPT in November 2022, then GPT-4 in March 2023. That stretch was OpenAI's golden age. For eighteen months, almost any conversation about AI anywhere in the world referenced something OpenAI had shipped that week.
What OpenAI sells today
Open ChatGPT in 2026 and what you see is a platform with several layers.
The chat models. The GPT-4o family is the default: fast, multimodal, cheap. It handles text, image, and audio in the same request. Above it sits o1 and o3, models that spend extra compute "thinking" before answering — better at math and complex reasoning at the price of latency (a hard problem can take half a minute).
DALL-E 3. Image generation built into ChatGPT. It remains one of the strongest general-purpose options in the market, with the edge of letting you refine the prompt conversationally in the same window.
Advanced Voice mode. Voice conversation with low latency and natural intonation. It's currently the closest commercial thing to "talking to an AI" that exists.
Canvas and SearchGPT. Canvas is a collaborative document editor where the model drafts and you edit side by side. SearchGPT is integrated web search. Both are signals of a strategy that stopped being "chatbot" and is increasingly "a suite of applications."
Distribution. GPT runs inside Microsoft Copilot for Office, inside GitHub Copilot, inside Slack AI, inside Figma, inside hundreds of API integrations. OpenAI doesn't win only because the product is good: it wins because it's already where you work, without you opening a new tab.
The question that defines OpenAI in 2026
If I had to pick one question to understand where this company is headed, it would be this: can a company hold a stated safety mission while the business model depends on growing subscriptions and shipping capability faster every quarter?
In May 2024 Ilya Sutskever, co-founder and chief scientist, resigned. Two days later, Jan Leike, co-lead of the long-term alignment team, posted a thread on X saying he'd left OpenAI because "safety culture and processes have taken a backseat to shiny products." The superalignment team was formally dissolved shortly after.
Sutskever founded Safe Superintelligence Inc. Leike went to Anthropic. Neither move was incidental.
None of this means OpenAI is unfit to use. It means the tension that produced the Altman firing weekend and the Sutskever-Leike exit isn't resolved: it's still the rubber band stretching every future decision.
Which tools are you using today that you weren't two years ago? If the answer has shifted, you're describing exactly the move underway in the market. The useful next step is to understand how AI models are measured so you don't pick by marketing, and to read the AI race for the full map without fandom.
Ask anyone who doesn't work in tech when they first heard about artificial intelligence. You'll get an anecdote back. "A friend showed me something on her phone." "My kid told me to try this." Almost certainly, the anecdote took place between December 2022 and February 2023.
That winter, something odd happened. In two months, a hundred million people created a ChatGPT account. For scale: TikTok took nine months to reach that number. Instagram took two and a half years.
ChatGPT didn't grow that way because it was technically better than everything before it. It grew that way because someone decided to put a language model behind a text box, free, on the internet, with no waitlist. That someone was OpenAI.
The company that bet on the open door
OpenAI started in 2015 in San Francisco. Sam Altman, Ilya Sutskever, Greg Brockman, and a small group of researchers founded it. Elon Musk put in money and walked away in 2018. The original structure was a non-profit with a stated mission: build AI that benefits all of humanity.
In 2019 they needed real money to train large models and Microsoft came in with an investment that grew over the years into the billions. With that money they trained GPT-3 in 2020 — technically striking but only useful for developers, since it shipped as an API.
The clever play came in November 2022. They took an improved version (GPT-3.5), wrapped it in a chat interface, and opened the doors for free. No waitlist. No technical explanation. Just a text box.
What it's winning, what it's paying
OpenAI won mindshare. Today, when a regular person says "the AI," they're almost always thinking of ChatGPT. That cultural lead is real.
It also has the widest product line: DALL-E for images, Voice mode for talking, Canvas for drafting documents, and it's plugged into Microsoft Office, Slack, and hundreds of applications.
What it's paying: a recent history of turbulent governance. In November 2023 the board fired Sam Altman on a Friday; he was back on Tuesday because nearly the whole team threatened to leave with him. In May 2024 the two leads of the long-term alignment team left OpenAI, one of them (Jan Leike) posting publicly that the safety culture had taken a back seat to shiny products.
What to take away
Three things worth holding onto:
- OpenAI invented the moment, not the technology. The transformer architecture is Google's, pre-training is older than either. What OpenAI did was take all of that and put it within reach of anyone. That gesture changed how we think about AI.
- Going first has upsides and costs. ChatGPT won the cultural distribution. It also carries every debut mistake — hype, talent departures, public boardroom fights. Competitors who came later (Anthropic among them) learned from it.
- For everyday use ChatGPT is a fine choice; for high-stakes work, not always. If you need AI to help with an important contract or financial analysis, the more serious companies picked other tools. If you need to explore ideas, generate images, or have a voice conversation, ChatGPT does the job.
On Friday, November 17, 2023, around 12:30 PM San Francisco time, Sam Altman joined a Google Meet the board had requested. The call ran twenty-three minutes. When it ended, Altman was no longer CEO of OpenAI.
The board's statement was clipped: the directors had concluded Altman "had not been consistently candid in his communications." They didn't say what about. Inside the company, no one knew. Outside, nobody knew either.
What followed was a weekend that put on public display, for the first time, the contradiction OpenAI is built on. Monday the 20th, Microsoft — the main investor — announced Altman and Greg Brockman would join to lead a new internal lab. Tuesday the 21st, a letter signed by 702 of OpenAI's 770 employees threatened resignation if Altman didn't return. Wednesday the 22nd in the morning, the same board that had fired him on Friday agreed to bring him back. Of the four directors who voted to remove him, three stepped down over the following months.
Five days. That's the distance between "CEO removed for lack of candor" and "CEO reinstated with a board reset." That's the actual tempo of the company that made AI mainstream.
From 2015 to now, in paragraphs
OpenAI was founded in December 2015. Five co-founders signed the original letter: Sam Altman, Ilya Sutskever, Greg Brockman, Wojciech Zaremba, and John Schulman. Elon Musk and Reid Hoffman brought capital and sat on the initial board. The structure was a non-profit with a stated mission: build AI that benefits all of humanity.
In 2018, Musk left the board after an internal dispute. That same year Sutskever and Schulman started work on what would become GPT-2, and the team began to suspect that scaling the model was more powerful than inventing new architectures. Scaling, though, costs money that a non-profit can't raise.
The answer came in 2019 with a hybrid structure no one had tried before: the non-profit on top, a capped-profit subsidiary underneath, a ceiling on investor returns, and a clause saying the non-profit board can shut down the for-profit arm if the mission requires it. That same year Microsoft made its first large investment, a billion dollars. The investment kept growing: by 2023 it was reportedly ten billion, and analysts today put it above thirteen.
What that money bought: the ability to train GPT-3 in 2020, then ChatGPT in November 2022, then GPT-4 in March 2023. That stretch was OpenAI's golden age. For eighteen months, almost any conversation about AI anywhere in the world referenced something OpenAI had shipped that week.
What OpenAI sells today
Open ChatGPT in 2026 and what you see is a platform with several layers.
The chat models. The GPT-4o family is the default: fast, multimodal, cheap. It handles text, image, and audio in the same request. Above it sits o1 and o3, models that spend extra compute "thinking" before answering — better at math and complex reasoning at the price of latency (a hard problem can take half a minute).
DALL-E 3. Image generation built into ChatGPT. It remains one of the strongest general-purpose options in the market, with the edge of letting you refine the prompt conversationally in the same window.
Advanced Voice mode. Voice conversation with low latency and natural intonation. It's currently the closest commercial thing to "talking to an AI" that exists.
Canvas and SearchGPT. Canvas is a collaborative document editor where the model drafts and you edit side by side. SearchGPT is integrated web search. Both are signals of a strategy that stopped being "chatbot" and is increasingly "a suite of applications."
Distribution. GPT runs inside Microsoft Copilot for Office, inside GitHub Copilot, inside Slack AI, inside Figma, inside hundreds of API integrations. OpenAI doesn't win only because the product is good: it wins because it's already where you work, without you opening a new tab.
The question that defines OpenAI in 2026
If I had to pick one question to understand where this company is headed, it would be this: can a company hold a stated safety mission while the business model depends on growing subscriptions and shipping capability faster every quarter?
In May 2024 Ilya Sutskever, co-founder and chief scientist, resigned. Two days later, Jan Leike, co-lead of the long-term alignment team, posted a thread on X saying he'd left OpenAI because "safety culture and processes have taken a backseat to shiny products." The superalignment team was formally dissolved shortly after.
Sutskever founded Safe Superintelligence Inc. Leike went to Anthropic. Neither move was incidental.
None of this means OpenAI is unfit to use. It means the tension that produced the Altman firing weekend and the Sutskever-Leike exit isn't resolved: it's still the rubber band stretching every future decision.
Which tools are you using today that you weren't two years ago? If the answer has shifted, you're describing exactly the move underway in the market. The useful next step is to understand how AI models are measured so you don't pick by marketing, and to read the AI race for the full map without fandom.
On May 14, 2024, Ilya Sutskever posted a two-paragraph message on X saying he was leaving OpenAI after nearly a decade. He was the chief scientist, a co-founder, the person the academic establishment had considered for years to be the company's technical brain. His message was polite, thanked Altman and Brockman by name, and offered no reason.
Two days later, Jan Leike — the other co-lead of the superalignment team, in charge of the internal program OpenAI had announced in July 2023 to commit 20 percent of its compute to long-term alignment research — published a different thread. Nineteen posts. The sentence that ended up quoted everywhere: "safety culture and processes have taken a backseat to shiny products." Leike said he'd been fighting for resources for months. That superalignment had been "sailing against the wind." That he couldn't keep doing it.
In the following weeks, the superalignment team was formally dissolved. Its mandate was absorbed by other groups. Sutskever announced his new company on June 19, Safe Superintelligence Inc., with offices in Palo Alto and Tel Aviv. Leike joined Anthropic the same week.
That sequence — the co-founder scientist's resignation, the alignment co-lead's public thread, the team dissolution, the move to Anthropic — can't be read as an isolated episode. It's the most visible symptom of a structural tension OpenAI has been carrying since its business model changed.
Technical genealogy: from GPT-1 to o3
Understanding OpenAI technically means tracing the model sequence with the right lens. The point isn't parameter counts. The point is the kind of bet each release represents.
GPT-1 (June 2018) was a proof of concept built on the Transformer architecture from Vaswani et al (2017). 117 million parameters. Unsupervised pre-training on BookCorpus, supervised fine-tuning on NLU tasks. The result didn't shake the world, but it internally validated that the pre-train + fine-tune paradigm worked.
GPT-2 (February 2019) stepped up to 1.5 billion parameters. Fluent generation, paragraph-level coherence. OpenAI delayed the full model release over misuse concerns — the first public gesture of a "staged release" policy that later became a pattern.
GPT-3 (May 2020) reached 175 billion parameters and demonstrated emergent few-shot learning: the ability to pick up a new task from a few in-context examples without any fine-tuning. That finding reshaped the academic conversation. It wasn't that "bigger models are better"; it was that some capabilities appear only above specific scale thresholds.
InstructGPT (January 2022) introduced RLHF — Reinforcement Learning from Human Feedback — as a preference fine-tuning technique. Humans rank responses, a reward model learns the rankings, and the main model is optimized against that reward. This turned language models into instruction models: they no longer just predicted the next likely token, they predicted the next token a human would prefer.
ChatGPT (November 2022) was GPT-3.5 (an optimized version of InstructGPT) behind a chat interface. There was no technical breakthrough. There was a packaging breakthrough.
GPT-4 (March 2023) was multimodal (text + image input), 8K-32K token context, better benchmark reasoning. Internal architecture remains speculative — OpenAI stopped publishing exact technical details after GPT-3. Independent analyses suggest a Mixture-of-Experts structure.
GPT-4o (May 2024) is "omni": natively multimodal across text, image, and audio, with reduced processing latency and lower cost. It's the default model for most users.
o1 and o3 (2024-2025) represent an important philosophical fork: models trained to "think" longer before responding, spending extra inference compute on an internal reasoning chain before producing the final answer. On Olympic math benchmarks and competitive programming, o3 approaches human elite levels. The cost is latency: hard problems can take tens of seconds.
The pattern: OpenAI has been sharp at identifying when the frontier was in raw scale (GPT-3), when it was in human preference alignment (InstructGPT), when it was in cheap multimodality (GPT-4o), and when it was in inference-time reasoning (o1/o3). Every pivot was technically right at its moment.
Distribution as moat
What many outside analyses underweight is that OpenAI doesn't win on model quality, it wins on distribution. GPT runs inside Microsoft 365 Copilot, which has hundreds of millions of installed corporate seats. It runs inside GitHub Copilot, which crossed a million paid subscribers in 2024. It runs inside Slack AI, Figma, Notion, and an ecosystem of plugins and custom GPTs. The OpenAI API is the de facto first choice for startups building on top of AI — it's the AWS of models, with all the downstream lock-in that implies for keeping developers in the stack.
The most interesting move in this direction was the "custom GPTs" launched in November 2023: a platform of configurable agents where non-technical users can build a specialized assistant without code. That play turned ChatGPT from "application" into "platform" — and platforms capture more value over time than applications do.
The structural tension
The governance structure OpenAI invented in 2019 made sense in a world where the central question was "how do we finance ever-larger models without betraying the mission." In that world, the answer — non-profit controlling a capped-return subsidiary — was elegant.
The problem is that after ChatGPT the business model changed. The pressure no longer comes from training the next model; it comes from retaining and growing a base of tens of millions of paying subscribers. That shifts internal priorities. Release cycles get shorter. Investment in long-term alignment research competes — inside the same compute budget — with investment in the shiny features that justify next month's subscription.
The Sutskever-Leike episode is the first public symptom that this new regime isn't compatible with the version of the mission the non-profit signed in 2015. The board's thesis in November 2023 wasn't wrong on the diagnosis; it was premature in execution and lacked operational backing. Five days later, the regime they wanted to correct came out stronger than before.
My editorial read
OpenAI will keep dominating the mass-consumer segment for a good portion of the next decade. The mindshare, the Microsoft distribution, and the iteration speed are hard to replicate for that use case.
But the professional high-stakes market — legal, financial, medical, production code, contracts, strategy — is going to fragment. Anthropic and Google are capturing growing slices of that segment, and the migration of alignment talent from OpenAI to Anthropic over the past year suggests the long-term bet on who produces the most reliable models for high-stakes use is shifting.
What's true is that any "which AI is better" conversation in 2026 makes no sense without specifying what for. OpenAI invented the category. Being the one who flips the light switch carries a large upside: everyone knows your name. And it carries a structural downside: once the first enthusiasm fades and the room is lit, everyone can also see the mess.
