The Launch That Was Meant to Be a Research Preview

ChatGPT was released on 30 November 2022, described by OpenAI as a research preview, built on a model in the GPT-3.5 series that was already available through an API. The capability was roughly two years old and the instruction-following technique was ten months old. What changed on that day was that it had a text box and no bill attached.

What was launched, and on what

The announcement post was titled ChatGPT: Optimizing Language Models for Dialogue. Two statements in it are the ones that matter and both are usually dropped from the retelling.

• It was explicitly a research preview. OpenAI said it was making the system available to gather feedback on its strengths and weaknesses, and it was free during the preview.
• It was described as a sibling model to InstructGPT, trained with the same methods with differences in the data collection setup. InstructGPT had been published in March 2022 and was already the default in OpenAI’s API. The training technique was not new on launch day.

The post also stated the limitations in the announcement itself: that the system sometimes writes plausible-sounding but incorrect answers, that it is sensitive to phrasing, that it is often excessively verbose, and that it will guess at ambiguous queries rather than ask. Everything the following two years of commentary treated as a discovery was in the launch post.

What was actually new

The GPT-3 API had been available since June 2020. InstructGPT models had been in it since January 2022. The technical delta on 30 November was small. The product delta was enormous, and it decomposes into four things:

1. Conversation as the interface. The API required you to compose a prompt that a completion engine would continue. The chat interface let you ask, and it kept the previous turns in context so a follow-up worked. This converts a tool that requires a mental model of the machine into one that requires none.
2. No barrier to entry. Free, no API key, no billing details, no waitlist after the first days, no code. The audience went from people who could write a Python request to everybody.
3. The model was tuned to refuse and to hedge. That was widely disliked and it is a large part of why the launch was survivable. A raw completion model exposed to millions of people would have produced a very different first week.
4. Streaming output. Watching text appear a token at a time reads as thinking. It is a rendering decision with no effect on the answer, and its effect on how the product felt was substantial.

None of these is a research contribution. All four are the difference between a capability existing and a capability being used, and that is the actual subject of this page.

The growth numbers, with their sources

Two figures are quoted constantly and they have very different provenance. Keeping them straight is the point of this section.

The second figure is not obviously wrong and it is not a company statement, and the difference matters because it has been repeated for years as though OpenAI had announced it. It also did not stand for long: Meta’s Threads reached 100 million sign-ups within five days of its launch in July 2023. Whether that beats ChatGPT depends entirely on whether you count sign-ups or monthly actives, and on whether you count a product launched into an existing billion-user network as a comparable case. “Fastest-growing consumer app in history” is a claim about a definition rather than about a fact.

How unplanned it was

The strongest evidence that the launch was not expected to be a product event is in the framing itself: a research preview, free, without a pricing page, without a mobile app — the iOS app did not arrive until May 2023 — and on a model OpenAI had not considered significant enough to give a version number of its own.

Beyond that, OpenAI staff including Sam Altman have said publicly in the years since that internal expectations were low and that the launch was assembled quickly as a lower-stakes way of gathering feedback than releasing a more capable model. That is a paraphrase of statements made in interviews after the fact rather than a quotation, and retrospective accounts of a success are the least reliable category of historical evidence, so it is offered as what the participants say rather than as established fact. GPT-4 was released on 14 March 2023, roughly three and a half months later, which is consistent with the more capable model having been substantially finished at the time of the ChatGPT launch.

How a research preview became a product

The transition is visible in the release dates, and it took about six months from a free experiment with no revenue model to a platform.

Two of those dates are worth arguing about. The subscription in February meant the free tier had to be preserved, which fixed the economics of serving an enormous number of non-paying users into the product from the beginning. And the API price cut in March is the event with the largest downstream effect: the entire application layer that grew during 2023 is a consequence of a per-token price, not of a capability release.

The hundred days after

1. December 2022. Stack Overflow bans ChatGPT-generated answers within days, on the grounds that they are frequently wrong and cheap to produce — the first institutional response, and a precise statement of the problem that has dominated since.
2. January 2023. School districts and universities begin blocking it; the detection industry appears and is unreliable from the outset.
3. 6 February 2023. Google announces Bard. Its promotional material contains a factual error about which telescope took the first image of an exoplanet, and Alphabet’s share price fell sharply — around seven per cent — on 8 February. The causal link is the market’s inference rather than a demonstrated one, but the sequence is well documented.
4. 7 February 2023. Microsoft launches Bing Chat. Its initial version, which users found could be pushed into long unstable exchanges, was constrained within about a fortnight by limiting conversation length — an early and instructive demonstration that alignment behaviour degrades over a long context.
5. 1 March 2023. The ChatGPT API is released at a price roughly a tenth of the previous generation’s, which is what moved the technology from experiment to line item for a large number of companies.
6. 14 March 2023. GPT-4, released with a technical report that explicitly declines to disclose architecture, parameter count, training data or training compute — the formal end of the disclosure norms that had held since GPT-1.

The lesson, which is about distribution

The reason this launch is worth studying is not that a capable model appeared. It is that a capable model had been publicly available for two and a half years and almost nobody had used it, and then the same capability behind a text box reached a hundred million people in a quarter.

Every chatbot on the long list of chatbots before ChatGPT had the interface and lacked the capability. GPT-3 had the capability and lacked the interface. The launch is the point at which the two arrived in the same product, and the growth curve measures the size of the gap that had been sitting there unexploited. If there is a transferable lesson, it is that the interval between a capability being available and it being used can be years, and that the interval is a product problem rather than a research one.