ChatGPT launched publicly on November 30, 2022, as a free research preview at chat.openai.com. It was built on the GPT-3.5 series of large language models, wrapped in a conversational interface that remembered context within a session. The service reached roughly one million users in five days, a adoption curve that pressured every major technology company to ship a competing chatbot within months.

That date is the short answer. The fuller answer involves a model lineage stretching back to 2018, a quieter API release months earlier, a paid tier in February 2023, and a mobile app in May 2023. Each milestone changed who could use the system and what it could do. Understanding the sequence explains why the November 2022 date became the cultural marker while the underlying technology had been publicly available, in less accessible form, for years.

The mechanism: how a language model becomes a chatbot

A base language model is a next-token predictor. It reads a sequence of text and outputs a probability distribution over the next token, then repeats. Trained on a large corpus, it learns grammar, facts, and stylistic patterns as side effects of compression. What it does not learn, by default, is how to hold a conversation, follow an instruction, or decline a harmful request.

ChatGPT's contribution was not a new model architecture. It was a training pipeline layered on top of an existing model. The pipeline has three stages. First, supervised fine-tuning: human contractors wrote demonstration conversations, and the model was trained to imitate them. Second, a reward model: contractors ranked multiple model outputs from best to worst, and a separate network learned to predict those rankings. Third, reinforcement learning from human feedback, or RLHF: the model was optimized to produce outputs the reward model scored highly, with a penalty for drifting too far from its starting behavior.

The practical effect is a system that behaves like an assistant rather than a document completer. It answers in prose, asks clarifying questions, and refuses some requests. The tradeoff is that RLHF optimizes for human preference, not factual accuracy. A response that sounds confident and well-organized can still be wrong, and the reward model has no independent knowledge to check against. This is the mechanical root of hallucination, and it existed on day one.

A second design choice mattered just as much: the chat interface kept the conversation history in the prompt. Every reply was conditioned on everything said before it in that session. This made multi-turn reasoning possible and also created the failure mode where a model commits to an early error and defends it for the rest of the conversation.

In plain terms, ChatGPT was a familiar engine with a new steering system. The engine existed. The steering is what shipped in November 2022.

The release timeline in dated milestones

Date Milestone What changed
June 2018 GPT-1 paper published 117 million parameters; demonstrated pretraining plus fine-tuning
February 2019 GPT-2 announced, released in stages 1.5 billion parameters; staged release over concerns about misuse
May 2020 GPT-3 paper and API 175 billion parameters; text generation available to developers, not the public
November 30, 2022 ChatGPT research preview Free web chat on GPT-3.5; the public launch date
February 1, 2023 ChatGPT Plus USD 20 monthly subscription for faster access and priority during peak load
March 14, 2023 GPT-4 announced Multimodal input, improved reasoning, available to Plus subscribers and via API
May 18, 2023 iOS app released Mobile access; Android followed in July 2023

The gap between GPT-3 in 2020 and ChatGPT in 2022 is the part most timelines compress. For roughly two and a half years, the same underlying capability was reachable only through an API that required programming and a paid account. The November 2022 event was a distribution event, not a capability event. That distinction matters when people ask why the technology seemed to appear overnight.

The February 2023 Plus tier is the second most consequential date after the launch itself. It established the consumer subscription model that competitors copied, and it introduced a two-tier access system where free users hit capacity limits during demand spikes. The USD 20 price point became a de facto industry anchor.

Quantified context: adoption speed and capacity limits

The adoption numbers are the strongest evidence that the interface, not the model, drove the surge. Reports at the time described ChatGPT reaching an estimated one million users within five days of launch, with estimates of over 100 million monthly active users by January 2023. Those figures come from analyst estimates and press reporting, not audited company disclosures, so treat them as directional rather than precise.

Capacity constraints shaped the early experience in measurable ways. During peak periods in early 2023, free-tier users encountered "at capacity" errors and queue screens. The Plus subscription was explicitly positioned as a way to skip those queues. Response latency varied by load, and long conversations slowed as the context window filled.

The context window itself is a concrete number worth knowing. Early ChatGPT models handled roughly 4,000 tokens of combined input and output, which is about 3,000 words. GPT-4 variants later offered 8,000 and 32,000 token windows, and later models extended far beyond that. A token is roughly three-quarters of a word in English. The practical consequence: early ChatGPT could hold a long article or a short document in working memory, but not a book.

Training data cutoffs are the other number that determines what the model knows. Each model version has a date beyond which it has no training data, so it cannot know about events after that point unless given web access or documents in the prompt. Users who ask an early model about recent news are asking it to extrapolate, and it will often do so confidently and incorrectly.

Edge cases: what the launch date does not tell you

The single most misunderstood edge case is the difference between the model and the product. GPT-3.5 was available through the API before ChatGPT existed. Developers were building on it. The same weights, given a different prompt and interface, produce very different user experiences. Asking "when did ChatGPT come out" is a product question; asking "when did GPT-3.5 come out" is a model question with a different answer.

A second edge case is regional and organizational rollout. Availability varied by country, by age verification requirements, and by enterprise agreements. Some users gained access through Microsoft products, which integrated the underlying models under different branding, before or alongside direct ChatGPT access. The launch date is a global marker that smooths over a staggered reality.

A third edge case is version drift. The model serving ChatGPT has been updated repeatedly without a rename. Users comparing notes about "ChatGPT's" behavior in different months may be comparing different underlying models with different training cutoffs and different safety tuning. This is why benchmarks and personal impressions of the system often disagree.

A professional caveat worth stating plainly: the release date is well documented, but performance claims about any specific early version are hard to verify retrospectively. The model served to a given user on a given day depended on tier, load, and internal routing. Researchers who need reproducible behavior should work against pinned API model versions, not the consumer chat interface.

Misconceptions and red flags

Several widely repeated claims about the launch are wrong or misleading.

Myth: ChatGPT was the first chatbot. ELIZA in the 1960s, and later commercial assistants, preceded it by decades. ChatGPT was the first widely accessible chatbot whose outputs were fluent enough to be mistaken for human writing at scale. Myth: the November 2022 release made GPT-3.5 public. The API predated it, and the weights were never released; only access was. Myth: the launch date is also the knowledge cutoff. Training cutoffs are separate dates, and they differ by model version. Myth: version numbers map cleanly to capability. Marketing names and internal model identifiers do not always align, and incremental updates can change behavior without a new name. Red flag: any source that gives a precise user count for the first week without noting it is an estimate. Company-confirmed figures were scarce. Red flag: claims that a specific date "proves" a model can or cannot do something. Capability depends on the version, the prompt, and the task, not the calendar.

The counterintuitive detail is that the launch succeeded despite the model being openly described as a research preview with known flaws. The company published warnings about incorrect answers and biased outputs at release. Users adopted it anyway, because fluent conversation is a compelling demo even when the content is unreliable. Adoption outran accuracy, and that gap defined the following two years of product development.

Practical actions for dating and verifying claims

If you need to pin down which ChatGPT version you are using or were using, these steps produce verifiable answers rather than impressions.

Check the model selector or settings menu in the product. It names the model family, and the help documentation maps names to release notes with dates. Ask the model for its training cutoff and treat the answer as a claim to verify, not a fact. Cross-check against official release notes. For research, use dated API model snapshots rather than the chat interface. Snapshots are immutable and documented. When citing behavior in writing, record the date, the model name, and the access tier. Behavior differs across all three. Treat user-count and adoption figures from 2022 and 2023 as estimates unless a company disclosure is cited. Report the range, not a single number.

The threshold question is simple. If a claim about ChatGPT depends on a specific date, ask whether it is a product date, a model date, or a training cutoff date. Conflating the three is the most common error in popular timelines, and it is the error that makes the November 2022 date seem to explain more than it does.

For the launch itself, the verified facts are narrow and sufficient: public research preview on November 30, 2022, built on GPT-3.5, free at launch, with a paid tier added on February 1, 2023. Everything else is context.

Limitations of what the timeline can establish

A release timeline documents availability, not capability, safety, or fitness for any particular use. Benchmarks published at launch measure narrow tasks under controlled conditions and do not predict performance on your document, your domain, or your language. The models have changed continuously since launch, so any characterization of "ChatGPT" as a fixed thing is inaccurate by default.

The historical record also has gaps. Early capacity incidents, regional outages, and silent model updates were not always publicly documented. Retrospective accounts rely on press coverage and user reports, which skew toward dramatic events. Absence of a documented change is not evidence that no change occurred.

This article is general educational information about product history and model mechanics. It is not a recommendation to use or avoid any specific service, and it makes no claims about outcomes from using these tools. Readers with contractual, compliance, or data-handling questions about AI services should consult a qualified professional familiar with their jurisdiction and obligations.