ChatGPT came out on November 30, 2022, when OpenAI released a research preview built on the GPT-3.5 family of models. The chatbot reached users as a free web interface at chat.openai.com. Within five days it passed one million users, a milestone that took Facebook about ten months and Instagram roughly two and a half months. That gap between launch date and cultural impact is the real story of when ChatGPT came out.

The launch date is simple. The mechanism behind it, the model lineage that followed, and the reasons a chatbot trained on next-token prediction felt different from earlier assistants are where most explainers stop short. This one goes deeper.

The mechanism: why a text predictor behaved like an assistant

ChatGPT is a large language model wrapped in a conversational interface. The base model learns by predicting the next token in sequences drawn from a large text corpus. A token is a fragment of a word, often two to four characters in English. Next-token prediction forces the model to encode grammar, factual associations, and patterns of reasoning as a side effect of compression.

Raw prediction alone does not produce a helpful assistant. The critical layer is instruction tuning. Humans wrote demonstration conversations, and the model was fine-tuned on them. A second stage, reinforcement learning from human feedback, trained a reward model on human rankings of competing responses, then optimized the chatbot against that reward signal.

The interface change mattered as much as the model. Earlier GPT-3 access required API calls and prompt engineering. ChatGPT packaged the model behind a text box with conversation memory. Memory let users iterate on a draft across turns, which changed the perceived capability more than any single benchmark number.

One nuance: the November 2022 model was labeled GPT-3.5, a sibling of the GPT-3 family rather than GPT-4. OpenAI described the release as a research preview and warned that the model could produce incorrect or misleading information. That warning has remained accurate across every subsequent version.

Quantified context: dates, versions, and adoption

The timeline from research preview to platform spans roughly two years and four major model generations. Each row marks a public release date, not a training cutoff.

Date Release What changed
Nov 30, 2022 ChatGPT research preview (GPT-3.5) Free web chat, conversation memory
Mar 14, 2023 GPT-4 Multimodal image input, stronger reasoning
May 13, 2024 GPT-4o Faster, cheaper, voice and vision native
Jul 18, 2024 GPT-4o mini Low-cost tier replacing GPT-3.5
Sep 12, 2024 o1-preview Reasoning models that spend tokens thinking
Jan 31, 2025 o3-mini Cheaper reasoning tier
Apr 14, 2025 GPT-4.1 in API Long-context and coding improvements

Adoption numbers help calibrate the launch's significance. ChatGPT reached an estimated 100 million monthly active users within about two months, according to UBS analyst estimates reported in early 2023. That pace was faster than any prior consumer application at the time. The comparison is imperfect because ChatGPT launched with free access and no app-store gatekeeping.

A professional caveat: user counts reported by third parties are estimates, not audited figures. Treat marketing and analyst numbers as directional. The launch date itself, November 30, 2022, is documented by OpenAI's own announcement and is not in dispute.

Edge cases: what the launch date does not tell you

The launch date masks several distinctions that matter for anyone citing it.

The underlying GPT-3 model was available through OpenAI's API from June 2020, roughly 30 months before ChatGPT. The November 2022 release was not the first chatbot. It was the first large language model chatbot offered free to the general public at scale. Model training cutoffs differ from release dates. A model released in 2024 may have training data ending months earlier. Regional availability varied. Some countries gained access weeks or months after the US launch. The research preview label meant features could change without notice, and they did.

A counterintuitive detail: the most consequential design decision was arguably the free price. Charging nothing converted a research artifact into a mass-market product and generated the human feedback data that improved later models. The launch was as much a data-collection strategy as a product release.

Another edge case involves the name. ChatGPT and GPT-3.5 are not interchangeable. ChatGPT is the product, the interface, and the tuning pipeline. GPT-3.5 is the model family. Subsequent ChatGPT versions run on different models, which is why "when did ChatGPT come out" has one answer while "which model am I talking to" has many.

Misconceptions: four claims worth correcting

Misconception one: ChatGPT launched with GPT-4. It did not. GPT-4 arrived on March 14, 2023, about three and a half months later.

Misconception two: the model searches the web by default. The original ChatGPT had no browsing capability. Retrieval and browsing were added later in 2023 and have changed repeatedly since.

Misconception three: the model knows recent events. Training cutoffs and release dates are separate. A model can launch in 2025 with a knowledge cutoff months earlier.

Misconception four: a higher version number guarantees better answers on your task. Reasoning models trade speed for deliberation and can underperform faster models on simple lookups. Provider benchmark scores rarely transfer cleanly to a specific workflow.

For medical, legal, or financial questions, none of these misconceptions is the main risk. The main risk is treating fluent output as verified fact. Treat any figure the model produces as unverified until you check the primary source.

Practical actions: using the release timeline correctly

Concrete steps for anyone citing ChatGPT's launch or choosing a version.

Cite November 30, 2022 as the public launch date and distinguish it from the GPT-3 API date of June 2020. Check the model name in the interface before relying on a capability. Product names and underlying models diverge. Set a personal verification threshold. For any number that affects a decision, confirm it against a primary source before acting. Separate training cutoff from release date when judging whether the model can know an event. Use the cheapest adequate model for routine tasks. The GPT-4o mini tier exists precisely for high-volume, low-stakes work.

A range to hold in mind: the gap between a model's training cutoff and its release date is typically weeks to months. If you need information newer than the cutoff, use a retrieval-enabled tool or check the source yourself.

Limitations: what this timeline cannot establish

Release dates are documented facts. Their causal weight is not. The launch on November 30, 2022 did not happen in a vacuum. Years of scaling research, instruction-tuning methods, and prior chatbot failures set the stage. Attributing the uptake to one date oversimplifies a compounding process.

Provider-reported capabilities also age quickly. A benchmark result published at launch describes one model version on one test set. It does not promise performance on your documents, your language, or your edge cases. Third-party user estimates carry the same caveat.

The article is general educational information about product history and model mechanics. It is not a recommendation to use any specific tool for a specific task. For decisions with financial, legal, or safety consequences, consult a qualified professional in your jurisdiction.