GPT Timeline: Every Model from GPT-1 to o3

GPT-1 arrived in June 2018 with 117 million parameters. Its contribution was a recipe rather than a product: train a Transformer on vast amounts of unlabeled text using unsupervised pre-training, then fine-tune that base model for specific language tasks. OpenAI showed that a single pre-trained model could be adapted across many natural language problems instead of building a fresh system for each one.

That two-stage approach (pre-train at scale, fine-tune for the task) set the template every later GPT model would follow. GPT-1 was modest on its own, but it established the paradigm that GPT-2 and GPT-3 would scale, and it placed OpenAI on the path that the rest of this timeline traces.

OpenAI announced GPT-2 in February 2019 at 1.5 billion parameters, more than ten times the size of GPT-1. The model generated convincing passages of text, and OpenAI initially declined to release the full version, describing it as too dangerous because of the risk of mass-produced fake content. The full model was eventually released in November 2019.

The decision split observers: some read it as responsible caution about the dual-use nature of language models, others as a publicity move. Either way, GPT-2 produced the first major AI safety debate of the large language model era, and the phrase too dangerous to release became a reference point in arguments about how labs should publish powerful models.

GPT-3 followed in June 2020 with 175 billion parameters, roughly 100 times larger than GPT-2. Without any task-specific fine-tuning, it could write essays, code, and poetry and learn a task from a few examples placed in the prompt, a behavior known as few-shot learning. The API launched in beta and an ecosystem of startups grew on top of it, and the scaling hypothesis (that bigger models would simply get more capable) gained credibility.

The public turn came in November 2022. ChatGPT wrapped a GPT-3.5 model fine-tuned with reinforcement learning from human feedback in a plain chat interface. It reached one million users in five days and 100 million in two months, the fastest-growing consumer application to that point. The model underneath was familiar, but the conversational framing made the capability legible to people who had never touched an API.

GPT-4, released in March 2023, was multimodal: it accepted both text and images. It passed the bar exam around the 90th percentile, scored 1410 on the SAT, and showed clearly stronger reasoning and reliability than GPT-3.5. Its results on professional exams pushed many industries to take model capability seriously, and Microsoft committed more than 10 billion dollars to OpenAI.

GPT-4o, the omni model, arrived in May 2024. It processed text, audio, images, and video in a single model with near-instant responses, holding spoken conversations with emotional inflection and reacting to what a camera saw in real time. Where GPT-4 raised the reasoning ceiling, GPT-4o made multimodal interaction feel immediate, moving the interface from typed text toward voice and vision.

In September 2024 OpenAI released o1, a model trained to think before it answers by working through chain-of-thought reasoning at inference time. Instead of producing an answer immediately, o1 spends more compute on multi-step solutions to math, coding, and science problems, trading speed for accuracy and reaching PhD-level results on some scientific reasoning benchmarks. This introduced test-time compute scaling: a model can get smarter by thinking longer, not only by being larger.

o3 followed in 2025 as the successor to o1, with stronger reasoning and state-of-the-art results on many math and coding benchmarks. Paired with GPT-4o for fast conversational work, it gave OpenAI a two-model strategy: a quick responder and a deeper thinker. The reasoning turn marks where the GPT timeline stops being only about larger base models and starts pointing toward agentic systems that plan and work across steps.