GPT-2 to GPT-6: What Actually Changed, in Plain English

Quick answer: In seven years, OpenAI's models went from continuing text (GPT-2), to learning from examples in the prompt (GPT-3), to seeing images and following a chat format (GPT-4), to reasoning before answering (GPT-5), to working alongside you while tools run and taking corrections mid-task (GPT-6 Astra). The size of the brain matters less than how you work with it.

This post was prompted by a popular ByteByteGo graphic on LinkedIn, "GPT-2 to GPT-6: What Actually Changed". It's a great picture. We checked each step against OpenAI's own pages, and fixed a few details along the way.

2019: GPT-2, a text continuer

You gave GPT-2 the start of some text and it carried on. That was all it did. OpenAI's GPT-2 paper showed that a model trained only to predict the next word could sometimes do tasks it was never trained for ("zero-shot"). The full 1.5 billion parameter model came out in November 2019, months after the first announcement.

How you used it: write the start, hope for a useful ending.

2020: GPT-3, learning from examples

GPT-3 had 175 billion parameters. Its paper was literally called Language Models are Few-Shot Learners. You could show it a couple of examples ("dog → chien, cat → chat, bird → ?") and it picked up the pattern without retraining. This is called in-context learning. OpenAI opened it to developers through its first API.

How you used it: prompt engineering. Clever examples did most of the work.

2023: GPT-4, images and chat

GPT-4 (March 2023) could take an image as well as text, and came with 8K and 32K token context versions.

Two things often credited to GPT-4 actually came a little earlier. The system message, a hidden instruction that sets the rules, shipped with the ChatGPT API two weeks before. Fine-tuning with human feedback (RLHF) was already used for InstructGPT and ChatGPT. GPT-4 is where they all came together.

How you used it: a chat with a system prompt, plus pictures.

2025: GPT-5, thinking by default

GPT-5 (August 2025) made reasoning the normal way to answer. The model spends hidden "reasoning tokens" thinking first, and developers choose a reasoning effort from minimal to high. In ChatGPT, a router picks a fast model or a thinking model for you.

Again, the ideas weren't brand new. Reasoning effort began with o1 and o3 in 2024, and tool calling dates back to 2023. GPT-5 made them the default.

The context window is 400,000 tokens, with up to 128,000 of those as output.

How you used it: pick an effort level, give it tools, let it think.

2026: GPT-6 Astra, working with you while it runs

The model is officially called GPT-6 Astra. Three changes matter:

• Bigger context: a 1,050,000 token window, with up to 128,000 output tokens. Prompts over 272,000 input tokens cost more per token, so the full window isn't free. Effort now goes up to xhigh and max.
• Async tools: with async tool calling, the model can keep working after it starts a slow tool, and your app sends the result back later.
• Mid-turn steering: you can send new instructions while it is still working, and it adjusts without starting over. It can't take back text it has already sent, or cancel tools that already started.

The graphic also mentions "searchable notes". We couldn't find that in OpenAI's API docs. Reports describe it as an experimental Codex feature, so we've left it out.

How you use it: less like sending a letter, more like talking to a coworker you can interrupt.

What this means for you

1. Learn the new controls, not just the new name. Reasoning effort is now your main dial for speed, cost and quality: low for quick lookups, high or above for hard problems.
2. A big context window is a limit, not a goal. You can paste in a whole codebase, but longer prompts are slower and cost more. Our context engineering explainer covers why less is often better.
3. New features need new code. Steering and async tools need the Responses API (steering over WebSocket). Older apps don't get them for free.

Want the same story for the other big labs? Read Claude 1 to Claude 5 and Google's road from BERT to Gemini 4.