The history of GPT, from GPT-1 to GPT-6.1

Eight years separate a small research model from today's GPT-6 family. Here is the story in order: each major release, the date it arrived and what it changed for people who use AI.

6 min readUpdated

Eight years of GPT on one page

OpenAI published its first GPT model in June 2018 as a research result that few people outside the field noticed. Eight years later, GPT models run ChatGPT and countless other apps. This page follows that path release by release, from GPT-1 to GPT-6.1 Terra in October 2026, and explains what each step changed for ordinary users.

Dates below are OpenAI's announcement dates. A table with the main releases is near the end of the page.

The research years: GPT-1 to GPT-3 (2018 to 2020)

GPT-1 (June 2018) showed a simple idea: train a transformer network on a large amount of ordinary text first, then adjust it for specific tasks. The model had 117 million parameters and was a paper, not a product.

GPT-2 (February 2019) was more than ten times larger, at 1.5 billion parameters, and wrote paragraphs that sounded surprisingly human. OpenAI first released only smaller versions, citing concerns about misuse, and published the full model in November 2019. It was the last GPT model with openly downloadable weights until 2025.

GPT-3 (May 2020) jumped to 175 billion parameters. Its research paper showed that a big enough model could do new tasks from a few examples in the prompt, without extra training. OpenAI offered it through a paid API instead of releasing it, and the first wave of AI writing tools was built on top of it.

ChatGPT and GPT-4 (2022 to 2024)

On November 30, 2022 OpenAI opened ChatGPT to the public as a free research preview. It ran on a GPT-3.5 model trained to follow instructions and hold a conversation, and it turned language models from a developer topic into something everyone talked about.

GPT-4 (March 14, 2023) was a big step in reasoning and reliability, and the first GPT that could look at images. OpenAI's GPT-4 announcement reported strong results on professional and academic exams. GPT-4 Turbo followed in November 2023 with a longer context and lower prices.

GPT-4o (May 13, 2024), with "o" for omni, handled text, images and audio in one model and became the default in ChatGPT, including for free users. GPT-4o mini arrived in July 2024 as a cheap small model for apps.

Thinking models and GPT-5 (2024 to 2025)

In September 2024 OpenAI introduced o1-preview, the first of its reasoning models. These models work through a problem step by step before answering, which made them much better at maths, science and code. The o-series continued with o3-mini in January 2025 and o3 and o4-mini in April 2025.

Alongside them, OpenAI kept releasing classic GPT models: GPT-4.5 in February 2025, a very large model with a warmer writing style, and GPT-4.1 in April 2025, aimed at developers with a context window of a million tokens.

In August 2025 two events landed in the same week. On August 5 OpenAI released gpt-oss-120b and gpt-oss-20b, open-weight models under the Apache 2.0 licence. On August 7 it launched GPT-5, which brought the GPT and o-series lines together: one system that decided by itself when to answer quickly and when to think longer.

Faster updates: GPT-5.1 to GPT-5.5

After GPT-5, OpenAI moved to frequent point releases instead of rare big jumps. GPT-5.1 arrived in November 2025 and GPT-5.2 on December 11, 2025. In 2026 came GPT-5.3-Codex for programming in February, GPT-5.4 in March and GPT-5.5 on April 23.

During this period the size of a model was still shown with suffixes. GPT-5 itself came to the API in three sizes, called GPT-5, GPT-5 mini and GPT-5 nano, and GPT-5.4 mini and nano followed in March 2026. Developers picked the small versions for quick, cheap jobs such as sorting messages or pulling data out of forms, and the full model for harder work. The pace was fast enough that a model could be the newest one for only a few weeks.

Older models were retired from ChatGPT along the way. GPT-4o left ChatGPT in February 2026, and o3 and GPT-4.5 followed on May 28, 2026.

New names: GPT-5.6 Sol, Terra and Luna

On July 9, 2026 OpenAI released GPT-5.6 with a new naming scheme. The number marks the generation; a second word marks the tier. Sol is the most capable tier, Terra the balanced middle and Luna the fastest and cheapest. The tiers roughly replaced the earlier habit of adding "mini" and "nano" to model names, and each tier can be updated on its own schedule.

The new scheme explains names that look odd at first. GPT-6.1 Terra and GPT-6 Luna are both current models: Terra received its 6.1 update, while Luna has not yet needed one. When you read a model name today, the word tells you roughly how capable and how expensive it is, and the number tells you how recent it is.

GPT-6 and GPT-6.1 (autumn 2026)

The current generation came out in several steps within eight weeks:

  • September 3, 2026: GPT-6 Astra, a new flagship tier above Sol, for the most demanding professional work.
  • September 22: GPT-6 Sol and GPT-6 Luna, trained with methods similar to Astra's, at lower API prices than their GPT-5.6 versions.
  • September 29: GPT-6.1 Sol, which OpenAI describes as close to Astra on complex work at a much lower price.
  • October 7: GPT-6 in ChatGPT. Paid plans got GPT-6 Sol in the main chat, and from October 8 free users got GPT-6 Luna.
  • October 26: GPT-6.1 Terra, the new mid-tier model, strong at writing and reading images.

As of November 2026 there is no GPT-6.1 Astra and no newer open-weight GPT than gpt-oss. On GPT.me you can try GPT-6 Luna and GPT-6.1 Terra free in the chat.

Main releases at a glance

DateModelWhat it brought
June 2018GPT-1Pre-training on plain text, then task tuning
February 2019GPT-2Fluent paragraphs; staged release
May 2020GPT-3Learning from examples in the prompt; paid API
November 2022ChatGPT (GPT-3.5)Conversation for everyone
March 2023GPT-4Better reasoning; image input
May 2024GPT-4oText, images and audio in one model
September 2024o1-previewThinking before answering
August 2025gpt-oss, GPT-5Open weights again; one unified system
July 2026GPT-5.6Sol, Terra and Luna tiers
September 2026GPT-6 Astra, Sol, LunaNew generation and flagship tier
October 2026GPT-6.1 TerraNew mid-tier model

What changed along the way

Looking back, three trends stand out. Models went from text only to reading images and sound. They learned to think before answering, first in separate o-series models and then inside every GPT. And the price of a given level of quality kept falling, which is why sites like ours can offer a recent GPT model for free. One thing has not changed: every GPT model can still state wrong things with confidence, so checking important answers is as useful today as it was in 2022.

FAQ

What was the first GPT model?

GPT-1, published by OpenAI in June 2018. It was a research model, not a public product.

When did ChatGPT launch?

On November 30, 2022, as a free research preview running on a GPT-3.5 model.

Is any GPT model open source?

OpenAI released the weights of GPT-2 in 2019 and of gpt-oss-120b and gpt-oss-20b in August 2025, the latter under the Apache 2.0 licence. The GPT-6 models are available only through OpenAI's services and partners.

What is the newest GPT model?

As of November 2026 the most recent release is GPT-6.1 Terra from October 26, 2026. The most capable model is GPT-6 Astra.

Got a question? Ask it now

No account and no download. Open the chat, type what is on your mind, and the answer starts appearing in a second or two.

Start chatting