Why Does OpenAI Have So Many Models? The Full Story (2018–2026)
From the first GPT to GPT-6.1: every major OpenAI model family, what mini, nano, Pro, o-series and Codex mean, and how to choose by the job.
On this page
GPT-4o, o3-mini, GPT-4.1 nano, GPT-5-Codex, GPT-6 Astra. Open OpenAI’s model list and it looks like chaos. But the names follow a story: a few branches, each doing a different job, and a handful of suffixes that describe size and effort.
This post follows that story from 2018 to the end of October 2026, then gives you a simple way to read any model name. The timeline stops at October 27, 2026, and every date comes from OpenAI’s own announcements and model pages. A date marks the stage shown (a research preview, API release and so on), not necessarily when everyone got access.
Quick answer: the branches
| Branch | Its job | Names you’ll see |
|---|---|---|
| GPT | Language and general capability | GPT-3, GPT-4, GPT-4o, GPT-5, GPT-6 |
| o-series | Spending more effort on hard problems | o1, o3, o4-mini |
| Images | Creating and editing pictures | DALL·E, GPT Image |
| Audio | Listening and speaking | Whisper, TTS, Realtime, Transcribe |
| Video | Generating moving scenes | Sora |
| Codex | Software engineering | Codex, GPT-5-Codex |
| Embeddings | Comparing meaning for search | text-embedding-3 |
| Specialists | Moderation, safety, security, science | Omni moderation, gpt-oss-safeguard, Cyber, Rosalind |
ChatGPT is the product. The GPT model is the engine. Keep that difference in mind and the names get easier.
2018–2021: one idea, many directions
It started with a simple recipe: learn from a large amount of text, then adapt that foundation to specific tasks, like learning to read before studying subjects.
| Date | Model | What it showed |
|---|---|---|
| Jun 2018 | GPT | The pre-train-then-adapt recipe (a research model) |
| Feb 2019 | GPT-2 | Surprisingly coherent paragraphs; released in stages, full model in November. Fluent didn’t mean reliable. |
| Apr 2019 | MuseNet | The same “predict what comes next” idea, applied to musical notes |
| Apr 2020 | Jukebox | Raw audio, including rough singing |
| May 2020 | GPT-3 | 175 billion parameters, and few-shot learning: show a few examples in the prompt and it attempts a new task |
| Jun 2020 | Image GPT | Sequence learning applied to pixels |
| Jan 2021 | CLIP | Matched images to words |
| Jan 2021 | DALL·E | Went the other way: words to images |
| Aug 2021 | Codex (original) | Plain language to code; behind the early GitHub Copilot |
The lesson of this era: “predict the next piece” worked for words, notes, sound and pixels. That’s why OpenAI’s tree has so many branches.
2022: learning to follow instructions
| Date | Model | What changed |
|---|---|---|
| Jan 2022 | Text and code embeddings | Text turned into numbers you can compare for meaning. A search for “car repairs” can find “automobile maintenance”. These retrieve; they don’t chat. |
| Jan 2022 | InstructGPT | Human demonstrations and feedback shifted the model from continuing text to following a request |
| Apr 2022 | DALL·E 2 | Editing, variations and extending an image beyond its edges |
| Sep 2022 | Whisper | Open-source speech-to-text, multilingual, with translation into English |
| Nov 30, 2022 | ChatGPT (GPT-3.5) | A public chat interface anyone could try |
| Dec 2022 | text-embedding-ada-002 | A simpler embedding offering, working behind the scenes |
2023: GPT-4 and building blocks
| Date | Model | What changed |
|---|---|---|
| Mar 2023 | GPT-3.5 Turbo | Chat models in the API; the older ada, babbage, curie and davinci tiers gave way |
| Mar 2023 | GPT-4 | Better at complex instructions and reasoning; image input was phased in later |
| Jun 2023 | Function calling | A model could request a structured action, which the app then runs |
| Sep 2023 | DALL·E 3 | Followed detailed descriptions more closely, with ChatGPT integration |
| Nov 2023 | GPT-4 Turbo | A much larger context window (working space, not permanent memory) |
| Nov 2023 | TTS-1 and TTS-1 HD | Text to speech: the opposite job from Whisper |
2024: omni models and reasoning
| Date | Model | What changed |
|---|---|---|
| Jan 2024 | Embedding 3 small and large | A refreshed retrieval family |
| Feb 2024 | Sora (research preview) | Text to moving scenes |
| May 2024 | GPT-4o | “o” for omni: text, vision and audio in one family |
| Jul 2024 | GPT-4o mini | A smaller model for jobs that don’t need the biggest engine |
| Sep 2024 | o1-preview and o1-mini | Trained to spend more computation working through hard problems before answering |
| Sep 2024 | Omni moderation | Classifies text and images for safety categories |
| Oct 2024 | Realtime API (beta) | Low-latency voice conversation |
| Dec 2024 | Sora Turbo | Video generation for eligible users |
| Dec 2024 | o1 | The full reasoning model in the API |
The “thinking” in reasoning models is extra computation, not evidence of consciousness.
2025: the family tree explodes
| Date | Model | What changed |
|---|---|---|
| Jan | o3-mini | Small reasoning model with adjustable effort |
| Feb | GPT-4.5 | Broad knowledge and natural interaction (a different goal from the o-series) |
| Mar | Search and computer-use previews | Answers tied to retrieved info, and a model that reads screens and proposes actions |
| Mar | o1-pro | More computation for difficult questions |
| Mar | GPT-4o transcribe and mini-TTS | Specialist listening and speaking models |
| Apr | GPT-4.1, mini, nano | Instruction following, coding and a million-token context |
| Apr | o3 and o4-mini | Reasoning plus visual understanding and tool use |
| Apr | GPT Image 1 | Detailed instructions, text in images, editing |
| May | codex-1 | A new Codex that works on whole repository tasks and runs checks |
| Jun | o3-pro, deep research models | High-compute work, and tool-using research reports |
| Aug 5 | gpt-oss 120b and 20b | Open-weight models you can run yourself (Apache licence) |
| Aug 7 | GPT-5, mini, nano | Stronger reasoning, coding and tools in one family |
| Aug | GPT-Realtime, GPT-Audio | Generally available voice models |
| Sep | GPT-5-Codex, Sora 2 | Longer coding tasks; video with synchronised audio |
| Oct | GPT-5 Pro and mini branches | Higher-compute and lighter tiers |
| Nov | GPT-5.1, Codex-Max | A faster path when deep reasoning isn’t needed; long-running coding |
| Dec | GPT-5.2, GPT Image 1.5, GPT-5.2-Codex | Professional tasks, image updates, big refactors |
A note on Sora: Sora 2 is a historical milestone here. The Sora product ended in April 2026 and its API in September 2026.
2026: tiers get names
| Date | Model | What changed |
|---|---|---|
| Feb | GPT-5.3-Codex, Codex-Spark | Steerable coding, and a fast real-time variant |
| Feb | Realtime 1.5, Audio 1.5 | Refreshed voice models |
| Mar | GPT-5.3 Instant snapshot | The ChatGPT chat model in the API (not the same as GPT-5.3-Codex) |
| Mar | GPT-5.4, Pro, mini, nano | Built-in computer use and longer-context work |
| Apr | GPT-5.4-Cyber, GPT-Rosalind | Trusted-access models for security defenders and life sciences |
| Apr | GPT Image 2, GPT-5.5 | Flexible image sizes; carrying messy tasks through to completion |
| May–Jul | Realtime 2 and 2.1, GPT Transcribe | Translation, interruptions, noise, and spelled names and numbers |
| Jul | GPT-5.6 Sol, Terra, Luna | Named tiers: flagship, balanced and efficient |
| Sep | GPT-6 Astra | End-to-end work across reasoning, coding, research and documents |
| Sep | Image 2.5 Sunburst and Flare | Precision editing vs fast everyday generation |
| Sep | GPT-Live 1 | Live conversation while another model handles the deep work |
| Sep | GPT-6 Sol and Luna, GPT-6.1 Sol | Tiers below Astra |
| Oct 26 | GPT-6.1 Terra | Visual reasoning, precise computer use and clear writing |
How to read any OpenAI model name
| Part of the name | What it usually means |
|---|---|
| Version number (4, 5, 6.1) | The generation. A decimal is an update within it, not a new kind of AI. |
| o in GPT-4o | Omni: several input types in one model |
| o-series (o1, o3) | Reasoning models that spend more effort before answering |
| mini / nano | Smaller and faster for simpler, high-volume jobs, not a universal substitute |
| Pro | More computation for hard problems, same family |
| Turbo | An older label for faster, cheaper versions |
| Codex | Tuned for software engineering |
| Sol / Terra / Luna / Astra | Named tiers, from most capable to most efficient |
| Realtime / Audio / Transcribe / TTS | Live voice, request-based audio, speech-to-text, text-to-speech |
| -latest alias | A moving pointer. Pin a dated version when results must be repeatable. |
| preview | An early stage, not general availability |
Choose by the job
The useful question isn’t “which number is newest?” It’s “what do I need done?”
| You need… | Look at |
|---|---|
| A quick answer or simple, high-volume task | A mini, nano or efficient tier |
| A careful analysis of a hard problem | A reasoning model, or a Pro option |
| An image created or edited | The GPT Image family |
| A transcript or subtitles | Transcribe or Whisper models |
| A spoken voice | TTS or audio models |
| A live voice conversation | Realtime or Live models |
| Code changed across a project | Codex models |
| Search over your own documents | Embedding models |
The map in one paragraph: GPT expanded language and general capability. DALL·E and GPT Image developed visual creation. Whisper and the audio families developed listening and speaking. The o-series focused on reasoning. Codex specialised in software. And the latest GPT families pull more of those abilities together.
Try it: open OpenAI’s model list, pick any three names, and decode them with the table above.
