AI Wiki

GPT Model Timeline (GPT-1 to GPT-5.x)

AI ModelsGenerative AILarge Language Models
8 min read
Updated
Suggest editHistory
RawGraph

Last reviewed

Sources

19 citations

Review status

Source-backed

Revision

v1 · 1,678 words

OpenAI's GPT lineage runs from GPT-1 (June 2018, 117 million parameters, a 512-token context window) to the GPT-5.6 "Sol, Terra, Luna" family that entered limited preview on June 26, 2026. As of July 2026 the newest generally available model is GPT-5.5 (released April 23, 2026), while GPT-5.6 remains restricted to preview partners. The arc has three phases: a pure scaling race (GPT-1 to GPT-3, 117M to 175B parameters), a product and multimodal era (ChatGPT, GPT-4, GPT-4o), and, since o1 in September 2024, a "reasoning" era that GPT-5 unified into one model that automatically routes between fast answers and deliberate thinking.

Short version, if you just want the answer: OpenAI has publicly disclosed a parameter count for exactly three models (GPT-1, GPT-2, GPT-3). Every model from GPT-4 onward is marked "undisclosed" below, and any specific number you see for them (for example the widely repeated 1.8-trillion figure for GPT-4) is an unofficial third-party estimate, never confirmed by OpenAI. For the current default assistant use GPT-5.5; the o-series (o1, o3, o4-mini) was the reasoning-specialist line that GPT-5 folded in. This page tracks OpenAI's GPT and o-series specifically; for a cross-lab index of every major flagship, see the AI Model Release Timeline.

The GPT and o-series timeline at a glance

All models are developed by OpenAI and are proprietary (accessible via ChatGPT and the OpenAI API). "Undisclosed" means OpenAI has never published an official parameter count. Context is the maximum context window; where the API and ChatGPT differ, the larger API figure is shown. Last verified: July 2026.

DateModelParamsContext windowHeadline change
2018-06GPT-1117M512First generative pre-trained transformer: unsupervised pre-training plus supervised fine-tuning [1]
2019-02GPT-21.5B1KAbout 10x the scale; coherent long-form text; staged release over misuse fears [2]
2020-05GPT-3175B2KAbout 100x the scale; few-shot in-context learning; launched the OpenAI API [3]
2022-11GPT-3.5 / ChatGPTUndisclosed4KRLHF-tuned chat model; ChatGPT reached about 100M users in two months [4]
2023-03GPT-4Undisclosed8K / 32KMultimodal (image input); roughly 90th-percentile bar-exam performance [5]
2023-11GPT-4 TurboUndisclosed128KCheaper, 128K context, fresher knowledge cutoff, JSON mode (DevDay) [6]
2024-05GPT-4oUndisclosed128KNatively multimodal "omni" (text, vision, audio); real-time voice; free-tier flagship [7]
2024-09 to 12o1Undisclosed128K to 200KFirst reasoning model; scales test-time compute via long chains of thought [8]
2025-02GPT-4.5Undisclosed128KLargest non-reasoning GPT (codenamed Orion); shipped as a research preview [9]
2025-01 to 04o3 / o4-miniUndisclosed200KTool-using reasoning: models call tools inside the reasoning chain [10]
2025-08GPT-5Undisclosed400KFirst "unified" model: an auto-router picks fast vs thinking per query [11]
2025-11GPT-5.1Undisclosed400KWarmer, steerable personalities (8 presets); adaptive reasoning [12]
2025-12GPT-5.2Undisclosed400KSplit into Instant, Thinking, and Pro; accuracy and coding gains; shipped early [13]
2026-02GPT-5.3Undisclosed400KUnified the Codex and general branches; first OpenAI model on non-NVIDIA silicon [14]
2026-03GPT-5.4Undisclosed1M1M-token context; built-in computer use; lower error rate [15]
2026-04GPT-5.5Undisclosed1MAgentic-coding and scientific-reasoning gains; became the ChatGPT default [16]
2026-06GPT-5.6 (Sol, Terra, Luna)Undisclosedn/r (preview)Next-generation family in limited, access-restricted preview [17]

What is the GPT lineage, in order?

The full order is GPT-1 (2018), GPT-2 (2019), GPT-3 (2020), GPT-3.5 and ChatGPT (2022), GPT-4 (2023), GPT-4 Turbo (2023), GPT-4o (2024), o1 (2024), GPT-4.5 (early 2025), o3 and o4-mini (2025), GPT-5 (2025), then the rapid point releases GPT-5.1, GPT-5.2, GPT-5.3, GPT-5.4, GPT-5.5, and GPT-5.6 across late 2025 and the first half of 2026. GPT stands for generative pre-trained transformer, the architecture introduced with GPT-1 and built on the Transformer. The cadence compressed sharply over time: seven years separated GPT-1 from GPT-5, but OpenAI then shipped GPT-5.1 through GPT-5.6 in under a year.

When did GPT stop disclosing parameter counts?

GPT-3, at 175 billion parameters, was the last OpenAI model with an officially published size [3]. Starting with GPT-4, OpenAI has declined to disclose parameter counts, training-data size, or architecture details, citing competitive and safety reasons [5]. The number most often attached to GPT-4, roughly 1.8 trillion parameters in a mixture-of-experts configuration, comes from third-party industry reporting and leaks; OpenAI has never confirmed it, so it should be treated as an unofficial estimate rather than a specification. For every model from GPT-4 onward, the honest answer is "undisclosed," which is why the table above avoids putting a specific figure in those cells. GPT-3.5 is a borderline case: it derives from the 175B GPT-3 series, but the exact size of the deployed gpt-3.5-turbo model was never confirmed, so it is listed as undisclosed too.

How did context windows and model scale grow?

Context windows expanded far faster than disclosed parameter counts. GPT-1 handled 512 tokens and GPT-2 handled 1,024; GPT-3 reached 2,048 and the first ChatGPT ran at 4,096 [3][4]. GPT-4 launched with 8K and 32K variants, then GPT-4 Turbo jumped to 128K in late 2023, a level GPT-4o and GPT-4.5 kept [6][7][9]. The reasoning models pushed to 200K (o1's general-availability release and the o3 line) [8][10], GPT-5 opened at 400,000 tokens in the API [11], and GPT-5.4 and GPT-5.5 reached a full 1 million tokens in the API in 2026 [15][16]. In practice ChatGPT surfaces a smaller window than the API for the same underlying model, and advertised maximums include both input and output budgets (GPT-5-class models expose roughly 272K input tokens within a 400K total) [11].

What are the o-series reasoning models?

The o-series was OpenAI's separate line of "reasoning" models that spend extra compute at inference time to think through a problem before answering. It began with o1, previewed on September 12, 2024 and released in full on December 5, 2024 with a 200K context window [8]. o3-mini followed on January 31, 2025, and the full o3 shipped alongside o4-mini on April 16, 2025, adding agentic tool use inside the chain of thought [10]. The reasoning turn added a second axis of progress beyond raw pre-training scale, and every major lab raced on it. With GPT-5, OpenAI merged this line back into the main GPT branch: rather than choosing a separate model, users get one system that decides when to think longer. For the dedicated hub, see OpenAI o-series.

What changed across GPT-5 and the 5.x point releases?

GPT-5, released August 7, 2025, was OpenAI's first "unified" model, combining the fast responses of the GPT line with the deliberate reasoning of the o-series behind an automatic router [11]. The point releases then arrived roughly every one to two months. GPT-5.1 (November 12, 2025) focused on a warmer, more steerable personality with selectable presets and adaptive reasoning [12]. GPT-5.2 (December 11, 2025) split into Instant, Thinking, and Pro tiers and was reportedly accelerated by an internal "Code Red" after Google's Gemini 3 briefly took the lead [13]. GPT-5.3 (Codex on February 5, 2026; Instant on March 3, 2026) unified the coding and general branches and, via its Codex-Spark variant, became the first OpenAI model served on non-NVIDIA silicon [14]. GPT-5.4 (March 5, 2026) added a 1-million-token API context and built-in computer use while cutting error rates [15]. GPT-5.5 (April 23, 2026) improved agentic coding and scientific reasoning and became the ChatGPT default [16]. GPT-5.6, a three-model family named Sol (flagship), Terra (balanced), and Luna (fast), entered a limited, access-restricted preview on June 26, 2026 [17].

Which GPT model should you use now?

As of July 2026, GPT-5.5 is the strongest generally available OpenAI model and the ChatGPT default, so it is the safe pick for most work [16]. GPT-5.6 Sol is more capable on paper but is limited to preview partners, so it is not a practical choice for general users yet [17]. GPT-4o remains the ubiquitous, low-latency multimodal workhorse where full frontier reasoning is not required [7], and the standalone o-series models have largely been superseded by GPT-5's built-in thinking mode. For a broader comparison across labs (Anthropic, Google, xAI, and open-weight models), use the cross-lab AI Model Release Timeline.

References

  1. OpenAI, "Improving Language Understanding by Generative Pre-Training" (GPT-1), 2018. https://cdn.openai.com/research-covers/language-unsupervised/language_understanding_paper.pdf ↩
  2. OpenAI, "Better Language Models and Their Implications" (GPT-2), February 14, 2019. https://openai.com/research/better-language-models ↩
  3. T. Brown et al., "Language Models are Few-Shot Learners" (GPT-3, 175B), May 28, 2020. https://arxiv.org/abs/2005.14165 ↩
  4. OpenAI, "Introducing ChatGPT," November 30, 2022. https://openai.com/blog/chatgpt ↩
  5. OpenAI, "GPT-4" and GPT-4 Technical Report, March 14, 2023. https://openai.com/index/gpt-4-research/ ↩
  6. OpenAI, "New models and developer products announced at DevDay" (GPT-4 Turbo), November 6, 2023. https://openai.com/blog/new-models-and-developer-products-announced-at-devday ↩
  7. OpenAI, "Hello GPT-4o," May 13, 2024. https://openai.com/index/hello-gpt-4o/ ↩
  8. OpenAI, "Introducing OpenAI o1-preview," September 12, 2024; o1 general availability, December 5, 2024. https://openai.com/index/introducing-openai-o1-preview/ ↩
  9. OpenAI, "Introducing GPT-4.5," February 27, 2025. https://openai.com/index/introducing-gpt-4-5/ ↩
  10. OpenAI, "Introducing OpenAI o3 and o4-mini," April 16, 2025 (o3-mini, January 31, 2025). https://openai.com/index/introducing-o3-and-o4-mini/ ↩
  11. OpenAI, "Introducing GPT-5," August 7, 2025. https://openai.com/index/introducing-gpt-5/ ↩
  12. OpenAI, "GPT-5.1: A smarter, more conversational ChatGPT," November 12, 2025. https://openai.com/index/gpt-5-1/ ↩
  13. OpenAI, "Introducing GPT-5.2," December 11, 2025. https://openai.com/index/introducing-gpt-5-2/ ↩
  14. OpenAI, "Introducing GPT-5.3-Codex," February 5, 2026; GPT-5.3 Instant, March 3, 2026. https://openai.com/index/introducing-gpt-5-3-codex/ ↩
  15. TechCrunch, "OpenAI launches GPT-5.4 with Pro and Thinking versions," March 5, 2026. https://techcrunch.com/2026/03/05/openai-launches-gpt-5-4-with-pro-and-thinking-versions/ ↩
  16. OpenAI, "Introducing GPT-5.5," April 23, 2026; Fortune coverage, April 23, 2026. https://openai.com/index/introducing-gpt-5-5/ ↩
  17. OpenAI, "Previewing GPT-5.6 Sol: a next-generation model," June 26, 2026; Axios and VentureBeat coverage, June 26, 2026. https://openai.com/index/previewing-gpt-5-6-sol/ ↩
  18. OpenAI Help Center, "Model Release Notes." https://help.openai.com/en/articles/9624314-model-release-notes
  19. AI Wiki, "AI Model Release Timeline (2022-2026)." https://aiwiki.ai/wiki/ai_model_release_timeline

Improve this article

Add missing citations, update stale details, or suggest a clearer explanation.

Suggest edit