OpenAI's large language models
Jump to navigation
Jump to search
OpenAI's large language models are a series of large language models (LLMs) developed by the research lab OpenAI. These models, built on the Transformer architecture, have become a key factor in the development of generative artificial intelligence. Starting with the GPT-1 model, introduced in 2018, each subsequent generation—including GPT-2, GPT-3, GPT-4, and newer multimodal systems like GPT-4o and the O-series family—has demonstrated exponential growth in capabilities, scale, and impact.
Evolution of GPT Models
GPT-1 (2018)
The first model in the series, introduced in 2018.
- Parameters: 117 million.
- Architecture: 12-layer Transformer decoder.
- Training: Trained on the BookCorpus dataset (~7,000 unpublished books).
- Key Innovation: Demonstrated the effectiveness of the two-stage approach (pre-training + fine-tuning), laying the foundation for all subsequent models. It proved that a single model could be adapted to numerous NLP tasks without architectural changes.
GPT-2 (2019)
A significant scaling-up from GPT-1.
- Parameters: 1.5 billion (~10x more than GPT-1).
- Architecture: 48-layer Transformer decoder.
- Training: Trained on the WebText corpus (40 GB of high-quality text filtered from the internet).
- Key Innovation: Showcased impressive zero-shot learning capabilities, i.e., solving tasks without specific fine-tuning. It could generate long and coherent texts. Its release was accompanied by public debate about the risks of misuse, which led OpenAI to initially release only smaller versions of the model.
GPT-3 (2020)
The model that brought a breakthrough in the capabilities and public perception of LLMs.
- Parameters: 175 billion (~100x more than GPT-2).
- Architecture: 96-layer Transformer decoder.
- Training: Trained on a mixture of datasets totaling ~570 GB, including Common Crawl, books, and Wikipedia.
- Key Innovation: Emergence of strong few-shot learning abilities—the model could solve tasks after being given just a few examples in the prompt itself. GPT-3 was the first model OpenAI offered via a commercial API, kicking off a boom in generative AI startups.
InstructGPT and GPT-3.5 (2022)
A family of models focused on improving controllability and usefulness.
- Parameters: Comparable to GPT-3 (~175 billion).
- Training: The RLHF method was applied on a massive scale for the first time to teach the model to follow instructions better, be more truthful, and less toxic.
- Key Innovation: A sharp increase in the model's "alignment" and safety. The gpt-3.5-turbo model formed the basis of the first release of ChatGPT, which launched on November 30, 2022, and became a global phenomenon.
GPT-4 (2023)
A new flagship model that marked the transition to multimodality.
- Parameters: Not officially disclosed (estimates are ~1.7 trillion, possibly with a Mixture-of-Experts architecture).
- Architecture: Multimodal Transformer.
- Training: Trained on a vast corpus of text and images.
- Key Innovation: Multimodality—the ability to accept not only text but also images as input. It demonstrated human-level (and even superhuman) performance on many professional and academic benchmarks (e.g., the bar exam).
GPT-4 Turbo (2023)
An optimized and more affordable version of GPT-4.
- Parameters: Similar to GPT-4.
- Context Window: Increased to 128,000 tokens (~300 pages of text).
- Training: Updated knowledge (up to April 2023).
- Key Innovation: A significant reduction in API call costs, improved instruction-following, and more recent knowledge, making the power of GPT-4 accessible to a wider range of applications.
GPT-4o (2024)
The "omni-model," which natively processes multiple modalities.
- Key Innovation: Native multimodal processing of text, audio, and images in real-time within a single model. This enables very fast and natural responses, comparable to the speed of human conversation. GPT-4o made GPT-4-level capabilities available to free ChatGPT users.
O-series Family: o1 and o3 (2024-2025)
A new generation of models focused on developing reasoning abilities.
- o1 model (September 2024): Introduced as a significant step forward in cognitive functions, allowing it to solve more complex problems that require deep analysis and multi-step reasoning.
- o3 model (January 2025): A further development of the ideas in o1, with even higher scores on complex logic and math tests (e.g., 96.7% on the 2024 AIME exam).
- Key Innovation: A focus not just on text generation but on constructing logical chains (Chain-of-Thought) and solving complex problems, bringing AI closer to more abstract thinking.
Models
| No. | Date | Model | Note | Source | Developer | System Card |
|---|---|---|---|---|---|---|
| 1 | 2018-06-11 | GPT / GPT-1 | First GPT-line model; transformer-based language model trained with unsupervised pre-training and supervised fine-tuning. | OpenAI | — | — |
| 2 | 2019-02-14 | GPT-2 | Large unsupervised language model; released gradually because of misuse concerns. | OpenAI | — | — |
| 3 | 2020-05-28 | GPT-3 | 175B-parameter language model demonstrating strong few-shot learning. | OpenAI | — | — |
| 4 | 2022-01-27 | InstructGPT | Instruction-following GPT-3-based models trained with human feedback. | OpenAI | — | — |
| 5 | 2022-11-30 | ChatGPT / GPT-3.5 | Conversational model related to InstructGPT; launched as a research preview. | OpenAI | — | — |
| 6 | 2023-03-01 | GPT-3.5 Turbo | API model family used in ChatGPT; optimized for chat and lower cost. | OpenAI | OpenAI API docs | — |
| 7 | 2023-03-14 | GPT-4 | Large multimodal model accepting image and text inputs and producing text outputs. | OpenAI | OpenAI API docs | arXiv:2303.08774 CDN PDF |
| 8 | 2023-11-06 | GPT-4 Turbo | More capable and cheaper GPT-4 variant with a 128K context window. | OpenAI | OpenAI API docs | — |
| 9 | 2024-05-13 | GPT-4o | Omni flagship model for real-time reasoning across audio, vision, and text. | OpenAI | OpenAI API docs | arXiv:2410.21276 CDN PDF |
| 10 | 2024-07-18 | GPT-4o mini | Smaller and cheaper GPT-4o-class model for cost-efficient intelligence. | OpenAI | OpenAI API docs | Covered in GPT-4o System Card: arXiv:2410.21276 CDN PDF |
| 11 | 2024-09-12 | OpenAI o1-preview | First public preview model in the o-series, designed to spend more time reasoning before answering. | OpenAI | OpenAI API docs | CDN PDF |
| 12 | 2024-09-12 | OpenAI o1-mini | Separate smaller reasoning model; faster and more cost-efficient, especially for STEM and coding. | OpenAI | OpenAI API docs | arXiv:2412.16720 CDN PDF |
| 13 | 2024-12-05 | OpenAI o1 | Full o1 reasoning model; successor to o1-preview. | OpenAI | OpenAI API docs | arXiv:2412.16720 CDN PDF |
| 14 | 2024-12-05 | o1 pro mode / o1-pro | Higher-compute version of o1 for harder problems and more reliable answers. | OpenAI | OpenAI API docs | — |
| 15 | 2025-01-31 | OpenAI o3-mini | Small, fast, cost-efficient reasoning model with strong STEM, math, and coding performance. | OpenAI | OpenAI API docs | CDN PDF |
| 16 | 2025-02-27 | GPT-4.5 | Research preview GPT model focused on broader pre-training and non-reasoning intelligence. | OpenAI | OpenAI API docs | CDN PDF |
| 17 | 2025-04-14 | GPT-4.1 | API model with improvements in coding, instruction following, and long-context use. | OpenAI | OpenAI API docs | OpenAI System Card |
| 18 | 2025-04-14 | GPT-4.1 mini | Smaller and faster version of GPT-4.1. | OpenAI | OpenAI API docs | OpenAI System Card |
| 19 | 2025-04-14 | GPT-4.1 nano | Smallest and lowest-cost version of GPT-4.1. | OpenAI | OpenAI API docs | OpenAI System Card |
| 20 | 2025-04-16 | OpenAI o3 | Advanced reasoning model for coding, math, science, and visual reasoning. | OpenAI | OpenAI API docs | CDN PDF |
| 21 | 2025-04-16 | OpenAI o4-mini | Smaller reasoning model optimized for speed, cost, and strong multimodal reasoning. | OpenAI | OpenAI API docs | CDN PDF |
| 22 | 2025-06-10 | OpenAI o3-pro | Higher-compute version of o3 designed to think longer and provide more reliable answers. | OpenAI | OpenAI API docs | Covered by o3 System Card: CDN PDF |
| 23 | 2025-08-05 | gpt-oss-120b | Open-weight reasoning model released under the Apache 2.0 license. | OpenAI | OpenAI API docs | — |
| 24 | 2025-08-05 | gpt-oss-20b | Smaller open-weight reasoning model released under the Apache 2.0 license. | OpenAI | OpenAI API docs | — |
| 25 | 2025-08-07 | GPT-5 | Unified flagship model system with fast and thinking capabilities. | OpenAI | OpenAI API docs | arXiv:2601.03267 CDN PDF |
| 26 | 2025-08-07 | GPT-5 mini | Smaller GPT-5 API model balancing capability, cost, and latency. | OpenAI | OpenAI API docs | Covered in GPT-5 System Card: arXiv:2601.03267 CDN PDF |
| 27 | 2025-08-07 | GPT-5 nano | Smallest GPT-5 API model for lowest-latency and lowest-cost use cases. | OpenAI | OpenAI API docs | Covered in GPT-5 System Card: arXiv:2601.03267 CDN PDF |
| 28 | 2025-08-07 | GPT-5 Pro / GPT-5 Thinking Pro | Extended-reasoning GPT-5 variant for harder tasks. | OpenAI | OpenAI API docs | Covered in GPT-5 System Card: arXiv:2601.03267 CDN PDF |
| 29 | 2025-11-12 | GPT-5.1 Instant | More conversational GPT-5.1 chat model with improved instruction following. | OpenAI | OpenAI API docs | CDN PDF |
| 30 | 2025-11-12 | GPT-5.1 Thinking | GPT-5.1 reasoning model with more adaptive thinking time. | OpenAI | OpenAI API docs | CDN PDF |
| 31 | 2025-12-11 | GPT-5.2 Instant | Instant variant of GPT-5.2 for ChatGPT and API use. | OpenAI | OpenAI API docs | CDN PDF |
| 32 | 2025-12-11 | GPT-5.2 Thinking | Reasoning variant of GPT-5.2 with stronger scientific and mathematical performance. | OpenAI | OpenAI API docs | CDN PDF |
| 33 | 2025-12-11 | GPT-5.2 Pro | Pro variant of GPT-5.2 for harder reasoning and professional work. | OpenAI | OpenAI API docs | CDN PDF |
| 34 | 2026-02-05 | GPT-5.3-Codex | Specialized agentic coding model for Codex-style long-running software tasks. | OpenAI | OpenAI API docs | CDN PDF |
| 35 | 2026-03-03 | GPT-5.3 Instant | Updated everyday ChatGPT model focused on smoother and more useful answers. | OpenAI | OpenAI API docs | DSH PDF |
| 36 | 2026-03-05 | GPT-5.4 | General-purpose frontier model with native computer-use capabilities. | OpenAI | OpenAI API docs | DSH PDF |
| 37 | 2026-03-05 | GPT-5.4 Pro | Pro version of GPT-5.4 for more complex professional and agentic tasks. | OpenAI | OpenAI API docs | DSH PDF |
| 38 | 2026-03-17 | GPT-5.4 mini | Smaller and faster GPT-5.4-class model for coding, tool use, and multimodal reasoning. | OpenAI | OpenAI API docs | DSH appendix |
| 39 | 2026-03-17 | GPT-5.4 nano | Smallest GPT-5.4-class model for high-volume and low-cost API workloads. | OpenAI | OpenAI API docs | — |
| 40 | 2026-04-23 | GPT-5.5 | Flagship GPT-5.5 model for professional work, coding, research, and analysis. | OpenAI | OpenAI API docs | DSH PDF |
| 41 | 2026-04-23 | GPT-5.5 Pro | Pro version of GPT-5.5 for more demanding reasoning and professional tasks. | OpenAI | OpenAI API docs | DSH PDF |
| 42 | 2026-05-05 | GPT-5.5 Instant | Updated default ChatGPT model with improved accuracy, clarity, and personalization. | OpenAI | OpenAI API docs | DSH |
| 43 | 2026-06-26 | GPT-5.6 Sol | Limited-preview flagship model in the GPT-5.6 series. | OpenAI | — | DSH PDF |
| 44 | 2026-06-26 | GPT-5.6 Terra | Limited-preview balanced GPT-5.6 model for everyday work. | OpenAI | — | DSH PDF |
| 45 | 2026-06-26 | GPT-5.6 Luna | Limited-preview fast and affordable GPT-5.6 model. | OpenAI | — | DSH PDF |
External links
- Official OpenAI website
- Generative pre-trained transformer — Wikipedia
- OpenAI — Wikipedia
- List of large language models — Wikipedia
Literature
- Radford, A. et al. (2018). Improving Language Understanding by Generative Pre-Training. OpenAI.
- Radford, A. et al. (2019). Language Models are Unsupervised Multitask Learners. OpenAI.
- Brown, T. et al. (2020). Language Models are Few-Shot Learners. arXiv:2005.14165.
- OpenAI (2023). GPT-4 Technical Report. arXiv:2303.08774.
- Ouyang, L. et al. (2022). Training language models to follow instructions with human feedback. arXiv:2203.02155.