What “GPT” means
Generative pre-trained transformer: a large-language-model that is first pre-trained on large text corpora to predict the next token, then adapted to tasks. The approach comes from OpenAI’s 2018 paper “Improving Language Understanding by Generative Pre-Training” (Radford et al.), built on the Transformer architecture from “Attention Is All You Need” (Vaswani et al., 2017). The term is used generically; the US Patent and Trademark Office reportedly found “GPT” descriptive and generic and refused OpenAI’s registration (Wikipedia). Other vendors’ models (Gemini, Claude, Llama, DeepSeek, Qwen) are also transformer LLMs but are not sold as “GPT”.
OpenAI GPT timeline
| Model | Date | Notes |
|---|---|---|
| GPT-1 | June 2018 | first GPT; parameter count not given in the source I could open |
| GPT-2 | Feb 2019 | 1.5B parameters |
| GPT-3 | May 2020 | 175B parameters; few-shot learning without fine-tuning (arXiv 2005.14165) |
| ChatGPT | 30 Nov 2022 | chat product built on GPT-3.5 |
| GPT-4 | Mar 2023 | accepts image and text input, text output (GPT-4 Technical Report) |
| GPT-5.x | 2025-2026 | GPT-5.4 (Mar 2026), 5.5 (Apr), 5.6 (Jul) |
| GPT-6 | Sept 2026 | Astra (2026-09-03), Sol and Luna (2026-09-22), 6.1 Sol (2026-09-29) |
Dates and sizes up to GPT-3.5 are per Wikipedia (secondary); GPT-5.x/6 dates are from vault notes, not re-checked here.
Current models and API IDs (no prices here; see the vendor pricing page): gpt-5-and-gpt-6-family, gpt-6-dot-astra, gpt-6-dot-sol, gpt-6-dot-luna, gpt-6-dot-1-sol. Older notes: gpt-3-dot-5, gpt-4-dot-1, openai-4o. The lab itself: openai.
Key characteristics
- Transformer backbone: self-attention over the context window.
- Pretraining then alignment: post-training (RLHF and successors) turns the base model into an assistant.
- Scale: GPT-2 to GPT-3 grew from 1.5B to 175B parameters; no parameter count is listed for GPT-4 in the source used.
- In-context / few-shot learning: GPT-3 showed tasks can be specified in the prompt (see in-context-learning).
Applications and limits
Content creation, code generation, chatbots, translation, agents. Limits: hallucination, training-data bias, misuse for disinformation, cost. Regulatory context: ai-regulation-and-policy.
Sources
- https://en.wikipedia.org/wiki/Generative_pre-trained_transformer (release dates, GPT-2/3 sizes, trademark; accessed 2026-09-30)
- https://arxiv.org/abs/1706.03762 (accessed 2026-09-30)
- https://arxiv.org/abs/2005.14165 (accessed 2026-09-30)
- https://arxiv.org/abs/2303.08774 (accessed 2026-09-30)
- GPT-5.x/6 dates: vault notes gpt-5-and-gpt-6-family and KB-AI-AUDIT-2026-09.md Phase 3a (vendor docs checked by R2; accessed 2026-09-30)
- OpenAI’s own GPT-1/GPT-2 pages returned 403; GPT-1 parameter count not verified.