| GPT-5 |
2025 |
Hybrid (multi-model) |
Dynamic model routing, native multimodal, state-of-the-art coding and reasoning, “thinking” modes, enterprise APIs |
| GPT-4.1 |
2025 |
Not specified |
Outperforms GPT-4, faster and more efficient variants (mini, nano) |
| GPT-4o |
2024 |
Not specified |
Unified text, image, audio input/output; faster/cheaper; improved multimodal capability |
| GPT-4 |
2023 |
~1.76T (MoE est.) |
Multimodal (text+image), strong logical reasoning, larger context window (up to 128k tokens), safety improvements |
| GPT-3.5 |
2022 |
~175B (improved) |
Optimized for conversation, better instruction-following, more robust dialogue |
| GPT-3 |
2020 |
175B |
Few-shot/zero-shot learning, remarkable language generation, massive scale, improved fluency |
| GPT-2 |
2019 |
1.5B |
Larger model, improved generation quality, better coherence, trained on bigger dataset, use of advanced training tricks |
| GPT-1 |
2018 |
117M |
First Transformer-based model; general-purpose language modeling; moderate generation quality |