101 milestones in six eras
Research, companies and products, people and society, security incidents — each milestone dated, filterable by type, searchable by name.
Trend check: how context windows grew
A glossary term followed through time — how much text a model can read at once went from a paragraph to whole libraries. More trend checks to come.
| Year | Model | Context window | Roughly equals |
|---|---|---|---|
| 2019 | GPT-2 | 1,024 tokens | a long email |
| 2020 | GPT-3 | 2,048 tokens | a short article |
| 2022 | ChatGPT (GPT-3.5) | 4,096 tokens | an essay |
| 2023 | GPT-4 | 8k → 32k tokens | a long report |
| 2023 | Claude 2.1 | 200,000 tokens | a 500-page book |
| 2024 | Gemini 1.5 Pro | 1–2 million tokens | a small bookshelf |
| 2025 | Llama 4 Scout | 10 million tokens (claimed) | an encyclopedia |
~10,000× in six years — the quiet revolution behind “paste in the whole codebase”. See Context Window and Quadratic Attention Cost in the glossary.
Data state: 30 September 2026