Context window size history timeline
In just a few years, the amount of text you can feed a language model jumped from a couple of paragraphs to entire books. This interactive timeline plots the maximum context window of major models from 2020 onward, with the input price per million tokens overlaid so you can see that longer context arrived and got cheaper at the same time.
How it works
Each model is positioned by release date and drawn with a bar scaled to its maximum context window in tokens. Because the range spans from ~2,000 tokens (GPT-3) to over 1,000,000 (Gemini 1.5), the bars use a logarithmic scale so the early models remain visible next to the giants. The launch-era input price sits beside each entry to show the cost trend. You can add your own entry to drop a new release or an internal model onto the same curve.
Key milestones in the context race
2020 — GPT-3 at 2,048 tokens. This was the ceiling for general-purpose language models at launch. Two thousand tokens is enough for a few paragraphs of conversation or a short article — useful, but nowhere near enough for documents, books, or long code files.
2023 — Claude 2 and GPT-4 push past 100k. Anthropic shipped Claude 2 with a 100,000-token context window, roughly 75,000 words or a short novel. GPT-4 moved from 8k to 32k and then 128k on subsequent releases. This shift made single-document analysis genuinely practical without retrieval.
2024 — The million-token era. Gemini 1.5 Pro launched with a 1-million-token context, enough to process entire codebases or hours of video transcripts in a single call. Gemini 1.5 Flash and Claude 3 models added comparable long-context capabilities at lower price points, signalling that million-token windows were becoming a table-stakes feature rather than a headline differentiator.
What a bigger window actually unlocks
Each jump in context length opened genuinely new use cases that were impossible before:
| Window size | What became possible |
|---|---|
| 2k–8k | Short conversations, single-page document Q&A |
| 32k–100k | Full research papers, long code files, multi-document analysis |
| 100k–200k | Book-length documents, full codebase review, long transcripts |
| 1M+ | Entire repositories, multi-hour recordings, thousands of documents at once |
The practical limit is no longer window size for most tasks — it is the cost of filling the window and the model’s ability to retrieve and reason over information buried deep in the middle of very long inputs.
What the trend shows
- Roughly 10× every couple of years. From 2k → 8k → 32k → 128k → 1M, the ceiling has climbed in big steps.
- Cheaper per token, not just bigger. Input prices fell even as windows grew, which is what made long-context use cases practical.
- Diminishing real-world gains. A bigger window helps only if the model uses the middle of it well — retrieval and good chunking still matter.
- Context is now rarely the bottleneck. For most tasks the limit is reasoning quality and cost, not how much you can paste.