How the context window grew 4,883×
From 2K to 10M in 1,097 days, measured across 131 releases that disclose one.
6 of the 131 releases with a disclosed context window set a new maximum. The rest shipped under a ceiling someone had already reached — which is the useful shape here: the frontier moves in a few large steps, not continuously.
Every step of the frontier
A release appears here only if it disclosed a larger context window than anything before it. 149 of 280 records disclose no context window at all and cannot appear — see data quality.
The largest single jump
Gemini 1.5 Pro multiplied the frontier by 16×, from 128K to 2.1M, 30 days after GLM-4 set the previous mark. Both figures are traced to the labs' own announcements — follow them from Gemini 1.5 Pro.
What this cannot tell you
A context window is what the lab says the model accepts. It is not a claim about what the model uses well: retrieval quality across a full window is a benchmark question, and this dataset records no benchmark it has not traced to a primary source. Nor is the absence of a figure evidence of a small window — 149 records disclose nothing here, and most of those are proprietary models whose labs publish no specification at all.