Is the catch-up window actually shrinking?

I already talk about this in the introduction, but I want to see if the data actually supports it.

When GPT-4 came out, it took almost a year for another lab to release a model that reached a similar level. That feels very different from today, where closed and open-weight models seem to follow the frontier within weeks or months.

A first timeline

Scores come from our Artificial Analysis v4.1.1 snapshot observed on 15 August 2026.

What does “caught up” even mean?

This is where it becomes messy. Is a model caught up when it has the same benchmark score? When it is within two points? Within five? Or when it is simply good enough for the same task?

The answer changes the window. Sol followed Fable within two points after 30 days. Qwen followed within five points after 64 days. Those are not the same definition of catching up.

What the data says so far

It supports the feeling that competition moves much faster now. It does not yet prove a clean historical trend.

Older models were scored again using today’s index rather than recorded on the day they came out. Model families are also not directly comparable, and one benchmark does not tell us whether two models are equally useful for a specific task.

So for now, I would still say the catch-up window seems to be shrinking. I don’t think we can say more than that yet.

What we would need to test

We need one stable definition of caught up, historical scores measured in the same way and enough frontier releases to see whether the window consistently becomes shorter.