The catch-up lag that used to take 12–18 months has compressed to roughly 6–9 months — and the window keeps shrinking.
The shaded amber region is the catch-up window — the time it takes for the best open-source models to reach where closed-source was. A few key moments worth noting: The gap peaked at roughly 15–20 quality index points in early 2024. Then came a rapid compression driven by Chinese AI labs. The “DeepSeek moment” in early 2025 saw DeepSeek R1 demonstrate ChatGPT-level reasoning at significantly lower training costs , which was a major inflection point. By 2025–2026, the best open-source model (MiniMax-M2, scoring ~61) trails the best proprietary model (GPT-5, scoring ~68) by just 7 points, down from 15–20 points a year prior.