It's one data point, but I think token share for open models will continue to rise.

You can get more insights by looking more closely at the multi-agent systems/patterns being adopted across domains. A supervisor leads a bunch of execution subagents (frontier open models work beautifully here). This is one of the most effective and widely used patterns today, where subagents dominate token use (which you can get from frontier open models at a much better price and efficiency). This trend will continue. There really is no point in using a Fable 5 for everything, where cheaper models will do just fine and at a faster and cheaper rate.