Model providers should start publishing 'Usable Context' as a metric. Just because a model can support 10M tokens doesn't mean it can use all 10M effectively. Often, I find models >32K can only effectively use a smaller portion of their context window.

→ View original post on X — @mattshumer_,