Every release pushes the window higher and the per-token price lower, but past a couple hundred thousand tokens things get mushy. Vendor retrieval benchmarks claim near-perfect scores, yet I still watch models miss a line buried mid-document. Has anyone found a model that genuinely holds up across its full advertised window, or are we paying for spec-sheet numbers?
Long context is cheap now, but is it actually usable?
Replies
No comments yet — be the first to share your thoughts.