Even with the 1m context window, it looks like these models drop off significantly at about 256k. Hopefully improving that is a high priority for 2026.