You should be doing everything you can to keep context under 200k, ideally even 100k. All the models unwind so badly as context grows.