A slightly contrarian observation from measuring long-running Codex workloads:
I am no longer convinced that a larger context window is always the right metric for professional agent workloads.
In our audit, the expensive failure mode was not “the model cannot fit enough information.” It was that completed execution history remained economically active and was repeatedly processed during later orchestration.
We measured individual MAIN turns above 200k input tokens in the old workflow, including very large turns whose semantic purpose was essentially checking external progress.
A 1M window certainly postpones compaction, but it may also allow much more completed history to remain active before the system is forced to deal with it.
I would therefore love to see benchmarks report not only maximum context supported, but peak active context required per successfully completed task.
Bigger RAM and better memory management solve different problems.