2026-11-02 · America/Los_Angeles · 社区动态 · #19

Is "Codex gets slower in long sessions" a model problem or a context-budget problem?

Something keeps coming up in long coding sessions: two agents on comparable models, one stays flat while the other degrades — slower first token, longer turns, and by the end the context is basically unusable for anything that needs follow-through. Almost everyone frames this as a model speed comparison. I lean the other way: I think we're comparing context management, not inference. Three different things get called "slow," and they're not the same: slower first token (prompt length, cache miss, queueing) longer per-turn thinking (more history to re-read) more turns to finish (the model…

热度 44.1 / 100；排名与评分保留该期记录。