GPT 6.1 的 token 效率秘密：靠上下文剩余量施压？
Оригинальный заголовок: Gpt 6.1 token efficacy secret?
Заголовок и краткое изложение на выбранном языке ожидают перевода.
有用户观察到 GPT 6.1 似乎更省 token，猜测原因是后训练或注入机制让模型感知到上下文剩余量，从而产生压力。例如模型会主动提及"还剩约 26k""37000 token 上下文已用 28000、剩 18000"等剩余额度，并据此压缩工作范围。
Полный текст на выбранном языке ожидает перевода. Пока показан оригинал.
it seem like it use less token because it post trained or injected with small token amount left in context to pressure it.
E.g. i constantly see things like
" I’m considering how to tackle this project where I have about 26k left. "
" I'm considering my current progress, with around 18,000 tokens left in a 37,000-token context and 28,000 tokens used so far. Time is limited, so I really don't want to dwell on this. "
" Figuring out token limits
I need to figure out how fast I can work with the remaining tokens, which could be 18k, 24k, or even 30k. I think it’s best to keep the scope small for now. "
Источник: Reddit · ClaudeCode / Codex / VibeCoding · reddit.com