长上下文 MRCR 对比：Sapience 97.7 略胜 Astra 96.3

Aleksa Gordić (水平问题) @gordic_aleksa原推发布于 2026.09.15

在 X 上打开原推

i find these long-context results quite surprising (promising?) on OpenAI's MRCR benchmark @Sapience_Labs scores 97.7 whereas GPT-6 Astra scores 96.3 Sapience is a custom system that reads in only ~15K tokens and an open model (DeepSeek V4 Pro) is doing the answering (Astra has to read the whole thing to answer, i.e. from 500k-1M tokens) [not affiliated with them, @lvrzhn is the guy behind it, i just find it cool]