The most prestigious law school admissions discussion board in the world.
|
 |
 
|
NSAM vindicated against fuckwits who promote OpenAI and Grok (lol)
| https://i.imgur.com/YbMiOBE.png | 09/16/26 |
Poast new message in this thread
Date: September 16th, 2026 8:07 PM Author: https://i.imgur.com/YbMiOBE.png How much does context rot cost on agentic coding tasks? We prefilled 250k–500k tokens of context and ran GPT-5.6 Sol and Claude Opus 5 on terminal-bench tasks to find out.
Sol performance drops significantly at longer contexts, even when the context is completely unrelated to the task. Opus 5 holds flat on unrelated context but degrades when the context is related.
https://x.com/sanyamsatia/status/2100019768455750124 |