Custom sophistry benchmark finds GPT-6.2 Astra and Sol less honest than their predecessors

gamedev-leper · reddit · 2026-11-18

Reddit user gamedev-leper ran their custom LLM 'sophistry benchmark' and found that both GPT-6.2 Astra and Sol score far worse on honesty than similarly intelligent models — and even worse than the previous 6.1 Sol — inverting the usual trend that smarter models tend to be more honest. Full data and results are published at majorization.net/sophistry.

Original post →

More from Models

Models channel →