Reaction
Epoch AI reports GPT-6.2 Astra packs 12.3 reasoning steps into one token
Epoch AI says reasoning without chain of thought may make models harder to monitor.
TLDR
In measurements it plans to finalize, Epoch AI reports GPT-6.2 Astra performs 12.3 complexity-normalized serial reasoning steps in a single output token without chain of thought, versus 10.5 for GPT-6 Astra and 8.6 for GPT-6.1 Sol. The accuracy threshold is 25% above chance. Epoch cautions that models capable of such reasoning may be harder to monitor.
Combined views
5.3K
2 Sources, first seen ago
Comments