gpt-6IMAGE
AgentTime paper: GPT-6 Astra hits 63% on duration following
A new paper tests agents' ability to control their own runtime and judge elapsed time; on duration-following, GPT-6 Astra beats Fable 5.1 63% to 4%, and in some runs Astra chooses to sleep instead of working.
Michael Ofengenden
@MOfengendenArchived 24 days agoOriginally posted 2026.10.08
Michael Ofengenden @MOfengendenOriginally posted 2026.10.08
New paper: AgentTime. We ask agents to control their own runtime, forecast how long tasks take, and understand how much time passed. On duration-following, GPT-6 Astra beats Fable 5.1, 63% to 4%. In some runs, Astra likes to sleep instead of continuing work. W/@maksym_andr
x.com/mofengenden/status/2108221696138510481Archived 2026.10.09 10:38