Persimmon Ultra hits 50% fool rate in multi-user Turing test while frontier models score near 0%

niloofar_mire · x · 2026-11-21

Humansand released Multi-User Turing Test (K=8) results across three datasets: Persimmon v0.1 Ultra reaches 41.9% on TIDES, 33.9% on internal workspace conversations, and 49.7% on TutorMoments, with Pro close behind. Base Persimmon scores 19-21%, while frontier models — GPT 5.6, GPT 6 (Astra), Opus, Fable 5, Kimi K3, GLM 5.3 — mostly score under 1%, often 0%. With 50% representing human parity, this suggests models purpose-built for multi-user human simulation can nearly pass as people in group chats, while general frontier models cannot at all.

Related event: Humansand Launches Persimmon v0.1 Pro/Ultra, Doubling Turing Test Scores(9 posts)→

Original post →

More from Models

Models channel →