Persimmon makes AI users harder to fake

humans& released Persimmon, a 550B user model that fooled LLM judges 18.6% to 21.1% of the time in multi-user conversation tests, versus 0% to 2.7% for comparison models. Its research preview is built to simulate varied human behavior rather than optimize assistant task completion, with gated playground and API access. Builders should use it as evaluation input for agents that must handle hesitation, oversharing, and error recovery; our OSWorld evidence on stale agent benchmarks shows why synthetic users need versioned tests too.