Humans& announced Persimmon v0.1 Pro and Ultra on November 20, saying the new human-conversation simulation tiers are now available to the public. In a quote-post, @alexisjross said both are accessible through the company’s API without needing to ask to be taken off the waitlist. Less queue management, more simulated conversation. Source Source
Humans& says the tiers use test-time scaling—more computation during response generation—to make conversations more realistic. It says Ultra generates more output tokens, or units of text, than Pro, and Pro more than standard Persimmon v0.1. The launch text doesn’t detail the scaling mechanism or give token counts, prices or latency figures. Source Source
Humans& reports that Ultra scored 48% on its eight-turn Multi-User Turing Test, up from Persimmon’s 18%. The company describes 50% as the point where a model is indistinguishable from real people “according to our held-out judge.” Ultra’s reported score remains below that reference point. Source
The result doesn’t establish that people generally cannot distinguish Ultra from humans. The available post text doesn’t identify the judge or supply sample sizes, scoring details, statistical uncertainty or independent replication. The Turing-test label doesn’t get to skip the fine print. Source
Humans& also reports improvements for both tiers across three datasets: TIDES, TutorMoments and its internal workspace conversations. In another post, it says Pro and Ultra are better at following supplied profiles and simulating people accurately. Source Source
In a quote-post, @ericzelikman said making the user-model API widely available required substantial work on release safety. @alexisjross separately credited @polaris_7369 with developing safety filters. Source Source
Those posts don’t explain what the filters block or how reliably they work. Source Source
Comments