humans&, the company behind a new tutoring model called Ampersand Tutor (&Tutor), has opened it to developers through an API and says it out-teaches the big general-purpose chatbots. The claim comes from the company’s own blog at tutor.humansand.ai, and the method is the story: &Tutor never trained on a classroom. It trained on a pretend student.
That pretend student is Persimmon, which humans& calls a learner model. In plain words, it is an AI built to behave like someone who is learning. It guesses, gets things wrong, asks for help, gets frustrated, and now and then lands on an “ohhhh” moment. A tutor needs a huge number of practice conversations, and a simulated student can supply them without anyone’s child in the room.
The reward rule is simple. &Tutor is scored on how much the simulated student learns, not on any list of teaching tricks. According to humans&, habits like giving a hint, demanding rigor, and staying quiet all appeared on their own. In a sample exchange on the blog, the tutor does not hand over the answer. It asks Persimmon to multiply 10 by itself three times and see whether the result is 10, and the student’s realization follows.
The scores come from TutorMoments, a test assembled from recorded tutoring sessions. Experienced math teachers flag the moments where a tutor must choose between supporting a student and making them think harder. Each model then continues the conversation while an AI plays the student, and a language-model judge compares the reply with the teachers’ notes. With a plain instruction to tutor well, &Tutor scored 0.754. Its starting point, the open GLM-5.3 base, scored 0.500. GPT-6 Astra managed 0.634, and Claude Sonnet 5.5 managed 0.667. With a prompt that spells out the support-versus-rigor balance, &Tutor reached 0.919, just under GPT-6 Astra at 0.926, and posted the top rigor score of any model at 0.931.
The pretend student matters as much as the tutor. humans& ran the same training recipe with GLM-5.3 playing the learner instead of Persimmon. At low reasoning effort, the Persimmon-trained tutor ended at 0.659 and the other ended at 0.477.
To justify Persimmon, the company set simulated students beside real ones on four collections of tutoring conversations. It reports that Persimmon came closest to real learners on conversational moves, on emotions, and on error rates. On errors, the gap from real learners was 0.03 to 0.05, while the competing simulators sat between 0.16 and 0.33. Humans got 25 to 30 percent of their checkable statements wrong. Persimmon missed 19 to 33 percent, while every rival simulator missed 15 percent at most. Most simulated students, in other words, are too good at being students.
Independence is thin here. The post names Alexis Ross as an author, and the TutorMoments reference lists her among that benchmark’s authors too. The students in the test are AI, and the judge is an AI. humans& itself says the real test is how well &Tutor helps people learn over time, and it is asking education teams to run that test.
Until a school or a tutoring company does, 0.754 shows a tutor that matches teachers’ annotations in staged conversations. Anyone building a learning product on a general chatbot now has a reason to run the same comparison with real students before choosing a model.
Reported by the humans& blog (tutor.humansand.ai), in a post titled “Introducing &Tutor” by Alexis Ross and Manya Bansal; the page carries no publication date.
