humans& releases &Tutor: RL-trained on GLM-5.3 against its Persimmon learner model, it beats Claude Sonnet 5.5 and GPT-6 Astra on TutorMoments by default

Why it matters

【Key points】

humans& released &Tutor (API model ampersand-tutor-v0.1) on 2026-10-30, a GLM-5.3-based model RL-trained to teach, with reward tied to how much a simulated student learns; the student is humans&' Persimmon user model. On TutorMoments (real tutoring moments annotated by math teachers, LLM-judged), &Tutor scores 0.754 under a plain prompt vs 0.500 for base GLM-5.3, 0.667 Claude Sonnet 5.5, 0.634 GPT-6 Astra and 0.512 Gemini 4 Argon; with an evaluation-aware prompt it scores 0.919 vs GPT-6 Astra's 0.926 and leads on appropriate rigor (0.931). An ablation using GLM-5.3 as the learner reaches only 0.635, showing learner-model quality drives tutor quality. Available now via the Ampersand API at $1.00/M input, $0.20/M cached, $5.20/M output.

【Verifiable】

https://tutor.humansand.ai/blog/ ; https://ampersand.humansand.ai/docs/#tutor-model ; https://ampersand.humansand.ai/docs/PRICING.html

【Our read】

Unlike ChatGPT Study Mode, Claude's learning mode or Google LearnLM, which steer general models with prompts and pedagogy tuning, &Tutor optimizes directly for learning gains, and its learner ablation is the strongest evidence yet for humans&' user-model-as-environment strategy. But once graders' criteria are spelled out, GPT-6 Astra catches up, and all results are vendor-reported and simulation-based; real classroom trials are the real test.