AI Tutoring Has the Cleanest ROI Story in Education
The Bloom 2-sigma result keeps surfacing in district and campus AI conversations because it's not subtle — students working one-on-one with a tutor outperform students in conventional classroom-only instruction by roughly two standard deviations. Human tutors at scale have never been economically feasible. AI tutors are, and the early evidence (Khan Academy's Khanmigo deployments, GSU's iAdvise pilots, several state-DOE initiatives) suggests the effect is showing up in measurable outcomes.
The vendors that own this space — Khanmigo, MagicSchool, Curipod, Brisk Teaching, Riiid Tutor, plus the per-seat options like ChatGPT Edu and Microsoft 365 Copilot Edu — price per-student or per-teacher. Either pricing shape scales linearly with enrollment, not with the actual tutoring work done.
The actual tutoring work, priced by tokens, costs almost nothing. The math is the post.
What a Tutoring Session Actually Costs Per Token
A typical tutoring session — student asks a question, the agent responds with scaffolded guidance, the student responds, another exchange or two — is about 1,000 input tokens (context, student profile, prior turn, current question) and 1,500 output tokens (response with worked example or Socratic question). Cost per session on the major models:
| Model | Input ($/MTok) | Output ($/MTok) | $ per session | When to use it |
|---|---|---|---|---|
| Claude Opus 4.7 | $15 | $75 | $0.128 | Graduate-level, complex reasoning subjects |
| GPT-5 | $10 | $30 | $0.055 | AP / college-level subjects |
| Claude Sonnet 4.6 | $3 | $15 | $0.025 | Standard K-12 / undergrad tutoring workhorse |
| Gemini 3 Flash | $0.35 | $1.05 | $0.002 | High-volume quick Q&A, vocab, math-check |
| Claude Haiku 4.5 | $1 | $5 | $0.009 | Elementary, supplementary practice |
| Qwen 3 (self-hosted, multilingual) | ~$0 | ~$0 | ~$0 | Spanish, Mandarin, multilingual ELL programs |
| Llama 4 (self-hosted) | ~$0 | ~$0 | ~$0 | District-owned, FERPA-inside-perimeter |
Standard workhorse: 2.5 cents per session. Cheap hosted: two-tenths of a cent. Self-hosted: marginal cost is GPU time amortized across thousands of sessions per day.
Monthly Bills at Three Scale Tiers
- Small school (500 students, ~300 active tutoring users): ~3,000 sessions/month
- K-12 district (50,000 students, 8K active tutoring users): ~96,000 sessions/month
- University campus (30,000 students, 10K active tutoring users): ~120,000 sessions/month
Monthly cost using Claude Sonnet 4.6 vs the per-student edtech AI alternatives:
| Approach | Pricing shape | Small school (3K) | District (96K) | University (120K) |
|---|---|---|---|---|
| Specialty AI tutor (per-student) | ~$4–10/student/mo × enrollment | ~$2,500 | ~$250,000 | ~$150,000 |
| ChatGPT Edu | ~$25/teacher OR student | ~$1,250 | ~$1,250,000 | ~$750,000 |
| Microsoft 365 Copilot (Edu) | $30/user × enrollment | ~$1,500 | ~$1,500,000 | ~$900,000 |
| Direct API — Claude Sonnet 4.6 | Token-based | ~$75 | ~$2,400 | ~$3,000 |
| Direct API — Gemini 3 Flash | Token-based (cheapest) | ~$6 | ~$192 | ~$240 |
| ibl.ai self-hosted (Llama 4 / Qwen 3) | Flat license + GPU | ~$1,500 | ~$3,000–6,000 | ~$5,000–10,000 |
At district scale, the per-student edtech AI is ~70× more expensive than self-hosted; ChatGPT Edu is ~300× more expensive — for the same tutoring sessions delivered.
The Per-Student Pricing Trap
$4–10 per student per month feels modest until you multiply. A 50K-student district pays $200K–500K/month for an AI tutor most students touch occasionally and a few hundred use heavily. The vendor's pitch is "you pay a little for everyone so any student can use it" — but the bill scales with enrollment whether or not the students actually engage.
Token-based or self-hosted aligns the bill with the actual tutoring delivered. A district that gets 8K students using the tutor heavily this semester and 2K next semester pays for the actual sessions in both terms, not the headcount on the registrar's roster.
Why FERPA + Multilingual Equity Push Self-Hosted
Two specifically-education concerns that don't apply elsewhere:
FERPA scope on tutoring transcripts. A tutoring session log contains conversations about academic performance, learning struggles, accommodations needed — all FERPA-protected student record. Sending those transcripts to a managed AI vendor's cloud creates a FERPA data-processing relationship the district counsel reviews per vendor; every model update or sub-processor change re-papers it. Self-hosted keeps the transcripts inside the district's existing student-information-system perimeter.
Multilingual learners. Districts serving Spanish-, Mandarin-, Arabic-, or Haitian-Creole-speaking students need tutoring in the student's first language. Managed AI vendors typically support a fixed set of languages with a fixed translation quality. A self-hosted Qwen 3 or Llama 4 deployment lets the district configure language support to match its actual student population — and update the configuration when the demographics shift.
What Stays the Same, What Changes
Self-hosting AI tutoring doesn't mean rebuilding the school's edtech stack. The student-facing chat UI, the teacher dashboards, the parent-progress reports, the LMS integration (Canvas, Schoology, Google Classroom via LTI 1.3), the SIS integration (PowerSchool, Infinite Campus, Banner), the audit logs the registrar relies on — all stays managed by ibl.ai. The compute, the model, and the student data move inside the district or campus VPC.
What disappears: the $200K–1.5M/month per-student or per-seat bill at district scale.
What appears: a district- or campus-owned AI tutoring capability with a model-routing recipe the curriculum team designed:
- Opus for AP / graduate-level subjects requiring deep reasoning
- Sonnet for standard K-12 and undergrad tutoring (the bulk)
- Haiku for elementary practice and supplementary drilling
- Qwen 3 self-hosted for multilingual sections — Spanish, Mandarin, Arabic, others
- Gemini Flash or Llama 4 self-hosted for the highest-volume routine Q&A where pennies matter
Run the Numbers for Your School, District, or Campus
For the segment-wide cost-math context (K-12), see AI Cost Math for K-12 Districts: Per-Seat vs Usage-Based in 2026.
For higher ed specifically, see AI Cost Math for Higher Education: Per-Seat vs Usage-Based in 2026 plus the 5 higher-ed calculators (advising ROI, retention impact, financial aid time savings, enrollment ROI, LMS integration cost).
For multilingual student support, Qwen 3 for Education: Multilingual AI Tutoring covers running the multilingual model self-hosted.
For the deployment comparison side-by-side — K-12 — see Self-Hosted AI vs ChatGPT Enterprise for K-12. Higher ed: Self-Hosted AI vs ChatGPT Enterprise for Higher Education.
For the broader pricing landscape across every model and per-seat vendor, the hub: What Does AI Actually Cost in 2026?.
Why Family-Owned and New York Matters Here
For a district or campus, the AI tutoring vendor relationship is a multi-year commitment that touches FERPA-protected student records and core pedagogical decisions. ibl.ai is family-owned and operated from New York, NY — a long-term partner with a perpetual platform license and no investor exit pressure. The runtime is open source. The student data stays inside the district or campus network. The math works at a 500-student elementary school or a 200,000-student multi-campus system.