The Short Answer
At district scale — 50,000 students and 3,000 teachers — per-seat AI costs $75–90K per month, with ChatGPT Edu at about $75K for teachers and Microsoft 365 Copilot at about $90K, before per-student tutoring licenses add $200–500K a year. The same workloads through the Claude Sonnet API cost about $2,931 per month. Self-hosted on ibl.ai they run $3,000–6,000 per month all-in, with FERPA-protected student data inside the district's own VPC because you own all the code and the data.
Districts feel per-seat pricing twice: once on staff licenses and again on per-student edtech tutors. Both line items scale with enrollment, which is the one number a district cannot reduce.
The 50,000-Student District Math: Per-Seat Either Way You Slice It
A mid-size school district has 50,000 students and 3,000 teachers. The per-seat AI vendors price both populations, but rarely sensibly:
- Per-teacher: $25–30/teacher/month × 3,000 = $75,000–90,000 per month ($900K–1.08M/year)
- Per-student: typical edtech AI runs $5–15/student/year, which is "low" until you multiply: 50K × $10 = $500,000 per year, on top of the teacher seats
Either way, the per-seat shape doesn't fit a district. Tutoring usage is concentrated on a few hundred high-engagement students; lesson planning is concentrated on departmental leads; IEP drafting is concentrated on the special-ed team. Buying a seat for every student and every teacher means the district pays for the long tail to subsidize the curve.
Token pricing — or a flat district-license platform — aligns the bill to the actual work, and lets the district keep FERPA-protected data on infrastructure it already runs.
The math is the post.
What the Latest Models Actually Cost in 2026
Token pricing across the major providers, approximate as of mid-2026:
| Model | Provider | Input ($/MTok) | Output ($/MTok) | K-12 fit |
|---|---|---|---|---|
| Claude Opus 4.7 | Anthropic | $15 | $75 | Differentiated lesson plans, IEP drafting |
| Claude Sonnet 4.6 | Anthropic | $3 | $15 | Writing feedback, tutoring sessions |
| Claude Haiku 4.5 | Anthropic | $1 | $5 | Quick Q&A, math-check, vocabulary |
| GPT-5 | OpenAI | $10 | $30 | Standards-aligned content generation |
| Gemini 3 Pro | $3.50 | $10.50 | Long-context unit planning, Workspace fit | |
| Llama 4 (70B, self-hosted) | Meta (open weights) | ~$0 | ~$0 | District-owned tutoring at scale |
| Qwen 3 (self-hosted) | Alibaba (open weights) | ~$0 | ~$0 | Multilingual (Spanish, Mandarin, etc.) |
For self-hosted open-weight models, the marginal cost is GPU time. A reserved H100 instance ($1.50–3/hour) supports thousands of tutoring sessions a day across a district.
A Real Workload: Tutoring + Lesson Planning + IEP Drafting
A district-wide AI deployment typically covers three flagship workloads. (For a deeper per-session cost breakdown of the tutoring workload specifically — including a side-by-side against Khanmigo, MagicSchool, Curipod, and Brisk Teaching — see What AI Tutoring Actually Costs in 2026 (K-12 + Higher Ed).)
- Tutoring — say 8,000 active students using the tutor 3× per week (96K sessions/month, each ~1,000 input + 1,500 output tokens)
- Lesson planning — 3,000 teachers averaging 4 plans/month (12K plans, each ~500 input + 2,500 output tokens)
- IEP drafting — 1,500 IEPs district-wide annually (~125/month, each ~3,000 input + 4,000 output tokens)
Combined: roughly 102M input + 175M output tokens per month. Front-loaded onto a fraction of students and teachers — exactly the population a per-seat vendor invoices for everyone.
What it costs by deployment shape
| Deployment | Pricing shape | Monthly cost | Annual | Student-data posture |
|---|---|---|---|---|
| ChatGPT Edu (teachers only) | Per-seat ($25/teacher × 3K) | $75,000 | $900,000 | OpenAI cloud (FERPA-aligned DPA) |
| Microsoft 365 Copilot (Edu) | Per-seat ($30/teacher × 3K) | $90,000 | $1,080,000 | Microsoft cloud (FERPA-aligned DPA) |
| Specialized edtech AI tutor | Per-student (~$10/student × 50K) | $41,667 | $500,000 | Vendor cloud (varies, often FERPA-aligned) |
| Direct API — Claude Sonnet 4.6 | Token-based | ~$2,931 | ~$35,172 | Anthropic cloud (district DPA) |
| ibl.ai self-hosted (Llama 4 / Qwen 3) | Flat license + GPU | ~$3,000–6,000 | ~$36,000–72,000 | Inside the district's VPC / on-prem |
The ibl.ai row covers GPU instance, platform license, and ongoing support — the entire district workload, FERPA-protected student data never leaving the district's infrastructure, with a model choice the district controls (English plus Spanish via Qwen 3 if the district serves multilingual learners).
Why Per-Seat Pricing Fails Harder in K-12
Three structural reasons:
1. Both populations are large and most don't use AI heavily. 50K students × any per-student fee is a big number; 3K teachers × any per-seat fee is a big number. But actual high-engagement AI users — the students working with the tutor multiple times per week, the teachers building units in the AI planner — are a fraction of either count. Per-seat invoices the long tail.
2. FERPA is a procurement event for every vendor change. District counsel reviews the data-processing terms of every AI vendor that touches student records. A managed AI vendor that controls the model and the data path forces a new review every time the vendor updates terms, switches sub-processors, or gets acquired. A self-hosted stack inside the district's VPC stays under one FERPA review forever.
3. Equity requires district-level control, not vendor-level. Districts serving multilingual learners, neurodiverse students, or students with IEPs need to set the agent's behavior — what languages it supports, what reading levels, what accommodations, what cultural framing. Managed AI vendors offer "settings"; self-hosted lets the district configure agents the way the district's curriculum and IEP teams designed them.
What Stays the Same, What Changes
Self-hosting the runtime doesn't mean rebuilding the district's AI tooling. The chat UI, the teacher dashboards, the parent-facing reports, the SIS integration (PowerSchool, Infinite Campus), the LMS integration (Canvas, Schoology, Google Classroom via LTI 1.3), the audit logs — all of that stays managed by ibl.ai. The compute, the model, and the student data move inside the district's VPC.
What disappears: the $900K–1M/year per-seat line item. What appears: a district-owned AI capability with a model-routing policy the IT director controls — Opus for the IEP team, Sonnet for general tutoring, Qwen 3 for the Spanish-language sections, Haiku for high-volume Q&A.
Run the Numbers for Your District
For workload sizing and ROI modeling — tutoring engagement, writing feedback volume, IEP drafting time savings — start with the AI Help Desk Cost Savings Calculator (the math generalizes to most high-volume school workloads).
For the deployment comparison side-by-side — including FERPA / COPPA posture and district-controlled deployment — see Self-Hosted AI vs ChatGPT Enterprise for K-12.
For multilingual student support specifically, Qwen 3 for Education: Multilingual AI Tutoring covers running the multilingual model self-hosted.
For the under-13 data-privacy requirements specifically, see COPPA-Compliant AI for Schools.
The per-seat-vs-usage math isn't specific to K-12 — it plays out the same way in every sector. The cross-industry view is in What Does AI Actually Cost in 2026?.
Why Family-Owned and New York Matters Here
For a school district, the AI vendor relationship is a multi-year commitment that touches student records, IEP documentation, and parent-facing communication. 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's network. The math works at a 2,000-student elementary district or a 200,000-student urban system.
Related: 20,000 Students in the AI Challenge — Who Owns Their Data?
Related: Why AI Agent Security in K-12 Requires a Different Playbook
Related: The Karpathy Lesson for K-12: Teach Comprehension, Not Just Usage
Related: K-12 AI Vendor Subscriptions vs Infrastructure You Own
Related: FERPA Governs Data, Not Which Model Reasons About It