The 2026 All-Model Roundup: 15 Providers and One Selection Framework for Schools

One school, one usage pattern, and an annual model bill that could be US$7,410 or US$46 — a spread of more than 160 times. Neither number is a sales line: both come from applying the same usage estimate to the token prices each vendor published in September 2026. What sits between them is 15 providers, several dozen models, and a series of clauses that never appear on a quotation.

This is the full comparison we redo each quarter. It is not meant to crown a winner. It is meant to give panel heads and IT coordinators a checkable baseline when a vendor is sitting across the table.

The three variables that drive cost

First, context window — the model's working memory. A million tokens is now the mainstream, enough to read a complete set of school curriculum documents in one pass. But a large memory is not the same as using it well, and several vendors surcharge very long inputs.

Second, parameter count — the number of brain cells. The 2026 dividing line is no longer total parameters but active parameters: a mixture-of-experts model may hold 552B parameters yet activate only 8B to 16B at inference. That is how DeepSeek can offer a million-token context at $0.15 — it wakes only a small fraction of the brain cells per call.

Third, reasoning effort. All four US providers and most Chinese ones now expose a dial for how long the model thinks, and the thinking itself is billed at output rates. The same question at low and at max can differ several-fold on the bill. See the reasoning revolution.

Head-to-head: all 15 providers

Prices in USD per 1M tokens (input / output), from each vendor's official documentation as of September 2026.

Three points deserve emphasis:

• Cursor Composer is not an option for a school platform. It is a coding-specific model built into the Cursor editor with no inference API that can be called as a model endpoint — the official Cloud Agents API and SDK drive coding agents, not a chat model. Cursor is a perfectly reasonable teaching tool in an IT lesson, but a learning platform's model choice does not live there. See the Cursor Composer model page.
• "Some models open" is not the same as open. Qwen's flagship Qwen3.8-Max is a closed cloud API, and the open build of the same architecture uses a bespoke licence without image input. The GLM-5.3 flagship weights were still unpublished when this article was updated; what is public is GLM-5.2 and GLM-5.3-Flash. When a school requires self-hostability, it must name the model, not the vendor.
• "Not wired up" is a statement of fact, not a verdict. We do not natively integrate DeepSeek, Kimi, MiniMax, GLM, Doubao, Hunyuan or ERNIE; several of them can be self-hosted from open weights through Ollama.

Turning token prices into numbers a school recognises

Token prices mean nothing to a school; an annual bill does. The conversion below uses clearly stated assumptions: 800 students and 100 staff; 600 AI requests per school day; 3,000 input and 700 output tokens per request on average; 190 school days per year. These are our estimating assumptions, not vendor figures — adjust them to your own usage intensity.

Three readings come out of that table. First, choosing the tier matters more than choosing the vendor: staying with Anthropic but moving from Fable 5.1 to Sonnet 5 cuts the bill by eighty per cent. Second, the cheap tiers are usually sufficient for everyday school work, and the requests that genuinely need a flagship may be under a fifth of the total. Third, every one of these numbers moves: Gemini's introductory rate doubles in January 2027, DeepSeek's off-peak rate doubles at peak hours, and GLM-5.3-Flash's $0.075 / $0.25 is promotional against a $0.15 / $0.50 list price. A school wiring up the API itself absorbs all of that volatility.

Limits and blind spots the table does not show

A price list is not a total cost. A school must also budget for embeddings (the knowledge base), speech recognition and synthesis (oral practice), tool-call charges that some vendors bill per invocation, and the item most often forgotten — IT staff time. An integration without quota controls can burn a month's budget on one runaway batch job.

Benchmark scores are close to useless for a school. The widely cited evaluations target coding and competition mathematics, which correlate poorly with "mark a Form 3 English composition and give three specific pieces of feedback". The only reliable assessment is a blind test on your own materials. By the same logic, treat any claim that a model is "X times faster" with scepticism unless the test method and hardware are stated.

Every model hallucinates, and local Hong Kong detail is where it is worst. Curriculum codes, circular numbers, full school names, government department structures — these are the thinnest parts of any training corpus. We have seen models produce perfectly formatted curriculum codes that do not exist. Changing vendor does not fix this; RAG over the school's own documents plus teacher review does. See the school knowledge base.

Knowledge cutoffs are an invisible trap. The September 2026 flagships have cutoffs scattered between January and June 2026. Ask about "this school year" and the model may confidently answer about last year.

Cross-border transfer is a contractual question, not a technical one. The eight Chinese vendors in the table all serve their official APIs from the mainland, so calling them directly constitutes a cross-border data transfer requiring assessment under the Personal Data (Privacy) Ordinance. See the China model roundup.

Actionable guidance: a four-question framework

Rather than comparing vendor by vendor, answer four questions in order. Most schools have their answer by question two.

Question one: can this data leave the campus network? No → only open-weight models self-hosted on site are in scope (Qwen3.8-27B, Muse Glimmer 30B and the Mistral family are the realistic candidates), and the other eleven are irrelevant. Yes → go to question two.

Question two: do we need embeddings and speech? Yes, because you want a school knowledge base or oral practice → the primary provider has to be OpenAI or Azure OpenAI, since neither Anthropic nor Poe supplies those model types. No → go to question three.

Question three: which tier does this task belong to?

• High-frequency, low-difficulty work such as classification, tagging, reformatting and notice drafts → the cheapest tier, around Claude Haiku 4.5 or Gemini 3.8 Flash.
• The daily workhorse — lesson planning, question generation, first-pass marking, student conversations → a mid tier such as Claude Sonnet 5.
• Genuinely hard work — cross-document reasoning, multi-step maths and science explanation, cross-referencing policies → a flagship, and only here.

Question four: can this decision be reversed? If switching vendor would mean rebuilding the knowledge base, retraining teachers or losing historical records, the answer fails. A passing answer looks like this: switch in the admin panel, export everything in one click, model upgrades included in the annual fee.

A checklist you can run this week

1. Count how many times each AI feature was actually used last term and convert it, using the assumptions above, into annual costs at three tiers;
2. Blind-test three real school artefacts with vendor names masked and three panel heads scoring;
3. Ask your current or prospective vendor to confirm in writing whether model upgrades are included, whether token overages exist, and whether data export is free;
4. Write switchable providers and one-click full export into the procurement terms;
5. Where sensitive data is involved, assess on-premises deployment in parallel.

For the full set of pre-purchase questions, see 12 questions before buying a school AI platform.

Takeaway

The one thing a school should take from the 2026 model market is this: capability gaps are narrowing, price gaps are widening, and the pace of replacement has not slowed. In that environment, picking today's strongest model is close to meaningless. Picking a platform architecture that lets you swap any model out painlessly within six months is the actual procurement decision.

For the full provider reference pages and model timeline, see the model database.