Comparing 15 LLM Providers in 2026: Prices, Context Windows, and the One Decision a School Actually Makes

The vice-principal puts two quotations side by side on the table. Both say AI usage is included and the annual fees differ threefold. She turns to the IT panel head running the integration: "Same pupils, same pile of homework — explain to me how the token cost for one year can be US$46 or US$7,410. I have to justify that number to the IMC."

Both figures are real, and the only difference between them is which tier was chosen. The decision a school makes is not which model to run but who absorbs the price volatility, and what a generational change will cost.

Price has decoupled from capability

Capability gaps on comparable tasks are narrowing while price gaps widen. That is not a benchmark conclusion; it falls out of dividing one published price list by another. Which makes "always use the strongest model" an expensive habit that does not reliably buy a matching difference in the classroom.

Where the money goes

The first driver is reasoning effort. All four US providers and most Chinese ones now expose a dial for how long the model thinks, and the thinking is billed at output rates, so the same question at low and at max can differ several-fold while producing answers of similar visible length. None of that appears on a quotation; it appears on the monthly statement. See the reasoning revolution.

The second is input volume. A context window is a ceiling, not a free container: a million-token window is still billed per token at the input rate, and several vendors surcharge very long inputs. Re-pasting a full set of curriculum documents on every call costs far more than retrieving the relevant passage.

Why the cheap tier is cheap

Architecture explains most of it. The 2026 dividing line is no longer total parameters but active parameters: a mixture-of-experts model may hold 552B parameters and activate only 8B to 16B at inference. That is how DeepSeek offers a million-token context at $0.15. The rest is competition and promotional windows, and at least three prices below are introductory rates with increase dates already in the terms.

Where the fifteen stand

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

Row seven needs a note. Cursor Composer 2.5 is not a candidate for a learning platform: it is a coding-specific model inside the Cursor editor with no inference API callable as a model endpoint, and the official Cloud Agents API and SDK drive coding agents rather than a chat model. See the Cursor Composer model page.

Seven rows read "not wired up", which states a fact rather than a verdict. Several of those models can be self-hosted from open weights through Ollama, while calling their official APIs directly is a cross-border transfer — all eight Chinese vendors serve from the mainland — requiring assessment under the Personal Data (Privacy) Ordinance. See the China model roundup.

The small print on open weights

There is a contract's worth of distance between "some models open" and open. Qwen's flagship Qwen3.8-Max is a closed cloud API and the open build of the same architecture carries 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. A school writing "self-hostable" into a tender must name the model and the licence, or discover at delivery that it received something else.

The annual bill

A unit token price means nothing to a school; an annual figure gets into a budget paper. The conversion rests on 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 estimating assumptions, not vendor figures.

Note the spread inside one vendor: stay with Anthropic and move from Fable 5.1 to Sonnet 5, and the bill falls by eighty per cent. Renegotiating with a new vendor takes months; changing tier is one admin setting.

Note also how unstable that last column is. Gemini's introductory rate doubles in January 2027, DeepSeek's 0.15 / 0.60 is off-peak and 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 APIs itself absorbs all three movements.

Picking today's strongest model is close to meaningless; picking an architecture that can swap any model out painlessly within six months is the actual decision.

Costs the tables do not show

Both tables cover model calls and nothing else. A knowledge base needs embeddings and oral practice needs speech recognition and synthesis, both separately priced. Some vendors bill tool calls per invocation. The line most often omitted is IT staff time: a self-hosted inference environment is not a one-off project but something someone checks every month.

The fourth is quotas. An integration without usage ceilings can lose a month's budget to one runaway batch job, so per-day and per-user request limits belong in the implementation requirements.

Where models fail in a Hong Kong school

Benchmark scores are of little use here, concentrating on coding and competition mathematics that correlate poorly with marking a Form 3 English composition and giving three specific pieces of feedback. The dependable assessment is a blind test on your own materials: three real school artefacts, names masked, scored by three panel heads. On the same logic, discount any "X times faster" claim that omits test method and hardware.

Local detail is the second weakness. Curriculum codes, circular numbers, full school names and government department structures are the thinnest material in any training corpus, and we have seen models produce perfectly formatted curriculum codes that do not exist — the hardest error to catch, because it looks professional. Retrieval over the school's own documents plus teacher review fixes it; changing vendor does not. See the school knowledge base.

Third, the September 2026 flagships have knowledge cutoffs scattered between January and June 2026, so a question about "this school year" may be answered, with equal confidence, about the last one.

Four sentences for the contract

What protects a school is not a good tier judgement but four sentences in the agreement. One: model upgrades are included in the annual fee. Two: providers are switchable, and switching requires no knowledge-base rebuild, no retraining and no loss of historical records. Three: all data exports in full at any time as JSON, CSV or original files, free and without conditions. Four: where data cannot leave the campus network, the platform supports on-premises deployment with the same feature set as the cloud.

For that fourth sentence the realistic candidates are self-hosted open-weight models — Qwen3.8-27B, Muse Glimmer 30B and the Mistral family. Be clear what self-hosting buys: it answers where the data goes, not whether the answer is right. See on-premises deployment and 12 questions before buying a school AI platform.

Common misreadings

• The claim: expensive models are more accurate on everything. What actually happens: on lesson planning, question generation, summarising and first-pass marking, mid tier and flagship take careful reading to distinguish, and the share of requests genuinely needing a flagship may be under a fifth.
• The claim: open weights mean safe. What actually happens: open weights answer where the data goes, not whether the answer is correct. Self-hosted models hallucinate too, and arrive without a cloud vendor's content-safety classifiers, so gating and monitoring fall to the school.
• The claim: a long enough context window removes the need for retrieval. What actually happens: long inputs are billed per token, with surcharges at some vendors. Re-pasting whole document sets is worse on both cost and latency.

Three things to try on Monday

1. Pull last term's usage records, count the actual requests, convert them with the assumptions above into annual costs at three tiers, and put the three figures side by side on one page.
2. Run a blind test with three real school artefacts through one flagship and one mid tier, names masked, scored by three panel heads. It fits in a free period and usually changes the budget allocation.
3. Ask your current or prospective vendor for written confirmation on three points: whether model upgrades are included, whether token overages exist, and whether data export is free. The reply is procurement evidence.