Choosing Among Four US LLM Providers: OpenAI, Anthropic, xAI and Google for Schools

An IT coordinator uploads the same 22-page school assessment policy four times and asks each US flagship for a summary. The dearest run costs HK$12.47, the cheapest HK$0.94 — more than a 13x spread. Three panel heads score the outputs blind and none can reliably say which came from GPT-6 Astra at $10 per 1M input tokens and which from Gemini 3.8 Flash at $0.75.

The question for a school is never "which is strongest". It is "which is good enough on the tasks we actually run, at a bill we can control" — and those two answers rarely name the same vendor.

Three words people merge by mistake

Context window sets how much text the model can hold at once. GPT-6 Astra runs to about 1.05M tokens, Claude Fable 5.1 about 1M, Gemini 3.8 Flash about 1.05M, Grok 4.6 500K — more than enough for almost any school document. A large window is not the same as understanding well; dumping a full term of materials in one go makes the model more likely to latch onto the wrong detail and can trigger the long-input cliffs below.

Parameter count reflects latent capability, not a daily lesson-planning need. Using a flagship for a Primary 5 Chinese fill-in-the-blanks sheet is hiring a barrister to photocopy — it works, but the cost does not match the task.

Reasoning effort is the dial all four converged on in 2026: OpenAI's reasoning.effort, Anthropic's effort, Google's thinking_level, xAI's reasoning_effort. More thinking usually means better output, billed at output rates. See the reasoning revolution for more.

What the September 2026 bill actually looks like

From published vendor pricing, USD per 1M tokens:

OpenAI GPT-6 Astra (released 2026-09-03) is $10 input, $50 output, $1.00 cache read, knowledge cutoff 30 April 2026. Above 272K tokens the entire request bills at 2x input and 1.5x output — not just the excess.

Anthropic Claude Fable 5.1 (2026-09-01) is $10 / $50, cache read $0.25 (lowest of the four), cutoff June 2026, adaptive thinking on by default at high effort. Mid tier Claude Sonnet 5 is $2 / $10, confirmed as permanent pricing on 10 August 2026.

Google Gemini 3.8 Flash (2026-09-02) introductory $0.75 / $3.75, cache read $0.075; the pricing page states this holds through 31 December 2026 and becomes $1.50 / $7.50 from 1 January 2027. It is the only flagship here with native text, image, audio and video input.

xAI Grok 4.6 (2026-08-12) is $2 / $6, cache read $0.50, cutoff 1 February 2026. At 200K tokens the entire request bills at $4 / $12.

The dearest and cheapest flagship input prices differ by 13.3x, yet on lesson planning, summarising and reformatting the capability gap is often much smaller.

Each vendor from a school angle

OpenAI and Azure

What makes OpenAI hard to replace is not "GPT is strongest" but that one vendor supplies chat, embedding and speech models. A school knowledge base needs vector retrieval; oral practice needs speech recognition and synthesis. Lose either and you no longer have a complete platform — which is why OpenAI is the Edor.ai default. Sponsoring bodies on Azure can route the same models through Azure OpenAI and manage data location under their own subscription. See the OpenAI model page.

Anthropic

Claude is steadier on Chinese writing feedback and long-document rubric checks. Fable 5.1 defaults to thinking before answering, which helps judgement-heavy tasks at the cost of higher latency and a $50 output price level with GPT-6 Astra. Anthropic supplies no embedding or speech models; drop OpenAI entirely and you must arrange those separately. Run Sonnet 5 ($2 / $10) daily and escalate to Fable 5.1 for hard problems. See the Anthropic model page.

Google

Gemini 3.8 Flash suits grunt work such as turning two years of scanned notices into a structured list, at an introductory price far below the rest. Edor.ai has no native Gemini — calls go through Poe, and Poe offers no embeddings or speech. Invisible in a demo, obvious in week three. Schools on Google Workspace for Education can run both. Budget three years at post-January 2027 prices. See the Google Gemini model page.

xAI

Grok 4.6 sits mid-priced at $2 / $6, 500K context is enough for school documents, and the xhigh reasoning level is in place. What a school must judge is content posture: tight integration with X platform content means test before student use and write the result into policy. Also Poe-only. See the xAI Grok model page.

For a school, "native provider" is not marketing — it decides whether one platform covers chat, knowledge base and oral practice without a second key to maintain.

Two traps behind the quotation

Long-input cliffs. OpenAI and xAI use the same mechanic: cross the threshold and the whole request reprices. One coordinator pasted a full term of materials (~310K tokens) into GPT-6 Astra for a summary and the bill doubled — only then reading the 272K clause. The right pattern is RAG over relevant passages; see the school knowledge base.

Cache spreads. If every call reloads the same curriculum outline, Claude Fable 5.1 cache read at $0.25 is a quarter of GPT-6 Astra's $1.00; Gemini's introductory $0.075 is lower still. On repetitive school workflows the cache column drives spend more than the headline rate — though on Edor.ai schools pay a fixed annual fee (~800 pupils, 100 staff) and token movement is absorbed by the platform.

Local hallucination. All four invent EDB circular numbers, curriculum codes and full school names with perfect confidence. Price does not fix it; answers must come from the school knowledge base with teacher review.

Prices move both ways. OpenAI cut GPT-5.6 Luna by 80% on 30 July 2026 and Sol by over 20% on 21 August 2026; Gemini 3.8 Flash has already announced a doubling for January 2027. A school wiring the API itself cannot lock a budget.

Routing without locking to one vendor

In our deployments roughly eight in ten requests — question generation, summarising, reformatting, first-pass marking, notice drafts — are fine on a mid tier. Only cross-document reasoning, multi-step maths and science explanation, and tasks cross-referencing several policies justify escalation.

Practical matching:

• One vendor for chat, knowledge base and oral practice → OpenAI or Azure OpenAI
• Steady Chinese writing feedback → Anthropic, Sonnet 5 daily, Fable 5.1 for hard problems
• Bulk documents or media grunt work, already on Google Workspace → Gemini, budget at 2027 prices
• Mid cost and capability with a completed content-safety assessment → Grok
• Data cannot leave the campus network → none of the four; read open versus closed weights and run local models on Ollama

All four replaced a flagship at least once in the first nine months of 2026. At procurement ask: if we switch, does the teacher interface change, must the knowledge base be rebuilt, do historical records survive? All three answers should be no. The full 15-vendor picture is in the 2026 all-model roundup; vocabulary starts at lesson one of the LLM classroom.

Common misreadings

• The claim: A flagship costs 13x more, so output quality should be visibly 13x better. What actually happens: On lesson planning and summarising, blind panel scoring often finds no difference. The gap shows up in long reasoning, hard judgement and the bill, not every worksheet.

• The claim: Gemini is cheapest, so moving the whole school there solves everything. What actually happens: Introductory pricing doubles from 1 January 2027; Poe access supplies no embeddings or speech, so the knowledge base and oral practice still need OpenAI or Azure.

• The claim: A million-token context means paste the whole file for best accuracy. What actually happens: Long inputs can trigger whole-request surcharges and the model may focus badly. RAG over relevant passages is more accurate and cheaper.

Three things to try on Monday

1. Pull three real artefacts from a panel (a Chinese composition, a curriculum extract, a draft parent notice), run one prompt across all four flagships, mask vendor names and have three panel heads score — a one-page record useful for AI funding reporting.
2. Ask IT to confirm: if a single request exceeds 272K tokens (OpenAI) or 200K (xAI), does the platform auto-chunk or RAG to avoid the long-input cliff?
3. Add one line to procurement: "Providers and models are switchable without changing the teacher interface, school knowledge base or ai_audit_logs history."