Model comparisons, the reasoning revolution, open vs closed weights, agents and safety risks
From OpenAI, Anthropic, Google, xAI, Meta and Mistral to DeepSeek, Qwen, Kimi, GLM, MiniMax, Doubao, Hunyuan and ERNIE, plus the coding-only Cursor Composer. One school usage model, applied to every vendor's published pricing, puts all 15 on a single table — followed by a four-question framework for choosing between them.
Read moreDeepSeek costs one sixty-seventh of GPT-6 Astra per input token and is no weaker in Chinese. So why does a school platform not simply plug into it? This roundup compares DeepSeek, Qwen, Kimi, GLM, MiniMax, Doubao, Hunyuan and ERNIE, explains the crucial difference between open weights and a cloud API, and sets out the three routes that are genuinely workable under the Personal Data (Privacy) Ordinance.
Read moreA chatbot only talks. An agent acts — it looks things up, reads documents, calls systems and runs several steps in sequence. This article explains tool calling and the MCP standard in terms a teacher can use, compares published tool-call charges, and offers a should-automate and should-not-automate list. The question is not what AI can do, but which step a person must confirm.
Read moreIn September 2026 the four leading US flagships range from $0.75 to $10 per 1M input tokens — a 13x spread. This head-to-head compares them on the six axes a school actually cares about (price, context, Chinese-language handling, content safety, platform availability and cost cliffs) and explains why the strongest model is usually not the one a school should be running.
Read moreAn AI drafted a parent notice citing a curriculum guideline number that was perfectly formatted and did not exist. That is not a glitch; it is a structural property of language models. Using the OWASP Top 10 and the NIST AI Risk Management Framework as scaffolding, this article maps each risk to a real Hong Kong school scenario and sets out the guardrails and audit arrangements a school AI platform must have.
Read moreMany schools assume that buying a machine to run an open model must be cheaper. Put real usage through September 2026 published pricing and the answer usually reverses — the annual cloud API bill is too small for hardware to beat. This article works through both sides honestly, lists the open models that are realistic on school hardware, and identifies the point at which data policy tips the balance towards self-hosting.
Read moreModels in 2024 blurted out an answer. Models in 2026 think first, and decide for themselves how long to think. This article explains chain-of-thought, reasoning effort and adaptive thinking in terms a teacher can use, compares the four vendors' effort dials, and sets out what has genuinely changed for marking, maths and science teaching — and what has not.
Read moreSpecs, pricing, release timelines and platform support status for 15 major model providers
Read more28 lessons of teacher AI literacy: from Transformers and the three training stages to RAG, quantisation and local deployment
Read more18 lessons of classroom prompting: planning, question setting, marking, differentiation and student prompt literacy
Read more