OpenAI (GPT-6 Astra / GPT-5.6 family): the model schools meet first, and its role in a school platform

Short answer: OpenAI is the provider Hong Kong teachers meet first and the one Edor.ai uses by default. Its value is not that it is the newest or strongest, but that one vendor supplies chat, embedding and speech models — enough to run a complete school platform.

What this is: one vendor, three kinds of model a school needs

Most people know OpenAI through ChatGPT. For a school platform, though, what makes OpenAI hard to replace is that it supplies three model types at once:

• Language models (the GPT family): writing lesson plans, setting questions, marking, talking with students.
• Embedding models: turning school policies, curriculum outlines and teaching materials into vectors after chunking, so answers can cite the school's own documents. This is the foundation of a school knowledge base (RAG).
• Speech models: speech recognition (turning a student's reading aloud into text) and speech synthesis (modelling pronunciation) for Cantonese, Putonghua and English oral practice.

Anthropic and Poe currently offer only the first. That is why we recommend keeping an OpenAI or Azure OpenAI key even when a school wants Claude as its main model.

The current line-up: four lanes in a sentence each

The September 2026 line-up is best read as "one expensive tier and three affordable ones":

• GPT-6 Astra (3 September 2026): the current flagship, with a 1.05M-token context window, 128K output, and reasoning.effort adjustable from low to max. Pricing is $10 per 1M input tokens and $50 per 1M output tokens, with cached input at $1. Watch one cost trap: once a request exceeds 272K input tokens, the entire request is billed at double the input rate.
• GPT-5.6 Sol / Terra / Luna (9 July 2026): three co-existing tiers from the previous generation. Sol is the flagship, Terra is the balanced everyday choice (and what ChatGPT Free and Go users get), and Luna is the fastest and cheapest, cut by a further 80% on 30 July 2026.
• GPT-4.1 / mini / nano: the previous workhorse with a 1M-token context window, still served in the API at far lower cost than the flagship.
• The o-series (o1, o3, o4-mini): the first "think before answering" reasoning models. Adaptive thinking has now superseded them, but conceptually they started the whole reasoning revolution.

The full version and price comparison is in the automatically generated series table further down this page.

Strengths: the four that matter most to a school

1. The most complete ecosystem. Whether a teacher is looking for prompt templates or an IT coordinator is looking for integration documentation, OpenAI has the most material in both English and Chinese, which keeps training costs down.
2. Mature speech support. Cantonese recognition holds up in real classroom conditions — primary pupils reading aloud, unclear articulation, background noise — which is the main reason we chose it behind oral practice.
3. Deployable through Azure. Schools or sponsoring bodies already on Azure can reach the same models through Azure OpenAI while managing data residency and compliance documents in their own subscription.
4. Clear tiering. Expensive and cheap options sit inside one vendor, so a platform can send hard tasks to the flagship and routine work to a cheap tier. For cost control that is far more effective than switching vendors.

Limits: what to say plainly to a principal

• Prices move a lot. GPT-6 Astra's $10 / $50 is about 2.5 times GPT-5.6 Sol's promotional rate. A school wiring up the API itself will struggle to budget, which is why we charge a fixed annual fee and absorb model price movements.
• Long context is priced separately. Above 272K tokens the whole request doubles, so pasting a term's worth of material in one go is poor value. The right approach is RAG, retrieving only the relevant passages.
• It does not know your school. No general model knows your teaching schedule, marking conventions or circular formats. That gap has to be closed with a knowledge base and prompt templates.
• Hallucination has not gone away. Models state wrong facts in a confident tone, especially on local details (Hong Kong curriculum specifics, school names, regulatory wording). Every outward-facing document and marking result needs teacher review.
• The free tier is not for students. Consumer ChatGPT has age terms, no teacher monitoring, and no way for the school to review conversations. Student use needs a gated platform.

Where each tier fits

Teachers: planning, tiered worksheets, rubrics, draft letters to parents, tidying meeting notes. A Terra or GPT-4.1 mini class model is enough for these; the flagship is unnecessary.

Students: only inside a gated, monitored environment. Our tutor mode restricts it to hints rather than answers and leaves the hint-level ceiling to the teacher.

IT coordinators: if the school wires up the API directly, set usage caps and monitoring first. Without quotas, one runaway batch job can burn a month's budget.

Prompts worth testing

Copy these into a Edor.ai teacher tool or any GPT interface. Note the third line, the constraints — that is what makes the output usable:

You are a Hong Kong primary General Studies teacher. Design a 30-minute inquiry activity on "Hong Kong's water resources" for Primary 5.
Output: three activity objectives mapped to learning outcomes, five guiding questions from simple to complex, a student worksheet in table form, and teacher observation points.
Constraints: write in Hong Kong English conventions; do not give away answers; if any figure is uncertain, write "teacher to verify" rather than inventing it.

Below is a Secondary 3 student's English composition. Score Content, Language and Organisation out of 5 each, write two specific comments per criterion, then list three things this student should practise next.
Do not rewrite the composition — give feedback only. Quote the student's own sentences when referring to specific issues.
[paste composition]

Availability inside Edor.ai

OpenAI is a native provider and the platform default. Once an administrator adds one school API key, teacher tools, the AI tutor, student conversations, knowledge-base retrieval and oral practice all work, with no per-teacher accounts. A school can switch to Azure OpenAI instead and manage data residency in its own Azure subscription.

Two practical points. First, the key belongs to the school and can be rotated or revoked at any time. Second, the model choice is adjustable — if a school wants certain subjects to run on Claude, that can be switched per module in the admin panel.

Alternatives

• For steadier Chinese writing feedback, see Anthropic Claude.
• For the lowest cost on high volumes of documents, especially if you already use Google Workspace, see Google Gemini.
• To keep data entirely inside the campus network, see Meta open weights or Alibaba Qwen, and read the open versus closed cost and privacy trade-off.
• To understand what embedding, RAG and hallucination actually mean, start at lesson 1 of the LLM Classroom.