Anthropic Claude (Fable 5.1 / Opus 5 / Sonnet 5): the steadiest long-form marking model, and its two hard gaps

Of the four large American providers, Claude handles long documents and writing feedback most reliably and can be switched to inside Edor.ai — yet the detail procurement papers most often omit is that it has no embedding and no speech models, so it can be your main chat model without running a school platform on its own.

Anthropic was founded in 2021 by researchers who had left OpenAI, and it has one product line: Claude. For a school that must answer to an incorporated management committee, the selling point is safety as an auditable commitment — published system cards, published retirement dates, even the system prompt on claude.ai is public. In the classroom the recurring comment is that Claude is willing to write at length and willing to give reasons — but that does not replace OpenAI's role in the knowledge base and oral practice.

Safety as an auditable commitment

What distinguishes Anthropic at procurement time is not scale but documentation you can cite:

• Retirement dates and knowledge cutoffs are published for every model, giving an IT coordinator something official for a risk assessment.
• Safety evaluations and system cards are released publicly, which is easier to defend to a school council than a verbal assurance.
• Adaptive thinking is always on for Fable 5.1: the model decides how long to think with no manual switch — a different posture from GPT-6 Astra's adjustable reasoning.effort.

In subject work Claude's strength is long-form steadiness. Marking a Secondary 3 English composition, it tends to explain each deduction; summarising a forty-page school assessment policy, it is less prone to blending requirements from different chapters. Claude 3.7 Sonnet was the first mainstream model to expose extended thinking as a switch, and every vendor's reasoning mode today descends from it.

One bottom line should be stated plainly. Anthropic makes language models only. There is no embedding model and no speech recognition or synthesis. A complete school platform needs all three kinds, so Claude is a chat model you can run as your main choice, not a provider that can stand alone.

Model line-up at a glance

The September 2026 line-up reads from the hardest work downwards:

All three flagship tiers (Fable, Opus, Sonnet) offer a 1M-token context window, so length alone never forces an upgrade to the most expensive option. Fable 5.1's cache read rate of $0.25 is the lowest in the family, which matters when the same school document is queried repeatedly across a lesson.

API list price and the school budget

Anthropic charges developers by token with input and output priced separately. Fable 5.1's $10 / $50 matches GPT-6 Astra — a school wiring up the API itself faces the same budgeting difficulty.

Edor.ai is a fixed per-school annual subscription with AI usage included, no token charges and no per-seat pricing. The platform tiers by task difficulty: everyday teacher tools and student conversations run on Sonnet 5 or Haiku 4.5, and only long marking or cross-document work escalates. That controls cost more effectively than sending everything to the flagship or switching vendors entirely.

When a forty-page school policy is queried repeatedly, the gap between a $0.25 cache read and a $10 ordinary input is where Claude actually saves money in bulk work — not because the flagship itself is cheap.

Inside Edor.ai and the two gaps

Anthropic is a native provider. Once an administrator adds one school Anthropic API key, teacher tools, the AI tutor and student conversations can run on Claude, and it can be set per module — for example Chinese and English on Claude with everything else on OpenAI.

Gap one: no embedding model. A school knowledge base chunks documents, turns each chunk into a vector, and retrieves relevant passages by query. The embedding model does that conversion. With only an Anthropic key and no OpenAI or Azure key, the knowledge-base module cannot run.

Gap two: no speech models. Cantonese reading assessment, English oral practice and model pronunciation audio are all outside what Claude can do. For primary Chinese and English oral work that is a hard constraint.

Everything else matches the other providers. The key belongs to the school and can be rotated at any time, every call passes through four-layer gating and ai_audit_logs, and data can be exported in full for free. We suggest stating both gaps plainly in procurement papers and in the technical annex of an EDB funding application, so a missing module is not discovered at acceptance testing.

Where subject teachers should draw the line

Worth it for: long composition marking that quotes the pupil's own sentence before explaining the problem; consolidating school policy across documents; drafting letters and circulars where tone matters; summarising meeting notes; rewriting worksheets for differentiated groups. Sonnet 5 is enough without touching Fable 5.1.

Students: only inside a gated, monitored environment. In tutor mode Claude's willingness to explain is an advantage, but it can also say too much at once — teachers should put a length limit in the prompt as well.

Not worth it for: building the vector index behind a school knowledge base; Cantonese or English oral practice; expecting the model to know Hong Kong curriculum clause numbers or assessment detail without teacher review — hallucination remains; handling war history, violence in literature or adolescent health education without stating the teaching purpose in the prompt — legitimate material can be declined.

A Secondary 3 English teacher told a class teacher after marking: "Claude explained the connective problem in paragraph three clearly this time, but I still have to check the DSE descriptor myself." That is normal use, not a failure.

Copy this into an Edor.ai teacher tool to test:

You are a Hong Kong secondary Chinese-language teacher. Below is a Secondary 3 student's narrative composition. Score Content and Ideas, Structure and Organisation, and Language and Rhetoric out of 6 each.
Output: for each criterion quote one or two sentences from the student as evidence, then write two specific comments; finish with three practice priorities, each paired with one action the student can take immediately.
Constraints: write in Hong Kong Chinese conventions; do not rewrite any paragraph on the student's behalf; keep the total comment length under 400 characters; if a cultural allusion or quotation source is uncertain, mark it "teacher to verify" rather than inferring it.
[paste composition]

The trade-off for schools

• If long-form marking and cross-document consolidation are your main work, make Claude your chat model but keep an OpenAI or Azure key behind the knowledge base and oral practice.
• If you plan a full provider switch before EDB funding acceptance testing, write the embedding and speech replacement into the technical annex first — do not add an Anthropic key alone and switch everything else off.
• If Claude declines legitimate teaching material, put the year group, subject and teaching purpose in the prompt; if it still declines, switch that module to another model rather than assuming "safe" means "answers every classroom question."