Education and training
Education and training use three kinds of models together: a mainstream model writes questions and training content from the course material, a model from another vendor answers each question independently to check the key and assists grading against rubrics, and a course assistant answers learners from the course content. Teachers approve questions before they enter the bank and grades before they are released. For 10,000 questions, 20,000 written answers and 30,000 learner questions a month, at live prices in each model’s cheapest group, model costs are about ¥346 a month.
Prices updated Nov 13, 2026, 6:13 PM
Monthly cost estimate
Assumptions: 10,000 questions a month written 5 per request, so 2,000 requests at about 1,500 input tokens (material and requirements) and 1,500 output tokens; each of the 10,000 questions checked with about 400 input and 200 output tokens; 20,000 written answers graded at about 1,500 input tokens (rubric, question and answer) and 300 output tokens; 30,000 learner questions with about 3,000 tokens of material and a 400-token answer.
¥346.44per month
- Question writing¥42
- Answer checking¥24
- Written-answer grading¥99
- Course assistant¥181.44
The business problem
Teachers and trainers spend much of their time on repetitive work: several sets of questions at different levels for every topic, hundreds of short answers to grade one by one, and learner questions that the slides already answer. Corporate training also means turning policies and product manuals into lessons and quizzes.
An LLM writes questions quickly, but its answer key can be wrong and questions can stray beyond the material. Letting a model grade directly gives different scores for the same answer on different runs, with no clear reason for lost points.
What works is a model for first drafts and first passes, with teachers deciding: questions come only from the given material and a model from another vendor answers them independently to check the key; grading follows a written rubric item by item and quotes the learner; the course assistant answers only from course content. Nothing enters the question bank or reaches learners as a grade without a teacher.
Recommended models
Recommended models are picked from the live catalog; prices are each model's cheapest group, updated Nov 13, 2026, 6:13 PM. Click a model to see every group's price.
Questions and training content (mainstream model)
deepseek-v4.1-flashDeepSeek
input ¥2.8 / output ¥11.2 per 1M tokensChina Models group
Reads the course or training material and writes stems, options, answer keys and explanations with a difficulty level; it also turns policies and product manuals into handouts. The volume is modest but every question has to be right, so it uses a mainstream model; output is longer than input, so the output price weighs more here.
Answer checking and grading assistance (another vendor’s model)
gpt-5.4-miniOpenAI
input ¥1.5 / output ¥9 per 1M tokensStandard group
Answers each question without seeing the key, and questions where it disagrees with the writer go to a teacher first; for written answers it scores each rubric item and quotes the learner. A model from another vendor makes it less likely that one model’s mistake goes unnoticed. For photos of handwritten work, switch to an alternative that accepts images.
Course assistant (mainstream model)
qwen3.5-plusQwen
input ¥1.12 / output ¥6.72 per 1M tokensChina Models group
For each learner question, relevant passages from slides, handouts and answer explanations are retrieved first, and the model explains only from them and names the chapter. This has the most requests and is usually the largest cost, so the input price matters most; retrieval works as in the knowledge-base solution.
Reference architecture
- 1
Prepare course material
Turn textbooks, slides and handouts into text organized by chapter and topic, labeling each passage; for corporate training, do the same with policies, product manuals and procedures.
- 2
Generate questions by topic
The mainstream model writes questions for a given topic, type and difficulty using only the given material, returning stem, options, answer, explanation and difficulty as JSON; code checks the format and that every option is present.
- 3
Check answers independently
A model from another vendor answers without seeing the key. Matching answers go to the teacher’s review queue; mismatches are marked as disputed and reviewed first. Only teacher-approved questions enter the bank.
- 4
Grade with rubrics
Every written question gets a rubric listing each point and its marks. The model scores each item and quotes the learner; code caps each item at its marks and adds them up as a suggested grade for the teacher to confirm.
- 5
Course assistant
Index the course material as in the knowledge-base solution, retrieve the relevant passages for each question, and have the model explain only from them and name the chapter; for homework and exam questions it can be set to explain the approach without giving the answer.
- 6
Teacher spot checks
Teachers regularly review questions, grades and assistant conversations, and feed what they find back into the question requirements, rubrics and course material.
Cost estimate
Assumptions: 10,000 questions a month written 5 per request, so 2,000 requests at about 1,500 input tokens (material and requirements) and 1,500 output tokens; each of the 10,000 questions checked with about 400 input and 200 output tokens; 20,000 written answers graded at about 1,500 input tokens (rubric, question and answer) and 300 output tokens; 30,000 learner questions with about 3,000 tokens of material and a 400-token answer.
| Item | Model | Usage assumed | Cost |
|---|---|---|---|
| Question writing | deepseek-v4.1-flash | 2,000 × (1,500 input + 1,500 output tokens) | ¥42 |
| Answer checking | gpt-5.4-mini | 10,000 × (400 input + 200 output tokens) | ¥24 |
| Written-answer grading | gpt-5.4-mini | 20,000 × (1,500 input + 300 output tokens) | ¥99 |
| Course assistant | qwen3.5-plus | 30,000 × (3,000 input + 400 output tokens) | ¥181.44 |
| Total (per month) | ¥346.44 | ||
Calculated from live prices in each model's cheapest group, updated Nov 13, 2026, 6:13 PM.
Model usage only: teacher review, the question-bank system and retrieval are not included (embedding costs very little; see the knowledge-base solution). Photos of handwritten work are billed as input tokens that depend on image size; the usage field in each response is what you are billed on. Twice the learners or answers means roughly twice the matching lines.
Sample code
The mainstream model writes questions from the given material only, a model from another vendor answers them independently and disagreements go to a teacher first; for grading, each rubric item is capped at its marks in code and summed as a suggested grade for the teacher.
import json
import os
from openai import OpenAI
client = OpenAI(base_url="https://www.tutujin.com/v1", api_key=os.environ["TUTUJIN_API_KEY"])
def ask_json(model, system, user):
resp = client.chat.completions.create(
model=model,
messages=[{"role": "system", "content": system}, {"role": "user", "content": user}],
temperature=0,
)
raw = resp.choices[0].message.content
return json.loads(raw[raw.find("{"): raw.rfind("}") + 1])
def write_questions(material, topic, n=5):
return ask_json(
"deepseek-v4.1-flash",
"You write multiple-choice questions for a course. Use only facts stated in the material. "
'Reply with JSON only: {"questions": [{"stem": "", "options": {"A": "", "B": "", "C": "", "D": ""}, '
'"answer": "A", "explanation": "", "difficulty": "easy|medium|hard"}]}',
"Topic: " + topic + "\nWrite " + str(n) + " questions.\n\nMaterial:\n" + material,
)["questions"]
def solve(q):
"""A model from another vendor answers without seeing the key."""
options = "\n".join(k + ". " + v for k, v in q["options"].items())
return ask_json("gpt-5.4-mini", 'Answer the question. Reply with JSON only: {"answer": "A"}', q["stem"] + "\n" + options)["answer"]
def suggest_score(question, rubric, answer):
"""Rubric-based suggestion for a written answer; the teacher confirms or changes it."""
result = ask_json(
"gpt-5.4-mini",
"Score the student answer against each rubric item and quote the words that earn the points. "
'Reply with JSON only: {"items": [{"id": "R1", "points": 0, "evidence": "", "comment": ""}]}',
"Question:\n" + question + "\n\nRubric:\n" + json.dumps(rubric, ensure_ascii=False) + "\n\nAnswer:\n" + answer,
)
cap = {r["id"]: r["max"] for r in rubric}
total = sum(min(max(float(i.get("points", 0)), 0), cap.get(i.get("id"), 0)) for i in result["items"])
return {"suggested": total, "out_of": sum(cap.values()), "items": result["items"], "status": "teacher_to_confirm"}
with open("chapter-3.txt", encoding="utf-8") as f:
questions = write_questions(f.read(), "Reading a cash flow statement")
for q in questions:
# every question is reviewed by a teacher; disagreements are reviewed first
q["status"] = "teacher_review" if solve(q) == q["answer"] else "answer_disputed"
print(json.dumps(questions, ensure_ascii=False, indent=2))
rubric = [
{"id": "R1", "max": 2, "text": "Names operating, investing and financing activities"},
{"id": "R2", "max": 2, "text": "Explains why profit and operating cash flow can differ"},
]
print(suggest_score("Why can a profitable company run out of cash?", rubric, "Customers pay late, so..."))Try this with one key
Sign up, create an API key in the console, and put your base_url and key into the code above. For company purchasing or invoices, add our WeChat support JsnonoChat.

Scan the code or search the WeChat ID to add us
FAQ
What do 10,000 questions and 20,000 graded answers a month cost?
With 10,000 questions written and checked, 20,000 written answers graded and 30,000 learner questions a month, the cheapest groups come to about ¥346 a month. The course assistant is usually the largest line.
Are the answer keys of generated questions always right?
No. A model can miscalculate, write two correct options or draw on knowledge outside the material. That is why the sample has a model from another vendor answer independently and sends disagreements to a teacher first; every question still needs a teacher’s approval before it enters the bank.
Can model grades go straight to learners?
Not recommended. The model gives a suggested grade with evidence for each rubric item, and the teacher confirms or changes it before release. The more specific the rubric, the closer the suggestion gets to the teacher’s grade; before going live, compare on a batch of answers teachers have already graded.
Will the course assistant just hand out homework answers?
You can say in the system prompt that for homework and exam questions it explains the approach and the related topics without giving the final answer, answers only from course material, and says it does not know, pointing learners to the teacher, when the material does not cover a question.
Does this work for internal corporate training?
Yes. Treat policies, product manuals and procedures as the course material and use the same flow to produce handouts and quizzes and to run a Q&A assistant. For sensitive internal material, remove what is not needed first and assess the setup against your company’s data rules.