ChatGPT Cheat Sheet
9 groups · 38 topics · Updated
ChatGPT is OpenAI's AI assistant, and the same models can be called from code through the OpenAI API. This is an independent guide, checked against OpenAI's developer documentation in November 2026. Model names and prices change often, so confirm them on the official pages linked at the end.
Models
Which model to choose, and what it costs.
Current models
| Model | ID | Use it for |
|---|---|---|
| GPT-6 Astra | gpt-6-astra |
The most demanding work |
| GPT-6.1 Sol | gpt-6.1-sol |
Complex work at a lower cost |
| GPT-6.1 Terra | gpt-6.1-terra |
Fast work: images, writing, computer use |
| GPT-6 Luna | gpt-6-luna |
High volume, focused tasks |
- In the app you pick a model from a menu. In code you pass the ID.
- Start with Sol. Move up for difficulty, down for speed and cost.
Prices
US dollars per million tokens, for prompts up to 272K tokens.
| Model | Input | Cached input | Output |
|---|---|---|---|
| Astra | $10 | $1 | $50 |
| Sol | $2 | $0.10 | $10 |
| Terra | $1 | $0.05 | $5 |
| Luna | $0.10 | $0.01 | $0.50 |
- Output costs five times as much as input.
- These are API prices. The ChatGPT app is a separate subscription.
Limits
| Limit | Size |
|---|---|
| Context window | 1.05M tokens |
| Longest reply | 128K tokens |
- The context window is everything the model can see at once: instructions, history, tools and its own reply.
- A token is a chunk of text, often part of a word.
Other models
| For | ID |
|---|---|
| Making and editing images | gpt-image-2.6-sunburst |
| Live voice conversation | gpt-live-1 |
| Speech to text | gpt-transcribe |
App or API
| ChatGPT app | OpenAI API | |
|---|---|---|
| Used by | People, in a chat window | Programs, in code |
| Paid by | Subscription | Tokens used |
| Remembers the chat | Yes | Only what you send or link |
| Sign up at | chatgpt.com | platform.openai.com |
A ChatGPT subscription does not include API credit.
Prompting
How to ask so that the first answer is usable.
A good prompt
- Task. What you want done, in one sentence.
- Purpose. What the result is for.
- Audience. Who will read it, and what they already know.
- Material. The text, data or code to work on.
- Format. Length, layout and tone.
- Limits. What to leave out.
Write a 100-word product update for
customers who are not technical.
Explain the new export button below.
Friendly tone, no jargon, no emoji.Separate the material
Mark where pasted text starts and ends, so it is not read as an instruction.
Summarise the email between the
markers in three bullet points.
"""
...paste the email here...
"""- Triple quotes, tags or a heading all work. Use one style throughout.
Show an example
Rewrite each line as a short heading.
Input: we fixed the bug in login
Output: Login bug fixed
Input: you can now export to csv
Output:- One to three examples is enough.
- An example shows format and tone better than a description does.
Follow up
A chat is a conversation. Improve the answer instead of starting again.
| Say | To get |
|---|---|
| "Shorter. Keep the first point." | A tighter version |
| "As a table." | A different layout |
| "For a ten-year-old." | A simpler explanation |
| "What did you assume?" | Hidden guesses |
| "Give me three other options." | Alternatives |
| "What is the weakest part?" | A critique |
Fix a weak answer
| Problem | Try |
|---|---|
| Too long | State a length: "under 100 words" |
| Too generic | Add the audience and the purpose |
| Wrong format | Show the format with an example |
| Makes things up | Paste the source and say "use only this" |
| Out of date | Turn on search, or paste recent facts |
| Lost the thread | Start a new chat with a summary |
Check the answer
- It can be confidently wrong. Check facts, numbers, quotes and links.
- Ask for sources, then open them.
- Run any code it writes before you rely on it.
- For anything medical, legal or financial, treat it as a starting point, not advice.
The ChatGPT app
Features of the chat app, and how to use it safely.
Features
| Feature | Does |
|---|---|
| Custom instructions | Standing rules about you and how to answer, used in every chat |
| Memory | Keeps useful facts from earlier chats |
| Projects | Groups chats, files and instructions for one piece of work |
| GPTs | Saved assistants with their own instructions and files |
| Temporary chat | A chat that is kept out of your history and memory |
| Search | Looks things up on the web and cites sources |
| Deep research | Researches a topic at length and writes a report |
| Canvas | A side panel for editing a document or code together |
| Voice | Talk and listen instead of typing |
| Files and images | Upload a document, spreadsheet or picture to work on |
What you get depends on your plan and country.
Useful requests
| To | Ask |
|---|---|
| Summarise | "Summarise this in five bullet points." |
| Rewrite | "Rewrite this to be clearer and half as long." |
| Explain | "Explain this code line by line." |
| Compare | "Compare these two options in a table." |
| Extract | "List every date and amount in this text." |
| Plan | "Break this goal into steps for one week." |
| Review | "What is unclear or missing in this draft?" |
Custom instructions
About me: I am a web developer.
I use JavaScript and British spelling.
How to answer: be brief. Give code
first, then one line of explanation.
No apologies, no filler.- Find them in Settings, under Personalisation.
- Keep them short. They apply to every chat.
Shortcuts on the web
| Keys | Does |
|---|---|
| Shift + Enter | New line without sending |
| Ctrl or ⌘ + Shift + O | New chat |
| Ctrl or ⌘ + Shift + S | Show or hide the sidebar |
| Ctrl or ⌘ + Shift + C | Copy the last code block |
| Ctrl or ⌘ + / | Show all shortcuts |
Privacy
- Do not paste passwords, API keys or other people's personal data.
- Check Settings, Data controls, to see whether your chats are used to improve the models.
- Use a temporary chat for anything you do not want kept.
- At work, follow your company's rules on what may be shared.
Responses API
Calling the models from JavaScript.
Set up
npm install openai
export OPENAI_API_KEY="your-key"import OpenAI from "openai";
// reads OPENAI_API_KEY
const client = new OpenAI();- Keep the key on the server. Never ship it in browser code or commit it.
First request
const res = await client.responses.create({
model: "gpt-6.1-sol",
instructions: "You are a maths tutor.",
input: "What is pi?",
});
res.output_text;| Field | Meaning |
|---|---|
model |
Which model to use |
input |
The question, or a list of messages |
instructions |
The role and rules for this request |
Read the reply
res.output_text; // all the text, joined
res.status; // → "completed"
res.id; // to continue later
res.output; // every item, in order
res.usage; // tokens in and outoutputis a list of items: messages, reasoning and tool calls.output_textis the shortcut for the text alone.
A conversation
const next = await client.responses.create({
model: "gpt-6.1-sol",
previous_response_id: res.id,
input: "Explain it more simply.",
});Or send the history yourself:
input: [
{ role: "user", content: "I am Asha." },
{ role: "assistant", content: "Hi Asha!" },
{ role: "user", content: "Who am I?" },
],Images
input: [
{
role: "user",
content: [
{
type: "input_text",
text: "What is in this image?",
},
{
type: "input_image",
image_url: imageUrl,
},
],
},
],Two APIs
| Responses | Chat Completions | |
|---|---|---|
| Status | The main API: use for new code | Older, still supported |
| Call | client.responses.create |
client.chat.completions.create |
| Send | input |
messages |
| Read | res.output_text |
res.choices[0].message.content |
Shaping the reply
Streaming, reasoning effort, length and structured output.
Streaming
const stream =
await client.responses.create({
model: "gpt-6.1-sol",
input: "Tell me a story.",
stream: true,
});
const delta = "response.output_text.delta";
for await (const event of stream) {
if (event.type === delta) {
process.stdout.write(event.delta);
}
}| Event | Means |
|---|---|
response.created |
The reply has started |
response.output_text.delta |
A piece of text arrived |
response.completed |
The reply is finished |
error |
Something went wrong |
Reasoning effort
reasoning: { effort: "low" },| Level | Use it for |
|---|---|
none, minimal |
The fastest replies. Not every model accepts them |
low |
Quick, simple tasks |
medium |
The usual default |
high, xhigh |
Hard problems |
max |
When being right matters more than cost |
- Higher effort means more thinking, more tokens and a slower reply.
- Each model accepts only some levels. An unsupported one is a 400 error.
Reply length
max_output_tokens: 25000,if (res.status === "incomplete") {
res.incomplete_details.reason;
// → "max_output_tokens"
}- The limit covers reasoning as well as the visible reply. Set it too low and the reply can be empty.
JSON output
import { z } from "zod";
import { zodTextFormat }
from "openai/helpers/zod";
const Person = z.object({
name: z.string(),
email: z.string(),
});
const res = await client.responses.parse({
model: "gpt-6.1-sol",
input: "Asha, asha@example.com",
text: {
format: zodTextFormat(Person, "person"),
},
});
res.output_parsed.name;- The reply is made to match the schema.
- If the model refuses, there is no parsed result. Check before using it.
Tools
Letting the model call your code.
Define a function
const tools = [
{
type: "function",
name: "get_weather",
description: "Weather in a city now",
parameters: {
type: "object",
properties: {
city: { type: "string" },
},
required: ["city"],
additionalProperties: false,
},
strict: true,
},
];- The model never runs it. It asks, your code runs it, and you send back the result.
strict: truemakes the arguments match the schema exactly.
The tool loop
- Send the request with
tools. - Look in
res.outputfor items of typefunction_call. - Parse
arguments, which is a JSON string, and run your function. - Add the model's output items to
input, then afunction_call_outputfor each call. - Send again with the same
tools. Repeat until there are no more calls.
Send a result back
import { toResponseInputItems } from
"openai/lib/responses/ResponseInputItems";
input.push(
...toResponseInputItems(res.output),
);
for (const item of res.output) {
if (item.type !== "function_call") {
continue;
}
const args = JSON.parse(item.arguments);
input.push({
type: "function_call_output",
call_id: item.call_id,
output: "18°C, cloudy",
});
}call_idmust match the call being answered.outputmust be a string. UseJSON.stringifyfor an object.- Keep every output item, not only the messages. Dropping reasoning or tool items makes the next request fail.
Built-in tools
These run on OpenAI's servers. There is no loop to write.
| Tool | type |
|---|---|
| Web search | web_search |
| Search your uploaded files | file_search |
| Connect to an MCP server | mcp |
tools: [{ type: "web_search" }],Cost and speed
Paying less and waiting less.
Prompt caching
- It is automatic. There is nothing to turn on.
- It applies to prompts of 1,024 tokens or more.
- The start of the prompt must be identical to an earlier request.
- Put what stays the same first: instructions, tools, reference text. Put what changes last.
- Cached input costs about a tenth of the normal price, or less.
res.usage.input_tokens_details.cached_tokens;If that stays at 0, something near the start of the prompt changes every time, such as a date.
Usage
res.usage.input_tokens;
res.usage.output_tokens;
res.usage.total_tokens;Spend less
| Situation | Use |
|---|---|
| Same long prompt each time | Keep the start identical, for caching |
| The answer can wait | Batch API: half price |
| A simple task | Luna or Terra, or lower effort |
| Replies too long | Ask for a length in the prompt |
| History keeps growing | Start again from a summary |
Errors
What went wrong, and whether to try again.
Status codes
| Code | SDK class | Retry? |
|---|---|---|
| 400 | BadRequestError |
No |
| 401 | AuthenticationError |
No |
| 403 | PermissionDeniedError |
No |
| 404 | NotFoundError |
No |
| 409 | ConflictError |
Yes |
| 422 | UnprocessableEntityError |
No |
| 429 | RateLimitError |
Yes |
| 500 and up | InternalServerError |
Yes |
| None | APIConnectionError |
Yes |
Catch errors
const { RateLimitError, APIError } = OpenAI;
try {
await client.responses.create(params);
} catch (err) {
if (err instanceof RateLimitError) {
// wait, then try again
} else if (err instanceof APIError) {
console.error(err.status);
console.error(err.requestID);
} else {
throw err;
}
}- Check the most specific class first.
- Quote
requestIDwhen you ask for support.
Retries and timeouts
const client = new OpenAI({
maxRetries: 2, // the default
timeout: 600_000, // 10 min, in ms
});- The SDK already retries connection errors, 408, 409, 429 and 5xx.
- The timeout is in milliseconds.
Common causes
| Error | Usually means |
|---|---|
| 401 | The key is missing, mistyped or revoked |
| 404 | The model ID is wrong, or your account cannot use it |
| 429 | Too many requests, or the account is out of credit |
| 400 | A field is missing, misnamed or not allowed for this model |
More help
The official pages this sheet was checked against.
Official documentation
| Topic | Page |
|---|---|
| Models | Models |
| Prices | Pricing |
| Tools | Function calling |
| JSON output | Structured outputs |
| Reasoning | Reasoning |
| Caching | Prompt caching |
| JavaScript SDK | openai-node on GitHub |
| The app | OpenAI Help Center |