GPT API Basics Cheat Sheet

OpenAI API fundamentals — chat completions, streaming, function calling, token management, error handling, and rate limits for GPT-4, GPT-4o, and GPT-3.5 models.

Last Updated: May 1, 2025

Chat Completion Basics

POST https://api.openai.com/v1/chat/completions

Single endpoint for all chat models

model: 'gpt-4o'

Best overall: fast + capable + vision + function calling

model: 'gpt-4-turbo'

Larger context (128K), slightly slower, high quality

messages: [{role:'system',content:'...'},{role:'user',content:'...'}]

Message array — system contextualizes, user asks

temperature: 0.7

Creativity control — 0=deterministic, 1=creative, 2=max

max_tokens: 1024

Response length limit — count both input+output against model limit

stream: true

Server-sent events — response arrives token-by-token

Function Calling

tools: [{type:'function',function:{name,description,parameters}}]

Define available functions

tool_choice: 'auto'

Model decides when to call a function (default)

tool_choice: {type:'function',function:{name:'get_weather'}}

Force a specific function call

finish_reason: 'tool_calls'

Model wants to call a function — extract arguments, execute, send back

role: 'tool', tool_call_id, content: result

Send function result back to model

Token Management

Item	Description
`tiktoken library`	Count tokens before sending — avoid exceeding model limits
`Pricing (per 1M tokens)`	GPT-4o: $5 input / $15 output. GPT-4: $30/$60. GPT-3.5: $0.50/$1.50.
`System prompt tokens`	Count against input limit — keep under 500 tokens for efficiency
`Truncation strategy`	Summarize old messages rather than dropping — preserve conversation context
`Token limits`	GPT-4o: 128K. GPT-4: 8K/32K/128K. GPT-3.5: 4K/16K.

Error Handling

Item	Description
`401 Unauthorized`	Invalid API key — check key, billing, organization
`429 Rate Limit`	Too many requests — implement exponential backoff with jitter
`500 Server Error`	OpenAI side — retry with backoff, max 3 attempts
`Context Length Exceeded`	Input too long — trim messages or use model with larger context
`Content Filter`	Request flagged by safety system — rephrase or use moderation endpoint first

Pro Tip: Always use streaming for user-facing applications — it cuts perceived latency from 5+ seconds to near-instant. The first token appears in ~200ms with streaming.