A token is a chunk of text that the AI model processes. Roughly:

• 1 token ≈ 4 characters in English
• 1 token ≈ ¾ of a word
• 100 tokens ≈ 75 words
• 1,000 tokens ≈ 750 words (about 1.5 pages)

When you send a request to an AI API, the model counts your input tokens (what you send) and generates output tokens (what it returns). You pay for both — but at different rates.

Input vs Output Tokens: Why the Price Difference?

Every AI API has two prices:

• Input price: Cost per 1M tokens you send to the model (your prompt + context)
• Output price: Cost per 1M tokens the model generates (its response)

Output tokens always cost more — typically 3-6x the input price. Here's why:

1. Compute intensity: Generating each output token requires running the full model forward pass. Input tokens can be processed in parallel (batched), but output tokens must be generated one at a time.
2. Memory requirements: The model must maintain attention over all previous tokens while generating each new one.
3. Latency: Output generation is the bottleneck — users wait for it, so providers charge more.

The Cost Formula

Example: 1,000 input tokens + 500 output tokens on GPT-4o mini ($0.15/$0.60 per 1M):

Input cost:  1,000 ÷ 1,000,000 × $0.15 = $0.00015
Output cost:   500 ÷ 1,000,000 × $0.60 = $0.00030
Total per request:                        $0.00045

At 1,000 requests/day × 30 days = $13.50/month

Pricing Across 59 Models (Per 1M Tokens)

Key observation: The cheapest input tokens (Gemini Flash Lite at $0.075) are 400x cheaper than the most expensive (GPT-5.5 Pro at $30.00). The output spread is even wider at 600x. Model choice is the single biggest lever for controlling costs.

How Tokens Add Up in Real Applications

Chatbot (Simple Q&A)

System prompt:     200 tokens (fixed instructions)
User message:      100 tokens (the question)
Model response:    300 tokens (the answer)
Total:             600 tokens per request

Cost on GPT-4o mini: $0.00027/request
Cost on GPT-5:       $0.00375/request (14x more)

RAG Pipeline (Search + Generate)

System prompt:      300 tokens
Retrieved context: 2,000 tokens (5 documents)
User question:      100 tokens
Model response:     500 tokens
Total:             2,900 tokens per request

Cost on GPT-4o mini: $0.00174/request
Cost on GPT-5:       $0.02415/request (14x more)

Coding Assistant

System prompt:      500 tokens (code instructions)
Code context:      3,000 tokens (file contents)
User instruction:   200 tokens
Model response:   1,500 tokens (code generation)
Total:            5,200 tokens per request

Cost on Claude Sonnet 4.6: $0.039/request
Cost on GPT-5.5:           $0.0725/request (1.9x more)

5 Ways to Reduce Your Token Costs

1. Shorter prompts: Remove unnecessary instructions, use concise system prompts. Every token in your prompt costs money.
2. Conversation pruning: Don't send 50 messages of history. Keep the last 5-10 and summarize the rest.
3. Output limits: Set max_tokens to what you actually need. Most chat responses don't need 4,096 tokens.
4. Model routing: Use cheap models for simple tasks, expensive ones for complex reasoning.
5. Prompt caching: OpenAI and Anthropic offer prompt caching — identical prefixes cost 50-90% less.

Calculate Your Costs

Don't guess — calculate. Enter your exact usage into our calculator to see what every model costs you per month.

Try it free: APIpulse Cost Calculator — estimate your monthly spend across 67 models and 10 providers in 30 seconds.