If you're new to AI APIs, the pricing model can be confusing. Unlike traditional SaaS (flat monthly fee) or cloud computing (pay per compute hour), AI models charge per token. Here's everything you need to know.
What Is a Token?
A token is a chunk of text that the AI model processes. It's not exactly a word — it's more like a syllable or a common character sequence. In English:
- 1 token ≈ 4 characters or about 3/4 of a word
- "Hello, world!" = 4 tokens
- "The quick brown fox" = 4 tokens
- A typical email (200 words) ≈ 270 tokens
- A full page of text (500 words) ≈ 675 tokens
Why Do Output Tokens Cost More?
Most providers charge 2-5x more for output tokens than input tokens. The reason is computational: generating new text (output) requires running the model one token at a time, while processing input text can be done in parallel. More computation = higher cost.
| Model | Input $/M tokens | Output $/M tokens | Output multiplier |
|---|---|---|---|
| GPT-4o | $2.50 | $10.00 | 4x |
| Claude Sonnet 4 | $3.00 | $15.00 | 5x |
| Gemini 2.5 Pro | $1.25 | $10.00 | 8x |
| GPT-4o mini | $0.15 | $0.60 | 4x |
How Much Does It Actually Cost?
Prices are quoted per million tokens, but real-world costs depend on your usage. Here are some concrete examples:
- A single chatbot message (500 input + 300 output tokens on GPT-4o mini): $0.000255 — essentially free.
- 1,000 chatbot messages/day on GPT-4o mini: $7.65/month.
- 1,000 chatbot messages/day on GPT-4o: $105/month.
- Summarizing 100 documents/day (4K input, 500 output on Claude Sonnet): $58.50/month.
The Pricing Landscape
AI model pricing spans roughly 100x from cheapest to most expensive:
- Budget tier ($0.10-0.15/M input): GPT-4o mini, Gemini Flash, SDXL — great for simple tasks at scale.
- Mid tier ($1-3/M input): GPT-4o, Claude Sonnet, Gemini Pro — the sweet spot for most applications.
- Premium tier ($5-15/M input): Claude Opus, o1 — maximum capability for complex reasoning.
Ways to Save Money
- Use the smallest model that meets your quality bar. GPT-4o mini handles 90% of tasks at 1/17th the cost of GPT-4o.
- Prompt caching. Anthropic and OpenAI cache repeated system prompts at 75-90% off.
- Batch API. If you don't need instant responses, batch pricing is typically 50% off.
- Open-source models. Running Llama or Mistral on your own GPU can be cheaper at very high volumes.
Ready to estimate your costs? Try our AI Token Cost Calculator to compare pricing across 27+ models instantly.