AI API Costs Explained: What You Actually Pay for GPT, Claude & Gemini
The Hidden Language of AI Pricing
When OpenAI, Anthropic, or Google publish their API prices, they use a unit that most people have never encountered: dollars per million tokens. Unless you work in natural language processing, this number means almost nothing at first glance. Is $2.50 per million input tokens cheap? Is $75 expensive? And what even is a token?
Understanding AI pricing is becoming increasingly important. Whether you are building a chatbot, automating document processing, or just experimenting with AI APIs, the costs can range from a few cents to thousands of dollars per month — depending on your choices.
What Is a Token?
A token is the fundamental unit that language models use to process text. It is not a word, not a character, and not a syllable — it is something in between. Most modern tokenizers (like OpenAI’s tiktoken or Anthropic’s tokenizer) break text into subword pieces.
As a rough rule of thumb:
- English text: 1 token ≈ 4 characters ≈ 0.75 words
- German text: 1 token ≈ 3 characters ≈ 0.6 words (German words are longer on average)
- Code: varies widely, but typically more tokens per “word” due to special characters
This means a 1,000-word English article uses approximately 1,333 tokens. A full A4 page (about 250 words) is roughly 333 tokens.
How AI API Pricing Works
Every major AI provider charges separately for input tokens (what you send to the model) and output tokens (what the model generates). Output tokens are always more expensive because they require more computation — the model has to generate each token sequentially, while input tokens are processed in parallel.
Here is a snapshot of current pricing (as of February 2026):
| Model | Input ($/1M tokens) | Output ($/1M tokens) |
|---|---|---|
| GPT-4o | $2.50 | $10.00 |
| GPT-4o Mini | $0.15 | $0.60 |
| Claude Sonnet 4 | $3.00 | $15.00 |
| Claude Haiku 3.5 | $0.80 | $4.00 |
| Gemini 2.0 Flash | $0.10 | $0.40 |
The price difference between the cheapest and most expensive model is over 100x. Choosing the right model for your use case is the single most impactful cost optimization you can make.
The Real Cost of a Chatbot
Let’s make this concrete. Imagine you are running a customer support chatbot that handles 500 conversations per day. Each conversation averages 2,000 tokens (input + output combined, roughly split 60/40).
With GPT-4o:
- Daily tokens: 500 × 2,000 = 1,000,000 tokens
- Daily cost: ~$5.50
- Monthly cost: ~$165
With GPT-4o Mini:
- Same volume
- Daily cost: ~$0.33
- Monthly cost: ~$10
With Gemini 2.0 Flash:
- Daily cost: ~$0.22
- Monthly cost: ~$6.60
The difference is staggering. For many chatbot use cases, the cheaper models perform well enough. The key is testing whether the quality meets your requirements before committing to a premium model.
Hidden Cost Multipliers
Several factors can inflate your AI costs beyond the basic per-token pricing:
System prompts eat tokens silently. Every API call includes your system prompt. If your system prompt is 500 tokens and you make 10,000 calls per day, that is 5 million tokens per day just for instructions — before any user input or model output.
Conversation history grows fast. In multi-turn conversations, you typically send the entire conversation history with each new message. A 10-message conversation might send the first messages 10 times, the second 9 times, and so on. This quadratic growth is the number-one cause of unexpectedly high bills.
Retries and error handling. Failed requests that you retry still count. Rate limiting that causes retries doubles your cost for those requests.
Strategies to Reduce Costs
Choose the right model. Use the cheapest model that meets your quality bar. Many tasks (summarization, classification, extraction) work perfectly well with smaller models.
Truncate conversation history. Instead of sending the full history, keep only the last N messages, or summarize older messages into a condensed system prompt.
Cache repeated queries. If many users ask similar questions, cache the responses. Even a simple hash-based cache can dramatically reduce API calls.
Use batch APIs. Most providers offer batch processing at 50% discount. If your use case can tolerate higher latency (hours instead of seconds), batch processing is significantly cheaper.
Try It Yourself
Curious how much your AI project will cost? Use our API Cost Calculator to select a model and instantly see what your token usage will cost. For chatbot-specific estimates, try the Chatbot Cost Calculator — enter your expected conversation volume and get daily, monthly, and yearly projections.
Want to compare all models side by side? The Model Price Comparison table lets you sort by input price, output price, or context window to find the best fit.
Fun Fact: GPT-4’s training is estimated to have cost over $100 million. Yet a single API call to GPT-4o Mini costs about $0.0003 — roughly the price of a grain of rice. The economics of AI are strange: astronomical fixed costs, near-zero marginal costs.