AI token counter and API cost calculator

What is a token?

A model doesn't read words. It reads tokens: small chunks of text that can be a whole word, part of a word, a space or a single character. And that's what you pay for. Every API price is per million tokens, once for what you send in and once for what the model writes back.

I'm Robin. I've been coding since I was twelve, and the apps I build use AI features: Postel drafts social posts with AI, and NativeExpress, my React Native starter, ships AI chat, image generation and a scanner. So the question "what will this cost me per user" comes up for me all the time. This tool is how I'd answer it.

How many tokens is a word?

It depends on the model and the language, and every company counts a little differently:

• Gemini: Google says a token is about 4 characters, and 100 tokens are about 60 to 80 English words (Google's docs).
• Claude: Anthropic says a token used to be about 3.5 English characters, and that Claude 4.7 and later make about 30% more tokens from the same text (Anthropic's docs). So the same prompt is more tokens on Claude than you'd guess from the price alone.
• OpenAI: the counter above uses o200k_base, the tokenizer OpenAI publishes for GPT-5 and GPT-4o, so that number is a real count, not a rule of thumb.

English is the cheap case. Other languages, code and anything with lots of numbers or symbols usually take more tokens per word. If you're building for German or Spanish users, paste a real German or Spanish prompt above instead of guessing.

How to estimate your API bill

The math is simple: tokens in times the input price, plus tokens out times the output price, times how often it happens. That's what the calculator does. What people get wrong is what goes into it:

• Output costs more than input. On most models above, a token the model writes is five times the price of a token you send. A short prompt that asks for a long answer is mostly output.
• Chats resend everything. In a chat feature, every new message sends the whole conversation again. Message 20 costs a lot more than message 1.
• Your system prompt goes with every request. A long set of instructions you never see is paid for every single time.
• Thinking counts as output. Google's pricing page lists its output prices with thinking included, so a model that reasons before it answers bills that reasoning too.

There are ways to pay less. Anthropic lists cheaper prices for cached prompts and 50% off for batch requests that don't need an answer right away (Claude pricing). And a smaller model is often enough: for short replies, Haiku, Luna or Flash-Lite cost a fraction of the big models.

For an exact count on Claude, Anthropic has a token counting API that tells you how many tokens a message is on the model you plan to use.

The expensive mistake isn't the price per token

It's an API key inside your app. When an AI builds an app from scratch, the key often ends up right in the app, and anyone can pull it out in about five minutes. With that key they use your AI on your bill, and no calculator helps you then.

So in my apps, the AI calls happen on the server and the key never touches the phone. That's how NativeExpress ships its AI features, with a free monthly quota and a Pro gate the server enforces. If you're building your first app with AI, my course covers exactly this, from the basics to an app in the store.