What Is a Token? How AI Usage Cost Is Calculated

A token is the smallest unit an AI model uses to process text — roughly a fragment of a word, averaging about 4 characters. When you connect to a model like ChatGPT, Claude, or Gemini through an API, what you pay for isn’t calculated by word count or character count — it’s calculated by the number of tokens processed. That means the same meaning, written in different languages or phrased differently, can map to a different token count, and therefore a different cost.

How Tokens Are Counted — A Concrete Example

(This concept is part of the full Agentic AI Guide.) To make token counting concrete, take a short sentence: “Yapay zeka artık küçük işletmelerin pazarlama sürecinde aktif olarak kullanılıyor.” This sentence comes out to roughly 30-32 tokens in Turkish. Its English equivalent with the same meaning — “AI is now actively used in the marketing process of small businesses.” — comes out to roughly 15-16 tokens.

That’s roughly half, and it’s not a coincidence. Turkish is an agglutinative language: every suffix attached to a root (“-lerin”, “-inde”, “-arak”, and so on) usually doesn’t exist as a single unit in the model’s token vocabulary — it gets split into two or three pieces. On top of that, since large language models’ token vocabularies are trained predominantly on English text, English words usually fit into a single token while Turkish words often split into 2-4 pieces. The result: for the same job, a business generating or processing Turkish content typically consumes 30-60% more tokens than the English equivalent — meaning a higher bill for the same work.

Note: the exact number depends on the model you’re using (Claude, GPT, Gemini — each has its own tokenizer); the figures above are approximate and meant to show direction, not precision.

How Cost Is Calculated

Your API bill is actually made up of two separate numbers: the token count of the text you send (input) and the token count of the model’s response (output). The two are usually priced differently — output tokens can cost 3-5x more than input, since generation is computationally more expensive than reading.

There’s also a model tier to consider. Small/fast models are very cheap per token for simple tasks; large/powerful models cost far more per token for tasks that require complex reasoning — sometimes a 10-20x difference. Picking the right model alone can cut your monthly bill in half. For concrete examples of price differences between models, see What Is Claude AI, and What Does It Cost.

To roughly estimate your monthly AI cost, you just need to multiply three numbers: (1) how many requests/operations you run per month, (2) the average tokens going back and forth per operation (input + output), and (3) your chosen model’s price per token. For example, a system that auto-summarizes 5,000 customer emails a month, averaging 800 tokens per email (500 input + 300 output), works out to 4 million tokens a month — depending on the model, that can range from a few dollars to hundreds of dollars.

Why It Matters

The practical takeaway: a long prompt or a long document means more tokens every single time — meaning more cost. For a one-off task, this barely matters. But if you’re pasting the same large document (a product catalog, a brand guide, a 40-page FAQ) into every question, you’re paying that document’s token cost all over again each time.

The approach that actually works here is “load once, reuse repeatedly.” Tools like Claude Projects let you upload a document once and have the model keep it on hand for every conversation afterward — you don’t have to resend the document every time, which cuts token consumption significantly. I walked through this setup step by step in Setting Up a Company Assistant with Claude Projects.

In short: unnecessarily long briefs, large documents sent over and over, and the “explain everything from scratch every time” habit quietly inflate your monthly AI bill without you noticing.

An Observation From My Own Operations

I learned this firsthand managing 43 ad accounts with AI agents. I had an agent running the same bulk reporting job for all accounts every day; at first I was resending the account’s entire history to the model on every run because it “felt safer.” When the bill came in higher than expected, I dug into why: I was re-tokenizing and resending the same 500-600 KB of historical data every single day — when really only the last 24 hours of change mattered. Once I switched the process to “only send the diff,” the token cost of the same job dropped to a tenth, with no drop in output quality.

You Can Ask Me to Handle This

If you want to see where your AI usage cost is actually going, pick the right model/tier, or move large repeatedly-sent documents to a “load once, reuse” setup, you can ask me to handle it: this is typically a 2-hour to 2-week job, remote, billed hourly. See the AI consulting page for details.

Frequently Asked Questions

How many words is 1 token?

There’s no exact ratio, but in English roughly 1 token equals 0.75 words (averaging 4 characters). In Turkish that ratio is lower, because suffixes get split into separate tokens — the same word count generally needs more tokens.

Why isn’t token cost based on characters or words?

Because the model doesn’t process text as raw characters — it processes it through a pretrained token vocabulary. Compute is spent per token, so pricing is done on that same unit.

Is a smaller model always cheaper?

Per token, yes, by a wide margin. But on complex tasks, its error rate is also higher. The right approach is giving simple, repetitive tasks to the small model and judgment-heavy tasks to the large model — running everything on the large model unnecessarily inflates cost.

How do I forecast my monthly AI cost in advance?

You can get a rough estimate by multiplying monthly operation count by average tokens per operation (input+output) and the model’s token price. Once real usage starts, calibrate that estimate against the first week’s actual data.