Tokens - The Hidden Currency of AI

🪙 You've been using ChatGPT, Claude & Gemini — but do you know what you're actually paying for?

Let's talk about tokens — the invisible currency behind every AI conversation.

What is a token?

A token is the smallest chunk of text an AI model reads and processes. Not a letter. Not a full word. Something in between.

A few examples to make it click:

"Hello" → 1 token

"ChatGPT" → 2 tokens (Chat + GPT)

"Cybersecurity" → 3–4 tokens

"नमस्ते" (Namaste in Hindi) → 5–6 tokens ← this matters if you work with regional languages!

The rule of thumb: ~1 token ≈ 4 characters ≈ ¾ of a word (in English)

Why should you care?

Because every AI API charges you per token — separately for what you send in and what the AI generates back.

Here's how the three big players price it (per million tokens, June 2026):

Notice something? Output tokens always cost more than input tokens. That's because generating text is computationally heavier than reading it.

Real-world cost: What does a typical interaction actually cost?

Say you send a 500-word email to the AI (≈ 667 input tokens) and get a 200-word reply (≈ 267 output tokens).

GPT-4o mini → ~$0.0003 per call → $0.03 for 100 calls/month

Gemini Flash-Lite → ~$0.00013 per call → effectively free at this scale

Claude Opus 4.6 → ~$0.03 per call → $3.00 for 100 calls/month

For individual use? Negligible. For enterprise applications running millions of calls? This is where smart model selection saves thousands of dollars a month.

How to check token counts (free tools)

✅ ChatGPT / OpenAI → platform.openai.com/tokenizer — paste any text, see tokens highlighted in colour

✅ Gemini / Google → Google AI Studio (aistudio.google.com) → token count shown in the sidebar

✅ Claude / Anthropic → No UI tool, but the Anthropic API has a count_tokens method you can call with a free account

3 quick tips to be token-smart

1. Write tighter prompts. "Summarise in 3 bullets" uses fewer tokens than "Could you please kindly help me summarise this into three bullet points?"

2. Beware of regional language tokens. A Hindi prompt costs 2–3× more tokens than the equivalent English prompt. Plan accordingly if you're building multilingual apps.

3. Use the right model for the job. Haiku/Flash for classification and extraction. Sonnet/Pro for reasoning. Opus only when truly needed.

Consumer plans hide this from you (on purpose)

If you're on ChatGPT Plus, Claude Pro, or Google AI Pro (all ~$20/month), you don't see any token billing — it's a flat subscription. Token pricing only matters when you're using the API directly to build applications.

But understanding tokens makes you a smarter user regardless — it explains why your long prompts get expensive at scale, why the AI sometimes "forgets" earlier parts of a long chat (context window limits), and why the same question sometimes costs more in Hindi than English.

If you'd rather see this explained with live examples, I've covered this on my YouTube channel — watch it below.