Why Every Token Costs Money: AI Pricing, Explained

AI APIs publish prices like “$3 per million input tokens” — numbers so small they look free. They aren’t. Per-token pricing scales exactly like a metered utility, and every prompt you send is billed twice: once when the model reads it (input), and once when it answers (output, usually several times pricier).

How the meter runs

Say an agent re-sends a 3,000-token system prompt with every request, 500 times a day. That’s 1.5 million input tokens per day just to keep the instructions on the table — before any user content or any answers. At typical mid-tier rates, that recurring prompt alone can cost more per month than the actual work being done.

The structure repeats everywhere:

• Input tokens: your prompt, your documents, your conversation history.
• Output tokens: the reply, generated at a premium rate.
• Cached or repeated context: many providers now discount repeated blocks, but the discount is never zero.

The cheapest lever you control

You can’t negotiate the per-token rate. You can send fewer tokens:

1. Cut filler and pleasantries. “I was wondering if you could please…” reads as politeness to you and as billable tokens to the model. An assistant loses nothing without it.
2. Don’t re-send what the model already has. Restating the whole conversation in every turn is the classic leak.
3. Trim system prompts. They run on every single call. A 20% cut is a 20% cut on your biggest recurring line item.

A worked example

Our before/after demo takes a wordy 170-token research prompt down to 117 — a 31% cut that changes nothing the model is asked to do. Multiply that by thousands of calls, and prompt hygiene stops being a nicety and starts being engineering.

Trim your own prompts at supatoken.lol — free, no signup, and nothing leaves your browser.