What is AI API Cost?

Tokens, not words

AI APIs bill per token, not per word. 1 token ≈ 0.75 English words. A 300-word email is ~400 tokens. Code and non-English languages tokenize less efficiently — sometimes 2–3× more tokens per character.

Input vs output

Every request has an input cost (what you send) and an output cost (what the model generates). Output is almost always 3–5× more expensive per token than input.

Per-model rates

Cheaper models exist for simple tasks. Gemini Flash is ~30× cheaper than GPT-4o per token. Choosing the right model is the highest-leverage cost decision.

Hidden costs

System prompts re-billed each call. Retries on errors. Tool-call overhead. Streaming has no cost effect but matters for UX. None of these show up in marketing pricing pages.

A minimal worked example

Say you send a 300-word question (about 400 tokens) and get back a 200-word answer (about 270 tokens) on GPT-4o mini ($0.15/$0.60 per 1M). Input cost: 400/1,000,000 * $0.15 = $0.00006. Output cost: 270/1,000,000 * $0.60 = $0.000162. One call costs about $0.00022 — but at 500,000 calls a month that's roughly $110, which is why per-call intuition breaks down at scale.

How to model this as a scenario with your own volumes

Instead of guessing, define three numbers: average input tokens per request, average output tokens per request, and requests per month. Multiply each by the relevant per-1M rate for a candidate model, then repeat for two or three alternative models to see the spread. This is exactly the calculation a token-volume calculator like TokenAtlas runs — you supply the volumes, it applies the published rates.

Decision checklist for a first estimate

— Do you know your average prompt length in tokens, not words? — Have you separated input and output token counts, since they price differently? — Have you picked a realistic monthly call volume, including retries? — Have you compared at least two model tiers before committing?

What this explanation deliberately leaves out

This is a conceptual overview, not a billing reconciliation. It does not cover provider-specific discount programs, enterprise contracts, or how to pull real usage data from a live account — for that you need the provider's own dashboard or invoice. What's here is enough to build a first-pass cost model from stated or assumed token volumes.

Frequently asked questions

How do I know my token count?

Use the provider tokenizer or TokenAtlas's live counter — paste text and see tokens instantly.

Are there free tiers?

Google offers a generous Gemini free tier; OpenAI and Anthropic offer trial credits.

Can I cap my spend?

Yes — every provider supports monthly budget caps. Always set one.