APIs, model versions & pricing

By the end of this page, you'll be able to look at any AI API pricing page, estimate what a month of usage costs, and avoid the three most common billing surprises.

What an API actually is

An API (Application Programming Interface) is a contract: you send structured data (a prompt, some context), you get structured data back (a completion, a classification, a JSON object). No chat window, no UI — just a request and a response over HTTP.

The simplest possible call looks like this:

curl https://api.example.com/v1/chat \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-4o",
    "messages": [
      {"role": "user", "content": "Summarize this in one sentence: [text]"}
    ]
  }'

That’s it. Every AI tool you’ve ever used — ChatGPT, Copilot, your app’s AI feature — is doing some version of this under the hood.

Why input and output cost different

Most AI APIs charge per million tokens, with two rates:

	Rate	Why
Input tokens (your prompt)	Cheaper	The model reads these — lower compute
Output tokens (the answer)	2–5× more expensive	The model generates these one by one — sequential, slower, more compute

A typical pricing page:

Input: $3 per 1M tokens Output: $15 per 1M tokens

That 5× difference is normal and reflects the real cost asymmetry of generation vs. reading.

A worked example

Let’s say you build a tool that:

Sends a 2,000-token prompt (document + question)
Gets back a 500-token answer
Runs 1,000 times a day

Input:  2,000 × 1,000 = 2,000,000 tokens  → $3.00
Output:   500 × 1,000 =   500,000 tokens  → $7.50
Total per day:                                              ≈ $10.50
Total per month:                                            ≈ $315

That’s a real, usable number for a solo dev. Scale it by 10× and you’re in “we need to optimize” territory.

The three billing surprises

The system prompt is charged every time. If your app has a 2,000-token system prompt, that’s 2,000 input tokens on every single call. At high volume, the system prompt is often the biggest line item.
Cached input is cheaper — if you use it. Many providers charge less for input tokens they’ve seen before (prompt caching). If your prompts are repetitive, check whether caching is on.
Free tiers have ceilings. “Free” often means “free up to 50 requests/day” or “free for 30 days”. Know the ceiling before you build on it.

Habits that keep the bill predictable

Pin your model version. Models get deprecated; your code should fail loudly, not silently route to a new (possibly more expensive) model.
Set a monthly budget in the provider’s dashboard, with alerts at 50% and 80%.
Use a smaller model for high-volume tasks. A 10× cheaper model is usually good enough for classification, extraction, and routing. Save the expensive one for the hard 20%.
Log token counts per request. You can’t optimize what you don’t measure.

Try it: open your AI provider's usage dashboard. Find your top 3 most expensive requests this month. In 9 times out of 10, one of them is a system prompt that's 3× longer than it needs to be. Shorten it, and your bill drops.