OpenAI API Cost, Explained Properly

OpenAI bills per token, not per message. A token is a chunk of text — roughly 4 characters, or about 0.75 of a word in English. As a working rule:

• 750 words ≈ 1,000 tokens
• A typical article of 1,500 words ≈ 2,000 tokens
• A short paragraph ≈ 100 tokens

Every request bills two separate amounts:

• Input tokens — everything you send: your instruction plus any text you pasted or selected
• Output tokens — what the model writes back

Output is almost always priced higher than input, often several times higher. That single fact drives most of the practical cost advice below.

Official pricing is quoted per 1 million tokens, which makes everything look expensive until you do the division. The arithmetic:

cost = (input tokens ÷ 1,000,000 × input price) + (output tokens ÷ 1,000,000 × output price)

Work an example. Say you summarize a 1,500-word article (~2,000 input tokens) and get a 200-word summary back (~270 output tokens). Even at a relatively expensive rate of $5 per million input and $15 per million output, that is:

• Input: 2,000 ÷ 1,000,000 × $5 = $0.01
• Output: 270 ÷ 1,000,000 × $15 = $0.004
• Total: about 1.4 cents

On a cheaper small model, the same job costs a small fraction of that. Rates change often, so check the current numbers on OpenAI's pricing page and plug them into the same formula.

The real question is not 'is the API cheap' but 'is it cheaper than what I'd otherwise pay.' Take a flat $20/month subscription as the benchmark and work out your break-even.

If an average query costs you roughly a cent, $20 buys about 2,000 queries a month — around 65 every single day. Most people using AI for summarizing, rewriting, and explaining land nowhere near that, which is why per-token billing usually wins.

Where a subscription wins:

• You genuinely run very high volume every day
• You want features that only exist in the chat product, not the API
• You'd rather have one predictable bill than variable usage

Where pay-per-token wins:

• Bursty usage — heavy some weeks, nothing in others (quiet months cost you nothing)
• You want to switch between providers and models per task
• You want your data going straight to the provider with no middleman

Most overspending comes from a few avoidable patterns:

1. Send less input. Select the two paragraphs you care about instead of the whole page. Input tokens are what balloon when you paste entire articles.
2. Cap the output. Asking for 'a 3-bullet summary' rather than an open-ended answer directly reduces the more expensive half of the bill.
3. Match the model to the task. Use a small cheap model for summarizing, formatting, and translation; save the expensive reasoning models for problems that need them. This is usually the single biggest saving.
4. Don't resend context you don't need. Long back-and-forth threads re-send the whole conversation as input each turn, so costs climb quietly. Start fresh when the topic changes.

Estimating is useful; a hard cap is better. In the OpenAI platform under Settings → Limits, set a monthly budget that stops API calls when reached, plus an alert threshold below it so you get warned first.

Because the API is prepaid, your balance is itself a natural ceiling — you can simply not enable auto-recharge and top up manually when it runs out. If you do enable auto-recharge, pair it with a budget limit; that combination is the only way a bill can genuinely surprise you.

The platform's Usage dashboard shows spend by day and by model, which quickly reveals where your money actually goes — usually one model or one habit.

If you use your key through a browser extension, it helps to see cost at the point of use. Prompt Anything Pro includes a built-in token tracker alongside its model selector, so you can watch what a given prompt costs and switch to a cheaper model when the task doesn't need the expensive one. Your key stays in your browser and requests go straight to OpenAI, so what you see billed is exactly what you sent.