What AI actually costs: tokens, credits & messages, explained.

Every AI company prices differently — tokens, credits, "messages," per-seat — and the names are confusing on purpose. Once you understand the three or four models underneath, you can answer the only question that matters: what is this actually going to cost me? Here's the plain-English version, plus what a single prompt really runs you on each tool.

01 The 4 ways AI gets priced

Strip away the branding and almost every AI uses one of four pricing models:

• Per-token (pay-per-use / API). You pay for the text in and out, measured in "tokens." This is how the APIs charge — you pay only for what you use.
• Flat subscription with usage caps. The $20/month apps (ChatGPT Plus, Claude Pro, etc.). You pay a fixed price and get generous-but-limited usage — measured in "messages," not dollars.
• Credits. A prepaid bucket you spend down — one action costs N credits. Common for image/video generation, and now for agentic work: Microsoft's Copilot Cowork bills $0.01 per credit, where a single multi-step task runs anywhere from ~70 to 1,500+ credits.
• Per-seat. Business pricing — a flat monthly fee per employee (e.g., Microsoft 365 Copilot). The company pays per person, not per use.

02 What's a "token," really?

A token is just a chunk of text — roughly ¾ of a word in English. "Learning" might be one token; "antidisestablishmentarianism" is several. A rule of thumb: ~750 words ≈ 1,000 tokens.

Every prompt has two token counts that both cost money:

• Input tokens — everything you send (your prompt + any document or context).
• Output tokens — everything the AI writes back. Output is almost always priced higher than input — often 4–5× — because generating is the expensive part.

So a "cost per prompt" = (your input tokens × input price) + (the reply's output tokens × output price). That's the whole formula. Everything below is just plugging in numbers.

03 What each tool charges (June 2026)

Here's the honest, side-by-side. "Pay-per-use" is the API price per million tokens (input / output). "Subscription" is the standard individual plan. The last column is roughly what one medium prompt costs on the API (about 800 tokens in, 1,200 out).

*Perplexity prices around "searches," not raw tokens — Pro gives you a large daily allowance; the Sonar API charges per request plus tokens. Copilot is per-seat, so individual prompts have no marginal cost. Premium reasoning models (GPT-5.4, Claude Opus) cost more than the figures above. Prices as of June 2026 — always check current rates.

04 What a real prompt costs (worked examples)

Numbers feel abstract until you map them to real work. Using mid-range pricing:

• A quick question + short answer (≈250 in / 350 out): a fraction of a cent — well under $0.01 on any model.
• Summarize a 2-page document (≈800 in / 1,200 out): about $0.01–$0.02 on most models; under a tenth of a cent on Grok 4.1 Fast.
• Analyze a long report (≈2,500 in / 3,500 out): roughly $0.05–$0.07 on a mid model; more on premium reasoning models.

05 Subscription or pay-per-use — which is cheaper for you?

The honest decision rule:

• Light-to-moderate user (a few to a few dozen prompts a day)? Your API usage would likely cost less than $20/month — but the subscription buys you the polished app, no metering anxiety, and features like voice and image gen. Most people happily pay for that.
• Heavy daily user (hundreds of prompts, or long documents all day)? A flat subscription is the better deal — unlimited beats per-token once you cross a few dollars of daily usage.
• Builder / automating things? The API (pay-per-use) is the right tool — you only pay for what runs, and you can pick a cheap model (Grok 4.1 Fast, Gemini Flash) for high-volume jobs.