The same model is billed by the message or by the token

Start here

A reader who hears the word 'API' attached to ChatGPT, Claude or Gemini and assumes it is just the developer's way of saying 'the app' is missing a distinction: the API is a separate product, billed by a separate unit — a token, roughly a few characters of text — rather than the flat monthly price a subscription charges. Understanding that unit is what makes an API price comparable to anything, since 'per month' and 'per million tokens' cannot be compared without doing the arithmetic.

What the documents say

OpenAI's developer pricing page, retrieved 16 September 2026 (found after openai.com/api/pricing/ began redirecting to a business-seat page instead), lists its flagship model at $10 per million input tokens and $50 per million output tokens, alongside a cheaper tier at $0.20 input and $1.20 output per million tokens. Anthropic's API pricing documentation states Claude Sonnet 5 at $2 per million input tokens and $10 per million output tokens, noting this introductory price 'is now the standard price' after a scheduled increase was cancelled, alongside a discounted Batch API tier at half those rates. Google's Gemini API pricing page states its free tier's content is 'used to improve our products,' while the paid tier's is explicitly 'not used to improve our products,' and prices its Flash-tier model at $0.75 per million input tokens through 2026. All three price by the token rather than the message or the month, and tie a free or discounted tier to reduced capability or the use of submitted content.

Check this

The concrete check for a reader curious whether an API is relevant to them is simpler than the tables suggest: most people encounter an API not by calling it directly but through a third-party tool that bills per token behind the scenes. The check is whether a tool's own pricing page mentions API costs or 'bring your own key' — if so, it is a wrapper around this same token pricing, and the free-versus-paid tradeoff described above applies one level removed.

What holds and what fails

Token pricing holds as the accurate unit for any tool built on these APIs, and comparing per-token rates is a reasonable way to estimate a heavy workload's cost. It fails as a direct comparison to a subscription's flat fee without translating usage into an estimate, and it fails to predict a specific tool's price, since a wrapper adds its own markup on top. This bridge between two pricing systems is editorial; the vendor pages do not draw it themselves.

• Check whether a third-party AI tool bills its own subscription or passes through API token costs directly.
• Compare a free API tier's data-use terms against the paid tier's, since they can differ, as Google's page states.
• Estimate token volume for an actual task before assuming a per-million-token price is cheap or expensive.

A developer account is not a hidden discount on the same product; it is a different pricing unit for the same models, useful mainly to someone building or estimating costs for a tool, not someone simply choosing a chat app.