Most LLM price pages lead with dollars per million tokens. That number is useful, but it does not tell you what a request costs. The ratio between input and output decides the bill, and every product has a different ratio.
Pick a fixed call shape and the comparison becomes real. Take one request with 2,000 input tokens and 500 output tokens. If the input rate is i dollars per million and the output rate is o dollars per million, the cost is:
(2,000 x i + 500 x o) / 1,000,000
The output share is 500 x o divided by that total. When output is priced at four times input, output is half the cost of the call even though it is a fifth of the tokens. A table sorted by input price alone will point you at the wrong model.
This is the call shape used on AI API Cost Calc: 2,000 input and 500 output tokens, official per-million rates, one comparison across GPT, Claude, Gemini, Grok, DeepSeek and Mistral. The site also shows the minimum price per user at a target margin, because a per-call cost only matters after you divide by users and subtract what you keep.
Three things the per-token headline hides:
Cached input. Long system prompts and repeated context are often billed at a cache-read rate that is lower than fresh input. A cost page that lists one input number overstates the bill for chat products with a fixed prompt, and understates nothing. If the model has a cache rate, it belongs in the table.
Output length is a product decision. A support bot that answers in 120 tokens and a writing tool that answers in 900 tokens do not have the same cost, even on the same model. Fix the output length you actually ship, then compare.
Margin is not a markup. If you want to keep 70% of revenue, the price is cost divided by 0.30, not cost times 1.7. Small difference per call, large difference per thousand users.
The limits are worth stating. This model does not include retries, tool calls, embedding, storage, or failed requests you still pay for. Rates change, and a price table is only as good as its checked date. Treat the number as a floor for planning, then measure real usage in the provider dashboard after launch.
If you are choosing between two models, run both through the same call shape at the same margin. The one with the lower headline rate is not always the cheaper one.
See the live comparison at AI API Cost Calc.
Top comments (0)