API cost per feature: three numbers before you ship

The cost of an AI feature comes from three numbers: input tokens, output tokens and price per million tokens. Everything else is a guess, and guesses do not pass a budget review.

Three numbers, not one

• Cost per call: input plus output of a single request.
• Calls per user per month: read it from the log, do not estimate it from the product.
• Cost per feature: total calls times unit cost, plus retries and errors.

The sum, in code

const MODELS = {
  mini: { input: 0.4, output: 1.6 },
  reasoning: { input: 1.1, output: 4.4 },
  sonnet: { input: 3, output: 15 },
};

export function costPerCall(model, inputTokens, outputTokens) {
  const price = MODELS[model];
  return (inputTokens * price.input + outputTokens * price.output) / 1_000_000;
}

for (const name of Object.keys(MODELS)) {
  const one = costPerCall(name, 3000, 400);
  console.log(name, one.toFixed(5), (one * 10000).toFixed(2));
}

With 3,000 input tokens and 400 output tokens, one call costs 0.00184 USD on gpt-4.1-mini, 0.00506 on o4-mini and 0.015 on claude-sonnet-4-6. Ten thousand calls a month make 18.40, 50.60 and 150.00 USD. Save the block as cost.mjs and run it with node cost.mjs.

The list in a table

The gap between the small model and the reasoning one is 2.75 times on the same call: 0.00506 divided by 0.00184. Before paying it, a test on your case set must show that the dearer model solves cases the other one misses.

Context is the expensive line

Every turn resends the whole conversation: ten turns of 3,000 tokens of history send 165,000 tokens in total, not 30,000, because the sum grows with the square of the turns. Summarising or cutting the history costs less than moving to a smaller model.

What stays out

• Embeddings and images: billed separately, on different lists.
• Retries and errors: one retry after a timeout doubles that turn.
• Evaluations and tests: every eval run burns tokens like production.

The limitation

The sum holds for one model and one provider, and the list changes without notice: 0.40 and 1.60 USD per million is the value read on 1 November 2026, not a promise. Estimating tokens from characters is approximate, and the reasoning tokens of a reasoning model are billed as output but never appear in the text.

Where to continue

The cost of a full task is in an agent's cost per task, the weight of long context in long context and cost, the measured comparison between models in evaluating a model for code. To show the sum you need a pricing table or a project data table, inside dashboard components and 111,985 components.