Price a SaaS feature before execution

Translate a representative workload into exact COGS, credits, and margin guardrails without floating-point money errors.

What you will build

A server-side preflight that estimates the cost of one product action, converts the quote to integer micro-USD, and decides how many internal credits to reserve before running the model.

1. Model representative usage

Use percentiles from production traces, not a context-window maximum or a provider’s headline price. Keep distinct profiles for features whose input/output mix differs.

TypeScript

const usageProfiles = {
  summarizeDocument: {
    inputTokens: 80_000,
    cachedInputTokens: 0,
    outputTokens: 2_500,
  },
  regenerateSummary: {
    inputTokens: 80_000,
    cachedInputTokens: 72_000,
    outputTokens: 2_500,
  },
} as const;

2. Quote the deployment-specific COGS

TypeScript

const quote = await allm.pricing.calculate({
  deployment: "dep_openai_gpt_5_1_responses_global",
  ...usageProfiles.summarizeDocument,
});

console.log(quote.totalUsd); // exact decimal string
console.table(quote.lineItems);

3. Apply margin using integer money

TypeScript

function usdToMicros(value: string): bigint {
  const [whole, fraction = ""] = value.split(".");
  return BigInt(whole) * 1_000_000n +
    BigInt(fraction.padEnd(6, "0").slice(0, 6));
}

function microsToUsd(value: bigint): string {
  const whole = value / 1_000_000n;
  const fraction = (value % 1_000_000n).toString().padStart(6, "0");
  return `${whole}.${fraction}`;
}

const costMicros = usdToMicros(quote.totalUsd);
const priceMicros = (costMicros * 250n + 99n) / 100n; // 2.5× markup
const credits = (priceMicros + 9_999n) / 10_000n;     // $0.01 per credit

await creditLedger.reserve({
  customerId,
  feature: "summarize_document",
  credits,
});

Example output

JSON

{
  "feature": "summarize_document",
  "deployment_id": "dep_openai_gpt_5_1_responses_global",
  "estimated_cost_usd": "0.125000",
  "customer_price_usd": "0.312500",
  "credits_reserved": "32",
  "usage_profile": "p75_2026_07"
}

Production guardrails

Reserve using estimated usage, then settle with actual input, cached input, and output tokens.
Keep retries as separate cost events; do not pretend failed paid calls were free.
Version usage profiles and pricing policies so historical margins remain explainable.
Reject or require confirmation when estimated cost exceeds a feature budget.
Use private price books for negotiated rates instead of modifying public catalog data.