LLM Strategy - OURPAY

Record one event after each model call and include the structured _llm metadata. A meter can aggregate metadata._llm.total_tokens, input_tokens, or output_tokens, while Cost Insights reads the optional _cost object. Use a stable external_id for retries. If the same event is sent again, OurPay counts it as a duplicate instead of billing the customer twice.

TypeScript

import { createOurPay } from "@ourpay-dev/sdk/2026-10";

const ourpay = createOurPay({
  accessToken: process.env.OURPAY_ACCESS_TOKEN ?? "",
});

await ourpay.events.ingest({
  events: [
    {
      name: "llm.usage",
      external_id: "generation_01JXYZ",
      external_customer_id: "customer_123",
      metadata: {
        _llm: {
          vendor: "openai",
          model: "your-model-id",
          input_tokens: 100,
          cached_input_tokens: 10,
          output_tokens: 200,
          total_tokens: 300,
        },
        _cost: {
          amount: 123,
          currency: "usd",
        },
      },
    },
  ],
});

Python

import os

from ourpay.v2026_10 import OurPay

with OurPay(os.environ["OURPAY_ACCESS_TOKEN"]) as ourpay:
    result = ourpay.events.ingest(
        events=[
            {
                "name": "llm.usage",
                "external_id": "generation_01JXYZ",
                "external_customer_id": "customer_123",
                "metadata": {
                    "_llm": {
                        "vendor": "openai",
                        "model": "your-model-id",
                        "input_tokens": 100,
                        "cached_input_tokens": 10,
                        "output_tokens": 200,
                        "total_tokens": 300,
                    },
                    "_cost": {
                        "amount": 123,
                        "currency": "usd",
                    },
                },
            }
        ]
    )

    print(result.inserted, result.duplicates)

_cost.amount is expressed in cents and can be fractional for very small costs. For example, 123 is US

1.23 and `0.025` is US

0.00025.

Send the event only after the model provider returns final usage totals. Streaming applications should ingest once when the stream closes, not once per chunk.