Metering

meterTelemetry from better-supabase/ai-sdk is an AI SDK telemetry integration that records the tokens of every model call on the usage block. Register it once, and calls made anywhere in the app (the assistant, agents, a job, a batch) are metered on the tenant without passing usage around.

import { registerTelemetry } from "ai";
import { meterTelemetry } from "better-supabase/ai-sdk";

import { serviceUsage } from "@/lib/usage";

export function register() {
  registerTelemetry(meterTelemetry({ usage: serviceUsage }));
}

An agent with several steps records each step. The tenant comes from the org: tag that gatewayOptions puts in providerOptions.gateway.tags, then from runtimeContext.organizationId; a call without either isn't metered. Pass tenant(event) to read it from somewhere else, and meters to rename the meters.

Each row carries an idempotency key per call and step, so a retried event is counted once. Don't also record the same tokens with usageEntries, or they count twice. A failed write goes to onError; it never fails the model call.

Nightly spend reconciliationCopy Anchor Link

The per-call cost in the usage block comes from each response. The AI Gateway's spend report is the figure you are billed, and the two drift when a response arrives without a cost or a call fails after it was charged. spendReconciliation is a job handler that reads yesterday's report grouped by tag and records each tenant's cost, in micro-dollars, on its own meter:

import { gateway } from "ai";
import { spendReconciliation } from "better-supabase/ai-sdk";

export const handlers = {
  "ai.spend_reconcile": spendReconciliation({
    gateway,
    usage: serviceUsage,
    credentialType: "system",
  }),
};

Schedule it once a day after midnight UTC. The payload can name a day (YYYY-MM-DD) to reconcile another one. Each tenant's day is recorded once on ai.gateway_cost (set meter to change it), so running the job twice changes nothing. credentialType: "system" counts only the spend the app pays, leaving out calls made with a tenant's own key. Compare ai.gateway_cost with the per-call cost meter to find the drift.