Recipe — Aggregate AI cost per tenant

Scenario. You run a multi-tenant SaaS where every customer hits the same support agent. Finance needs a monthly invoice line per tenant, and engineering needs to see which tenants burn reasoning tokens (expensive) versus cache hits (cheap). One number — “we spent $4,210 on OpenAI” — answers neither question.

The fix: stamp every execute() with the tenant id, read the per-channel Usage.cost breakdown off the result, and roll it into a tenant-keyed ledger. Pricing comes from a ModelPricing registry configured once on the SDK, so the cost is computed at emit time and stored as a historical fact.

Configure pricing once, on the SDK

ModelPricing is USD per 1,000,000 tokens. Declare it on the SDK adapter as a registry keyed by model name; every model the SDK produces inherits the matching entry. With pricing present, every report carries a Usage.cost breakdown — without it, cost is undefined (honest absence, never a false zero).

import { OpenAISDK } from "@warlock.js/ai-openai";
import { ai } from "@warlock.js/ai";

const openai = new OpenAISDK({
  apiKey: process.env.OPENAI_API_KEY!,
  pricing: {
    // USD per 1M tokens — copy straight from the provider's pricing page.
    "gpt-4o-mini": { input: 0.15, output: 0.6, cachedInput: 0.075 },
    "gpt-4o": { input: 2.5, output: 10, cachedInput: 1.25 },
    // o-series prices reasoning at the output rate by default; set
    // `reasoning` only when the provider bills it separately.
    "o4-mini": { input: 1.1, output: 4.4, reasoning: 4.4 },
  },
});

const supportAgent = ai.agent({
  model: openai.model({ name: "gpt-4o-mini" }),
  systemPrompt: ai.systemPrompt().instruction("Answer the customer's billing question."),
});

Stamp every run with the tenant id

sessionId is the opaque grouping key the framework mirrors onto every report node (and every nested executable). Use the tenant id as the session id — then a flat query over your ledger never has to walk the report tree.

async function answerForTenant(tenantId: string, question: string) {
  const result = await supportAgent.execute(question, {
    sessionId: tenantId,
  });

  if (result.error) {
    // Cost is still real even on failure — the model was called. Record it
    // before returning so a tenant's failed runs still show up on the invoice.
    await recordSpend(tenantId, result);
    throw result.error;
  }

  await recordSpend(tenantId, result);

  return result.text;
}

Read the cost breakdown off the result

Usage.cost mirrors ModelPricing — input, output, and the optional cachedInput / cachedOutput channels. The token sub-channels (cachedTokens, reasoningTokens, cacheWriteTokens) are separate counters, present only when the provider meters them. There is no single scalar total: sum the populated cost fields, treating undefined as 0.

import type { AgentResult, ModelPricing } from "@warlock.js/ai";

/** Sum a per-channel cost breakdown into one USD scalar. */
function totalUSD(cost: ModelPricing | undefined): number {
  if (!cost) {
    return 0;
  }

  return (
    cost.input +
    cost.output +
    (cost.cachedInput ?? 0) +
    (cost.cachedOutput ?? 0)
  );
}

async function recordSpend(tenantId: string, result: AgentResult): Promise<void> {
  const { usage, report } = result;

  await tenantSpendRepo.create({
    tenantId,
    runId: report.runId,
    modelName: report.model.name,
    provider: report.model.provider,
    status: report.status,
    // Token channels — undefined when the provider doesn't meter them.
    inputTokens: usage.input,
    outputTokens: usage.output,
    cachedTokens: usage.cachedTokens ?? 0,
    reasoningTokens: usage.reasoningTokens ?? 0,
    // Cost channels — the historical USD breakdown captured at emit time.
    inputUSD: usage.cost?.input ?? 0,
    outputUSD: usage.cost?.output ?? 0,
    cachedInputUSD: usage.cost?.cachedInput ?? 0,
    totalUSD: totalUSD(usage.cost),
    startedAt: report.startedAt,
    endedAt: report.endedAt,
    durationMs: report.duration,
  });
}

Roll it up across composites with one middleware

Calling recordSpend by hand works for a single agent. The moment you wrap the agent in a workflow, supervisor, or orchestrator, the cleanest place to record is a middleware that fires once per execute() — it sees the rolled-up usage (own spend plus every child) and the sessionId off the run options.

import type { AgentMiddleware } from "@warlock.js/ai";

const tenantLedger: AgentMiddleware = {
  name: "tenant-ledger",
  execute: {
    after: async (ctx, result) => {
      const tenantId = ctx.options?.sessionId;

      if (!tenantId || !result.usage.cost) {
        return;
      }

      await tenantSpendRepo.create({
        tenantId,
        runId: result.report.runId,
        modelName: ctx.model.name,
        provider: ctx.model.provider ?? "unknown",
        status: result.report.status,
        inputTokens: result.usage.input,
        outputTokens: result.usage.output,
        cachedTokens: result.usage.cachedTokens ?? 0,
        reasoningTokens: result.usage.reasoningTokens ?? 0,
        totalUSD: totalUSD(result.usage.cost),
        durationMs: result.report.duration,
      });
    },
  },
};

const supportAgent = ai.agent({
  model: openai.model({ name: "gpt-4o-mini" }),
  middleware: [tenantLedger],
});

Attach tenantLedger to every agent and the ledger becomes a single source of truth — one row per run, already tenant-keyed and priced.

Query the monthly invoice

With the breakdown stored as columns, the invoice and the engineering view are two queries over the same table:

-- Per-tenant monthly spend (the invoice line).
SELECT tenant_id,
       SUM(total_usd)      AS spend_usd,
       SUM(reasoning_tokens) AS reasoning_tokens,
       SUM(cached_tokens)  AS cached_tokens,
       COUNT(*)            AS runs
FROM tenant_spends
WHERE started_at >= date_trunc('month', now())
GROUP BY tenant_id
ORDER BY spend_usd DESC;

-- "Which tenants are reasoning-heavy?" — reasoning tokens as a share of output.
SELECT tenant_id,
       SUM(reasoning_tokens)::float / NULLIF(SUM(output_tokens), 0) AS reasoning_share
FROM tenant_spends
WHERE started_at >= now() - interval '7 days'
GROUP BY tenant_id
ORDER BY reasoning_share DESC;

Production notes

Cost is a stored fact, not a re-derivation

Usage.cost is computed when the report is built, from the pricing table in effect at that moment. Store the USD columns, not just the token counts — if you re-derive cost later against today’s ModelPricing, a mid-month price change silently rewrites last week’s invoice. The framework captures it as history precisely so your ledger stays accurate after the upstream table shifts.

Unpriced runs are undefined, not 0

A run on a model with no pricing entry leaves usage.cost as undefined. The middleware’s if (!result.usage.cost) return; guard skips those rows rather than booking a false $0. Aggregators merge only defined channels, so one unpriced child never erases a priced sibling’s cost in a composite’s rolled-up total. If you want every run on the ledger, add an explicit pricing entry for every model you run — absence is the signal that you forgot one.

cachedTokens is a subset of input

usage.cachedTokens counts the input tokens served from the provider’s prompt cache; they are already included in usage.input. computeCost prices them at cachedInput (falling back to the full input rate when the provider publishes no cache rate) and the rest of input at full rate. Don’t double-count by adding cachedTokens to input — read it only to see how much the cache saved.

Related

• Pick a provider — ModelPricing, the token channels, and capability flags.
• Cost tracking — the per-session dashboard view of the same data.
• Run agent — the usage / report envelope execute() returns.