AI Agent Cost Tracking

Track cost at the event level, roll it up to an agent run, and divide by successful outcomes. A monthly provider invoice can tell you what you spent; it cannot tell you which workflow, tool, retry, or customer created that spend.

What to Track

Create one cost event for every billable or capacity-consuming action. Model calls are the obvious source, but production agents also pay for embeddings, vector storage, search APIs, browsers, code sandboxes, and retry loops.

A Cost Event Schema

Use immutable events rather than updating one running total. Events can be replayed when a provider changes pricing, and they preserve failed attempts that would otherwise disappear from success-only metrics.

{
  "trace_id": "tr_123",
  "run_id": "run_456",
  "step": 3,
  "event_type": "model_call",
  "provider": "example-provider",
  "model": "example-model",
  "input_tokens": 2840,
  "output_tokens": 412,
  "cached_input_tokens": 900,
  "cost_usd": 0.008976,
  "latency_ms": 1840,
  "success": true,
  "timestamp": "2026-08-07T10:30:00Z"
}

Store the raw usage fields and the calculated cost. Raw usage lets you recompute historical spend; the stored cost preserves what your system believed at execution time. Do not put prompts, secrets, or raw tool output in a cost event.

Cost Formulas That Matter

• Cost per run: the sum of model, tool, memory, and compute events sharing a run ID.
• Cost per successful task: total spend divided by successful outcomes, including the spend from failed runs.
• Retry tax: retry-event spend divided by total spend.
• Tool efficiency: successful tool results divided by tool cost and latency.
• Unit margin: revenue or internal value per task minus fully loaded agent cost.

Estimate model budgets before launch with the token budget calculatorand memory infrastructure with the memory cost calculator.

Instrument Every Step

Put telemetry around the provider adapter, not inside individual prompts. The wrapper should record success and failure, use the provider's returned usage when available, and attach the current trace and run IDs.

async function callModel(input: ModelInput, trace: TraceContext) {
  const startedAt = performance.now();
  try {
    const response = await provider.generate(input);
    await costEvents.write({
      ...trace,
      event_type: "model_call",
      provider: response.provider,
      model: response.model,
      input_tokens: response.usage.inputTokens,
      output_tokens: response.usage.outputTokens,
      cost_usd: price(response.usage, response.model),
      latency_ms: performance.now() - startedAt,
      success: true,
    });
    return response;
  } catch (error) {
    await costEvents.write({
      ...trace,
      event_type: "model_call",
      cost_usd: 0,
      latency_ms: performance.now() - startedAt,
      success: false,
    });
    throw error;
  }
}

Apply the same wrapper to tools and memory operations. If a request times out after the vendor accepted it, record the event with an unknown or estimated cost and reconcile it later instead of silently assigning zero.

Dashboards and Budget Alerts

Start with distributions, not only totals. Track p50, p95, and p99 cost per run by workflow and outcome. A stable monthly total can hide a small group of looping runs that create poor user experiences and future budget risk.

• Stop a run when its hard token or monetary cap is reached.
• Warn when p95 cost per successful task rises above its baseline.
• Alert when retries exceed a fixed share of daily spend.
• Compare prompt, model, and workflow versions before deployment.
• Reconcile event totals with provider invoices on a schedule.

Common Tracking Mistakes

Do not attribute all spend to the final successful run, ignore failed tool calls, or mix list price with discounted invoice price in one metric. Do not use tokens as a substitute for dollars when a workflow uses several models with different prices.

Cost is also not quality. A cheaper run that completes the wrong task is waste. Join cost events with task evaluations so optimization preserves completion rate and safety.

Frequently Asked Questions

Next Steps

Map events onto the complete perceive-reason-plan-act-feedback loop, then set budgets using the two calculators above.

ai agent cost tracking — return to the complete AI agent architecture guide.