What should I use to track token usage and cost per request?

To track token usage and cost per request, use the API response usage fields and log them alongside your request metadata.

What to capture per request

Store:

• request_id or your own trace ID
• model name
• input tokens
• output tokens
• total tokens
• estimated cost
• latency
• user/session ID if relevant

Where to get token usage

Most LLM APIs return a usage object in the response, for example:

• prompt_tokens
• completion_tokens
• total_tokens

If you’re using a newer Responses-style API, the exact field names may vary, but the idea is the same: read the usage data from the response and persist it.

How to calculate cost

Use the model’s published pricing:

• cost = (input_tokens × input_rate) + (output_tokens × output_rate)

For example:

• 1,000 input tokens at $5 / 1M tokens
• 500 output tokens at $15 / 1M tokens

Cost:

• input: 1000 / 1,000,000 × 5 = $0.005
• output: 500 / 1,000,000 × 15 = $0.0075
• total: $0.0125

Good practice

• Centralize this in middleware/interceptors so every request is tracked
• Log raw usage first, compute cost later if pricing changes
• Separate usage by:
  • model
  • environment
  • endpoint/use case
  • tenant/customer

If you want a ready-made approach

Use:

• OpenTelemetry for traces/metrics
• a database or warehouse for usage records
• a dashboard tool like Grafana, Metabase, or Looker

Minimal schema example

{
  "timestamp": "2026-10-09T12:00:00Z",
  "request_id": "abc123",
  "model": "gpt-4.1",
  "input_tokens": 1200,
  "output_tokens": 300,
  "total_tokens": 1500,
  "estimated_cost_usd": 0.018,
  "latency_ms": 842
}

If you want, I can also give you:

1. a Python example,
2. a Node.js example, or
3. a database schema + logging middleware pattern.