ReferenceReference

Usage fields

The exact shape of usage in each protocol.

By the end: You can read token counts from any response format.

The shape of usage depends on the protocol, not on the model vendor.

OpenAI — Chat Completions

"usage": {
  "prompt_tokens": 28,
  "completion_tokens": 54,
  "total_tokens": 82,
  "prompt_tokens_details": { "cached_tokens": 0 }
}

OpenAI — Responses

"usage": {
  "input_tokens": 28,
  "output_tokens": 54,
  "total_tokens": 82
}

Anthropic — Messages

"usage": {
  "input_tokens": 28,
  "output_tokens": 54,
  "cache_creation_input_tokens": 0,
  "cache_read_input_tokens": 0
}

Anthropic's input_tokens excludes cache reads and writes. The gateway adds them back so the billed input is the true total, and prices the cached portion at the cached rate.

Gemini

"usageMetadata": {
  "promptTokenCount": 28,
  "candidatesTokenCount": 54,
  "totalTokenCount": 82,
  "cachedContentTokenCount": 0
}

Whatever the field names, the deduction is computed as described in usage fields and metering.