LLM API Response

Updated: 2025-11-25

When you make an API call, you receive a structured response object named ChatCompletion. Let us break down what each part means using the response from the "Hello World" example in src/llm_api_access.py:

ChatCompletion(
  id='chatcmpl-CfaRxgTbF2CfDMo63LJwICeQBjODl',
  object='chat.completion',
  created=1764027597,
  model='gpt-5-nano-2025-08-07',
  service_tier='default',
  system_fingerprint=None,
  choices=[
    Choice(
      index=0,
      finish_reason='stop',
      logprobs=None,
      message=ChatCompletionMessage(
        content='Hello World',
        role='assistant',
        refusal=None,
        annotations=[],
        audio=None,
        tool_calls=None))],
  usage=CompletionUsage(
    prompt_tokens=11, 
    completion_tokens=203, 
    total_tokens=214, 
    prompt_tokens_details=PromptTokensDetails(
      audio_tokens=0, 
      cached_tokens=0),
    completion_tokens_details=CompletionTokensDetails(
      accepted_prediction_tokens=0, 
      audio_tokens=0, 
      reasoning_tokens=192, 
      rejected_prediction_tokens=0)))

Top-Level Fields

Here are the top-level fields in the response:

Field	Description	Example
id	Unique identifier for this API request	'chatcmpl-CfaRxgTbF2CfDMo63LJwICeQBjODl'
object	Type of object returned	'chat.completion'
created	Unix timestamp when the response was created	1764027597
model	The specific model version that generated the response	'gpt-5-nano-2025-08-07'
service_tier	The service tier used for the request	'default'
system_fingerprint	Backend configuration identifier	'fp_12345abc' or None
choices	Contains the model's response(s)	Array of Choice
usage	Tracks how many tokens were used	CompletionUsage

The choices Array

This field contains an array of possible responses. In most cases, there is only one choice (index 0):

Field	Description	Example
index	Position in the choices array	0
finish_reason	Why the model stopped generating • stop: natural completion • length: hit max_tokens limit • content_filter: content was filtered	'stop
logprobs	Log probabilities for tokens	None (unless requested)
message	The response and metadata from the model	ChatCompletionMessage

The message Object

This field contains the actual response from the model:

Field	Description	Example
content	The actual text response from the model	'Hello World'
role	Who generated this message	'assistant', 'user', 'system'
refusal	Message if model refused to respond	None or 'I cannot help with that'
annotations	Citations or references (RAG/retrieval)	[] (usually empty)
audio	Audio response data	None or audio object
tool_calls	Tools/functions the model wants to use	None or list of tool calls

The usage Object

This is critical for understanding API costs. Every API call consumes tokens, which you pay for:

Field	Description	Example
prompt_tokens	Number of tokens in your input (prompt)	11
completion_tokens	Number of tokens in the model's response	203
total_tokens	Total tokens used (prompt + completion)	214
prompt_tokens_details	Breakdown of tokens used in a prompt	PromptTokensDetails
completion_tokens_details	Breakdown of tokens used in a completion	completion_tokens_details

Tokens are pieces of words. Roughly:

• 1 token ≈ 4 characters in English

• 1 token ≈ ¾ of a word

• 100 tokens ≈ 75 words

You pay separately for prompt_token (input) and completion_tokens (output). For gpt-5-nano (as of 2025-12-01):

• Input: $0.05 per 1M tokens ($0.00000005 per token)

• Output: $0.4 per 1M tokens ($0.0000004 per token)

Thus, the total cost is:

• Input: 11 tokens x $0.00000005 = $0.00000055

• Output: 203 tokens x $0.0000004 = $0.000081

• Total: $0.00000055 + $0.00000812 = $0.00008175