Grok models support a "thinking" mode that allows them to perform step-by-step reasoning before providing a final answer. This is useful for tasks requiring transparent logic, like mathematical proofs, intricate code debugging, or multi-step agent planning.
Grok guidance
- The reasoning token count is outputted in the
reasoning_tokensfield, separated from thecompletion_tokens. - The response doesn't have the
reasoning_contentfield. All response texts are outputted in thecontentfield. - Grok reasoning models don't support
reasoning_effort.
Example request:
curl
-X
POST
\
-H
"Authorization: Bearer
$(
gcloud
auth
print-access-token )
"
\
-H
"Content-Type: application/json"
\
https://aiplatform.googleapis.com/v1/projects/test-project/locations/global/endpoints/openapi/chat/completions
-d
'{
"model": "xai/grok-4.1-fast-reasoning",
"messages": [{
"role": "user",
"content": "Who are you?"
}],
}'
Example response:
{
"choices"
:
[
{
"finish_reason"
:
"stop"
,
"index"
:
0
,
"logprobs"
:
null
,
"message"
:
{
"content"
:
"I am Grok, an AI assistant built by xAI..."
,
"role"
:
"assistant"
}
}
],
"created"
:
1775523905
,
"id"
:
"knTMaJC0EJfM5OMP7I3xkAk"
,
"model"
:
"xai/grok-4.1-fast-reasoning"
,
"object"
:
"chat.completion"
,
"system_fingerprint"
:
"fp_39c5j0a324"
,
"usage"
:{
"completion_tokens"
:
50
,
"completion_tokens_details"
:{
"accepted_prediction_tokens"
:
0
,
"audio_tokens"
:
0
,
"reasoning_tokens"
:
124
,
"rejected_prediction_tokens"
:
0
},
"cost_in_usd_ticks"
:
0
,
"num_sources_used"
:
0
,
"prompt_tokens"
:
663
,
"prompt_tokens_details"
:{
"audio_tokens"
:
0
,
"cached_tokens"
:
654
,
"image_tokens"
:
0
,
"text_tokens"
:
663
},
"total_tokens"
:
837
}
}
What's next
- Learn about Function calling .
- Learn about Structured output .
