diff --git a/src/lighteval/models/endpoints/litellm_model.py b/src/lighteval/models/endpoints/litellm_model.py
index b69764ebc..20fb87d04 100644
--- a/src/lighteval/models/endpoints/litellm_model.py
+++ b/src/lighteval/models/endpoints/litellm_model.py
@@ -294,10 +294,14 @@ def greedy_until(
 
             for response, context in zip(responses, contexts):
                 result: list[str] = [choice.message.content for choice in response.choices]
+                reasonings: list[str | None] = [
+                    getattr(choice.message, "reasoning_content", None) for choice in response.choices
+                ]
 
                 cur_response = ModelResponse(
                     # In empty responses, the model should return an empty string instead of None
                     text=result if result[0] else [""],
+                    reasonings=reasonings,
                     input=context,
                 )
                 results.append(cur_response)
diff --git a/src/lighteval/models/model_output.py b/src/lighteval/models/model_output.py
index 3663d88bd..fcb96099e 100644
--- a/src/lighteval/models/model_output.py
+++ b/src/lighteval/models/model_output.py
@@ -127,6 +127,7 @@ class ModelResponse:
     text: list[str] = field(default_factory=list)  # The text of the response
     output_tokens: list[list[int]] = field(default_factory=list)  # Model generations
     text_post_processed: list[str] | None = None  # The text of the response postprocessed
+    reasonings: list[str | None] = field(default_factory=list)  # The reasoning content of the response
 
     # Model logprob outputs
     logprobs: list[float] = field(default_factory=list)  # Log probabilities of the response