diff --git a/litellm/responses/litellm_completion_transformation/transformation.py b/litellm/responses/litellm_completion_transformation/transformation.py
index e8aacac9e67..b2382de3de5 100644
--- a/litellm/responses/litellm_completion_transformation/transformation.py
+++ b/litellm/responses/litellm_completion_transformation/transformation.py
@@ -2826,8 +2826,10 @@ class LiteLLMCompletionResponsesConfig:
             if hasattr(prompt_details, "text_tokens") and prompt_details.text_tokens is not None:
                 input_details_dict["text_tokens"] = prompt_details.text_tokens
 
-            if hasattr(prompt_details, "audio_tokens") and prompt_details.audio_tokens is not None:
-                input_details_dict["audio_tokens"] = prompt_details.audio_tokens
+            for modality in ("audio_tokens", "image_tokens", "video_tokens"):
+                value = getattr(prompt_details, modality, None)
+                if value is not None:
+                    input_details_dict[modality] = value
 
             cache_write_tokens = getattr(prompt_details, "cache_write_tokens", None) or getattr(
                 prompt_details, "cache_creation_tokens", None
diff --git a/litellm/responses/utils.py b/litellm/responses/utils.py
index 599e978df6a..09739b18cc4 100644
--- a/litellm/responses/utils.py
+++ b/litellm/responses/utils.py
@@ -1179,6 +1179,7 @@ class ResponseAPILoggingUtils:
                     audio_tokens=getattr(response_api_usage.input_tokens_details, "audio_tokens", None),
                     text_tokens=getattr(response_api_usage.input_tokens_details, "text_tokens", None),
                     image_tokens=getattr(response_api_usage.input_tokens_details, "image_tokens", None),
+                    video_tokens=getattr(response_api_usage.input_tokens_details, "video_tokens", None),
                     cache_write_tokens=getattr(response_api_usage.input_tokens_details, "cache_write_tokens", None),
                 )
         completion_tokens_details: CompletionTokensDetailsWrapper | None = None
diff --git a/litellm/types/llms/openai.py b/litellm/types/llms/openai.py
index dfafe27e0a1..ad7b3c068ad 100644
--- a/litellm/types/llms/openai.py
+++ b/litellm/types/llms/openai.py
@@ -1288,7 +1288,9 @@ class OutputTokensDetails(BaseLiteLLMOpenAIResponseObject):
 class InputTokensDetails(BaseLiteLLMOpenAIResponseObject):
     audio_tokens: int | None = None
     cached_tokens: int = 0
+    image_tokens: int | None = None
     text_tokens: int | None = None
+    video_tokens: int | None = None
 
     model_config = {"extra": "allow"}
 
diff --git a/tests/test_litellm/responses/litellm_completion_transformation/test_litellm_completion_responses.py b/tests/test_litellm/responses/litellm_completion_transformation/test_litellm_completion_responses.py
index 3c78bbf79d7..6ebeb9ca2af 100644
--- a/tests/test_litellm/responses/litellm_completion_transformation/test_litellm_completion_responses.py
+++ b/tests/test_litellm/responses/litellm_completion_transformation/test_litellm_completion_responses.py
@@ -2852,6 +2852,39 @@ class TestUsageTransformation:
         assert response_usage.input_tokens_details.cached_tokens == 100
         assert getattr(response_usage.input_tokens_details, "cache_write_tokens", None) == 800
 
+    def test_transform_usage_preserves_input_modality_tokens(self):
+        """Regression: the bridge dropped image and video input tokens.
+
+        Vertex reports prompt tokens split by modality, so a Live session that sends
+        camera frames arrives with image_tokens set. InputTokensDetails declared only
+        audio/cached/text, so those tokens were folded into text and lost their
+        attribution, and any per-modality rate could never apply to them.
+        """
+        usage = Usage(
+            prompt_tokens=300,
+            completion_tokens=10,
+            total_tokens=310,
+            prompt_tokens_details=PromptTokensDetailsWrapper(
+                text_tokens=20, audio_tokens=80, image_tokens=150, video_tokens=50, cached_tokens=0
+            ),
+            completion_tokens_details=CompletionTokensDetailsWrapper(text_tokens=10),
+        )
+
+        response_usage = LiteLLMCompletionResponsesConfig._transform_chat_completion_usage_to_responses_usage(
+            chat_completion_response=usage
+        )
+        details = response_usage.input_tokens_details
+        assert details is not None
+        assert getattr(details, "image_tokens", None) == 150
+        assert getattr(details, "video_tokens", None) == 50
+        assert getattr(details, "audio_tokens", None) == 80
+
+        from litellm.responses.utils import ResponseAPILoggingUtils
+
+        back = ResponseAPILoggingUtils._transform_response_api_usage_to_chat_usage(response_usage.model_dump())
+        assert back.prompt_tokens_details.image_tokens == 150
+        assert back.prompt_tokens_details.video_tokens == 50
+
     def test_transform_usage_with_reasoning_tokens_gemini(self):
         """Test that reasoning_tokens from Gemini are properly transformed to output_tokens_details"""
         # Setup: Simulate Gemini usage with thoughtsTokenCount