fix: avoid counting video urls as text

litellm/litellm_core_utils/token_counter.py

@@ -724,12 +724,7 @@ def _count_content_list(
                     image_url, use_default_image_token_count
                 )
             elif c["type"] == "video_url":
-                video_url = c.get("video_url", "")
-                if isinstance(video_url, dict):
-                    video_url = video_url.get("url")
-                if video_url is None:
-                    video_url = ""
-                num_tokens += DEFAULT_IMAGE_TOKEN_COUNT + count_function(str(video_url))
+                num_tokens += DEFAULT_IMAGE_TOKEN_COUNT
             elif c["type"] in ("tool_use", "tool_result"):
                 num_tokens += _count_anthropic_content(
                     c,

tests/test_litellm/litellm_core_utils/test_token_counter.py

@@ -948,6 +948,21 @@ def test_token_counter_with_video_url():
         tokens_str > DEFAULT_IMAGE_TOKEN_COUNT
     ), f"Expected default video token budget, got {tokens_str}"
 
+    messages_base64_str = [
+        {
+            "role": "user",
+            "content": [
+                {
+                    "type": "video_url",
+                    "video_url": "data:video/mp4;base64," + ("A" * 4000),
+                }
+            ],
+        }
+    ]
+    assert (
+        token_counter(model="gpt-4o", messages=messages_base64_str) == tokens_str
+    )
+
     messages_empty_url = [
         {
             "role": "user",