fix: tokenize text-source documents instead of using flat fallback

litellm/litellm_core_utils/token_counter.py

@@ -681,17 +681,32 @@ def _count_anthropic_document(
     Anthropic shape:
         {"type": "document", "source": {...}, "title": ..., "context": ..., "citations": ...}
 
-    We count the text-bearing fields ("title", "context") with the model
-    tokenizer and add DEFAULT_IMAGE_TOKEN_COUNT for the opaque source payload
-    (typically a PDF — we cannot dimension it). Skipped: "type",
-    "cache_control", "citations" (configuration metadata, not prompt content).
+    Source variants (per the Anthropic Citations API):
+        - {"type": "text", "media_type": "text/plain", "data": "<inline text>"}
+        - {"type": "base64", "media_type": ..., "data": ...}
+        - {"type": "url", "url": ...}
+        - {"type": "file", "file_id": ...}
+
+    For text sources we tokenize source["data"] with the model tokenizer
+    (it's plain prompt content). For base64 / url / file sources the
+    payload is opaque (typically a PDF — we cannot dimension it) and we
+    add DEFAULT_IMAGE_TOKEN_COUNT as the fallback.
+
+    Title and context are always counted. Skipped: "type", "cache_control",
+    "citations" (configuration metadata, not prompt content).
     """
     tokens = 0
     for field in ("title", "context"):
         value = content.get(field)
         if isinstance(value, str) and value:
             tokens += count_function(value)
-    if content.get("source") is not None:
+    source = content.get("source")
+    if isinstance(source, dict) and source.get("type") == "text":
+        data = source.get("data")
+        if isinstance(data, str) and data:
+            tokens += count_function(data)
+    elif source is not None:
+        # base64 / url / file / unknown — opaque payload, use the fallback.
         tokens += DEFAULT_IMAGE_TOKEN_COUNT
     return tokens
 

tests/test_litellm/litellm_core_utils/test_token_counter.py

@@ -1130,6 +1130,40 @@ def test_token_counter_with_anthropic_document_file():
     assert tokens > 0, f"Expected positive token count, got {tokens}"
 
 
+def test_token_counter_with_anthropic_document_text():
+    """
+    Anthropic-native document block with source.type='text' tokenizes the
+    inline text in source['data'], not the flat DEFAULT_IMAGE_TOKEN_COUNT
+    fallback used for opaque (base64/url/file) payloads.
+    """
+    inline_text = "Lorem ipsum dolor sit amet, consectetur adipiscing elit. " * 200
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {
+                    "type": "document",
+                    "source": {
+                        "type": "text",
+                        "media_type": "text/plain",
+                        "data": inline_text,
+                    },
+                    "title": "Long inline doc",
+                }
+            ],
+        }
+    ]
+
+    tokens = token_counter(
+        model="anthropic/claude-sonnet-4-5-20250929", messages=messages
+    )
+    # The inline text alone is well over a thousand tokens — substantially
+    # more than the 250-token fallback used for opaque sources.
+    assert (
+        tokens > 500
+    ), f"Expected text-source document to tokenize source['data'], got {tokens}"
+
+
 def test_token_counter_with_image_inside_tool_result():
     """
     Image block nested inside tool_result.content is counted via recursion.