diff --git a/litellm/llms/anthropic/chat/handler.py b/litellm/llms/anthropic/chat/handler.py
index 2fb29b32a61..9d027d6fb6f 100644
--- a/litellm/llms/anthropic/chat/handler.py
+++ b/litellm/llms/anthropic/chat/handler.py
@@ -820,7 +820,13 @@ class ModelResponseIterator:
                     "type"
                 ]
                 if content_block_start["content_block"]["type"] == "text":
-                    text = content_block_start["content_block"]["text"]
+                    # Anthropic Messages streaming spec: the initial text on a
+                    # content_block_start text block is conventionally "" and the
+                    # real content arrives via subsequent content_block_delta
+                    # chunks. Some Anthropic-compatible upstreams omit the field
+                    # entirely rather than sending "", so default to "" to avoid
+                    # aborting the stream with KeyError.
+                    text = content_block_start["content_block"].get("text", "")
                 elif (
                     content_block_start["content_block"]["type"] == "tool_use"
                     or content_block_start["content_block"]["type"] == "server_tool_use"
diff --git a/tests/llm_translation/test_anthropic_completion.py b/tests/llm_translation/test_anthropic_completion.py
index 371b27c5b21..cc3eebdf000 100644
--- a/tests/llm_translation/test_anthropic_completion.py
+++ b/tests/llm_translation/test_anthropic_completion.py
@@ -262,6 +262,25 @@ def test_anthropic_tool_streaming():
             assert tool_use["index"] == correct_tool_index
 
 
+def test_anthropic_content_block_start_text_missing_field():
+    """
+    Regression for #28067: some Anthropic-compatible upstreams omit the
+    optional `text` field on a content_block_start text block. The actual
+    content arrives via subsequent content_block_delta chunks, so the parser
+    must tolerate the missing field instead of raising KeyError and aborting
+    the stream.
+    """
+    response_iter = ModelResponseIterator([], False)
+    chunk = {
+        "type": "content_block_start",
+        "index": 0,
+        # 'text' field intentionally omitted (observed in the wild)
+        "content_block": {"type": "text"},
+    }
+    # Should not raise.
+    response_iter.chunk_parser(chunk=chunk)
+
+
 def test_process_anthropic_headers_empty():
     result = process_anthropic_headers({})
     assert result == {}, "Expected empty dictionary for no input"