### Note on root cause (verified vs inferred)

**Verified by reading code and by a production traceback:** - `stream_chunk_builder` (`litellm/main.py:8342`) does `first_chunk_with_choices = next((c for c in chunks if c["choices"]), None)` and later `if len(chunk["choices"]) == 0` (`litellm/main.py:8357`) with no `isinstance(c, dict)` guard. If any element of `chunks` is `bytes`, `c["choices"]` raises `TypeError: byte indices must be integers or slices, not str`, and the function's own `except` (`litellm/main.py:8591-8598`) re-raises it as a generic `litellm.APIError(status_code=500, message="Error building chunks for logging/streaming usage calculation")`. - It is invoked inline at stream completion for logging/usage assembly from `litellm/litellm_core_utils/streaming_handler.py` (e.g. `stream_chunk_builder(chunks=self.chunks)` at `:2109`), which is why the exception surfaces to the client through `proxy_server.async_data_generator` rather than being swallowed by a background logging task. - The Anthropic-messages streaming logging paths deal in raw SSE bytes: `litellm/llms/anthropic/experimental_pass_through/messages/streaming_iterator.py:104-114` collects `collected_chunks: List[…