## What happened

For streaming responses, custom callbacks that read `kwargs["response_cost"]` in `async_log_success_event` intermittently (for us: *usually*) observe `None` even though the call succeeded, litellm computed a real cost, and token usage is fully populated. Any cost/spend tracking built on that key silently records $0.

In our proxy deployment (`litellm[proxy]==1.83.14`, Python 3.13), **80.7% of streaming success rows (245,562 of 304,148)** recorded a zero cost with real token counts. The rate was load-independent but strongly model-correlated (~93–97% for gpt-5.x streams, ~0% for Claude streams), consistent with a scheduling race rather than a data problem.

## Root cause

Line references are into `litellm==1.83.14`.

1. **Both success handlers fire concurrently for a finished stream.** `litellm_core_utils/streaming_handler.py:2209-2224`: when a stream completes, `async_success_handler` is scheduled as an asyncio task and `success_handler` is simultaneously submitted to the thread-pool `executor`. This happens even when no sync callbacks are registered.

2. **Both handlers mutate the same dict.** Each handler reads and writes `self.model_call_details`…