## Summary

For deployments using LiteLLM's generic OpenAI-compatible passthrough provider (`litellm_params.model = "openai/<upstream-model>"` with a custom `api_base`), successful completions are logged to `LiteLLM_SpendLogs` with **inconsistent deployment identity**: some rows correctly carry the deployment's `model_id` and full `model` string (with the `openai/` prefix intact), while other rows for calls to the **exact same deployment** have `model_id` blank/empty and `model` stripped of the `openai/` prefix.

Cost computation itself appears correct in both cases (verified — see below). The bug is specifically that the spend-log row's deployment identity is lost for a subset of calls, which makes any dashboard/analysis that groups or filters by `model_id` (the standard way to view spend by model) silently miss real, correctly-priced spend.

## Environment

- `litellm==1.95.0` (`ghcr.io/berriai/litellm:main-latest`) - Proxy mode, `STORE_MODEL_IN_DB=True`, Postgres backend - Reproduced on two independent deployments, so this isn't specific to DB-stored vs YAML-defined models: 1. A DB-stored model (`POST /model/new`) with `litellm_params.model: "openai/anthropic/claude-sonnet-5"`…