### What happened?

**Env:** LiteLLM Proxy **v1.92.0**, `STORE_MODEL_IN_DB=True`. An OpenAI `sora-2` model is registered with a **per-model** `litellm_params.api_key` (there is **no** global `OPENAI_API_KEY` in the environment).

- `POST /v1/videos` (create) **works** — it correctly uses the per-model key by deployment. - `GET /v1/videos/{id}` (status) and `GET /v1/videos/{id}/content` (download) **fail** with: - `invalid_api_key` → - `APIConnectionError: 3 validation errors for VideoObject (id / object / status — Field required)`

**Root cause:** the `video_id` returned by create decodes to:

``` litellm:custom_llm_provider:openai;model_id:;video_id:video_... ```

i.e. **`model_id` is empty**. On status/content the proxy has no `model_id` to resolve the deployment (and thus the per-model key) from, so it falls back to a provider default key that doesn't exist → `invalid_api_key`.

Scope: **video only** (any async job + retrieve). Text / image / audio are single-shot and work fine with per-model keys.

### Analysis (against `main`)

- `video_status` and `video_content` (`litellm/proxy/video_endpoints/endpoints.py`) set `data["model"]` **exclusively** from `resolve_model_name_fr…