[Bug]: Response cost not reported for fine-tuned OpenAI models with custom pricing in config

[TeddyAmkie]

### What happened?

**Summary**  
When using fine-tuned OpenAI models and specifying `input_cost_per_token` / `output_cost_per_token` in the `model_info` section of the config YAML, LiteLLM does not report `response_cost` in logs or in the `x-litellm-response-cost` header. Instead, it either shows `0.0` or picks the cost from the `base_model` rather than the overridden custom pricing.

---

### Minimal Repro Setup

1. **Config (`config.yaml`)**

```yaml
model_list:
  - model_name: openai/ft:gpt-4.1-2025-04-14:*
    litellm_params:
      model: openai/ft:gpt-4.1-2025-04-14:*
      api_key: os.environ/OPENAI_DEFAULT_API_KEY
    model_info:
      id: openai-fine-tuned
      input_cost_per_token:  0.000003
      output_cost_per_token: 0.000012

litellm_settings:
  json_logs: true

2. Run LiteLLM proxy

litellm --config config.yaml

3. Send a test request

curl http://localhost:4000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer sk-<proxy-key>" \
  -d '{
    "model": "openai/ft:gpt-4.1-2025-04-14:my-org:example",
    "messages": [{"role": "user", "content": "Hello!"}]
  }' -i

---

Expected Behavior

• LiteLLM should calculate response cost using the configured input_cost_per_token / output_cost_per_token.
• response_cost should be present in logs.
• x-litellm-response-cost header should match the calculated value.

---

Actual Behavior

• response_cost is logged as 0.0 (or absent).
• No x-litellm-response-cost header is returned.
• If base_model is set, LiteLLM falls back to base model pricing instead of the custom values.
• Example:
  • Configured: 0.000003 * 26 + 0.000012 * 165 = 0.002058
  • Reported: 0.0025725 (base model pricing).

---

Additional Context

• Works correctly with non-fine-tuned models (e.g. bedrock/anthropic.claude-3-5-haiku-20241022-v1:0).
• @ishaan-jaff confirmed repro on Jul 30.

[wojciech-mikolajczyk-yohana]

What are the plans for fixing that?