Title: [AI Gateway] Missing usage.prompt_tokens_details.cached_tokens in /v1/chat/completions

Description

Vercel AI Gateway routes requests to models with automatic or explicit prompt caching (e.g. zai/glm-5.3, moonshotai/kimi-k3, deepseek/deepseek-v4-flash). While the Vercel dashboard and GET /v1/report endpoint accurately record cached_input_tokens, the OpenAI-compatible runtime endpoint (POST https://ai-gateway.vercel.sh/v1/chat/completions) drops the standard OpenAI prompt_tokens_details field.

Impact

Any tool, proxy, or SDK using the OpenAI compatibility standard (including LiteLLM, Langfuse, and corporate billing tools) receives prompt_tokens_details: null. This causes 5x+ phantom cost over-reporting in downstream accounting tools because callers cannot see which prompt tokens received cache-read discounts.

Affected Endpoints

1. POST https://ai-gateway.vercel.sh/v1/chat/completions (Non-streaming)
2. POST https://ai-gateway.vercel.sh/v1/chat/completions (Streaming final chunk with stream_options: {"include_usage": true})
3. POST https://ai-gateway.vercel.sh/v1/responses (OpenResponses endpoint)

Expected Behavior

POST /v1/chat/completions should adhere to the OpenAI usage schema:

1
"usage": {
2
  "prompt_tokens": 12847,
3
  "completion_tokens": 308,
4
  "total_tokens": 13155,
5
  "prompt_tokens_details": {
6
    "cached_tokens": 10920
7
  },
8
  "completion_tokens_details": {
9
    "reasoning_tokens": 0
10
  }
11
}
12
Reference
13
- Vercel Anthropic Messages endpoint (POST /v1/messages) already correctly exposes cache_read_input_tokens.
14
- Vercel Custom Reporting (GET /v1/report) already tracks cached_input_tokens.
15
- Please propagate this existing metadata into the usage.prompt_tokens_details object of /v1/chat/completions.