Hi Vercel team, could the AI Gateway Batch team confirm whether a supported Batch route can preserve this Responses profile on our existing Gateway account? Our current endpoint is https://ai-gateway.vercel.sh/v1/responses.

1
{
2
  "model": "openai/gpt-6.1-sol",
3
  "background": false,
4
  "max_output_tokens": 32000,
5
  "parallel_tool_calls": false,
6
  "providerOptions": {
7
    "gateway": {
8
      "only": [
9
        "openai"
10
      ]
11
    }
12
  },
13
  "reasoning": {
14
    "effort": "xhigh",
15
    "mode": "standard"
16
  },
17
  "service_tier": "default",
18
  "store": false,
19
  "stream": true,
20
  "text": {
21
    "format": {
22
      "type": "json_object"
23
    }
24
  },
25
  "tool_choice": "none",
26
  "tools": [],
27
  "truncation": "disabled"
28
}

The complete instructions and input must remain unchanged. Only the separate Batch submission may change stream to false; the model ID, OpenAI-only routing and account must stay the same.

Could you clarify:

1. Supported route and request fidelity: Does Gateway support OpenAI Files/Batches or a raw Responses option in SDK Batch? Please provide the API/schema, required versions and any access requirements. Our review of the published Responses serializer and tool preparation found omitted background:false, dropped tools:[]/tool_choice:"none", and added encrypted-reasoning include. Can a supported route avoid those changes?

2. Original receipts: Can we retrieve original per-item request/response bodies and raw result records, beyond normalized SDK results? We need correlation with our item IDs and Gateway/provider batch, request and response IDs where available; effective model/provider; terminal status including partial failures; and complete usage, including cache and reasoning tokens. Which fields are guaranteed, and for how long are they retained?

3. Billing and recovery: What authoritative original-account evidence supports independent per-item charge reconciliation, including currency, discounts, cache usage, fees and partial failures? If billing is batch-level, how can each item be reconciled without estimates? What idempotency, retry/fallback and interrupted-submission recovery guarantees prevent duplicate work or charges?

Documentation or sanitized request/result examples would help. If these requirements are unsupported, please identify the gaps.