One gateway, three providers: request shapes compared

August 1, 2026 · 6 min read

LLM-ops platform

Your app sends one POST /gateway/chat/completions request with an OpenAI-shaped body; changing providers is a one-field edit to model. What that field hides is more interesting: underneath, Anthropic, OpenAI, and Gemini each expect a genuinely different request — a different auth header, different sampling-param nesting, and a different key for the same tool definition. We took one fixed request — a system message, a user message, and one tool definition — and traced it through all three adapters that ship today to show exactly what each one rewrites and why the gateway needs to do it at all.

The one request every path starts from

{
  "model": "demo-model",
  "messages": [
    { "role": "system", "content": "You are a concise, friendly support agent." },
    { "role": "user", "content": "My order has not arrived yet — can you look it up?" }
  ],
  "temperature": 0.2,
  "max_tokens": 200,
  "tools": [
    {
      "type": "function",
      "function": {
        "name": "lookup_order",
        "description": "Look up an order by its id.",
        "parameters": { "type": "object", "properties": { "orderId": { "type": "string" } }, "required": ["orderId"] }
      }
    }
  ],
  "tool_choice": "auto"
}

That's the whole client-facing contract, regardless of which model answers it. Below is what each adapter actually builds from it, captured by running the gateway's own adapter code directly against this exact request.

Anthropic: system hoisted out, tools renamed, no `Authorization` header

POST https://api.anthropic.com/v1/messages
x-api-key: <key>
anthropic-version: 2023-06-01

{
  "model": "demo-model",
  "messages": [
    { "role": "user", "content": "My order has not arrived yet — can you look it up?" }
  ],
  "max_tokens": 200,
  "system": "You are a concise, friendly support agent.",
  "temperature": 0.2,
  "tools": [
    { "name": "lookup_order", "description": "Look up an order by its id.", "input_schema": { "type": "object", "properties": { "orderId": { "type": "string" } }, "required": ["orderId"] } }
  ],
  "tool_choice": { "type": "auto" }
}

The system message never appears in messages at all — it's pulled out and joined into a top-level system string.
max_tokens is required; the adapter defaults it to 1024 if a caller omits it, since Anthropic's API rejects a request without one.
Auth is x-api-key, not Authorization: Bearer — plus a mandatory anthropic-version header pinning the API version.
A tool's parameters becomes input_schema — same JSON Schema, different key.

A live call through the gateway to an Anthropic-backed model with this same system and user message came back through the same normalization path as the other two:

{"id":"msg_011CdbQLdD7d1TBhx2cp7izz","model":"claude-haiku-4-5-20251001","object":"chat.completion","created":1785558766,"choices":[{"index":0,"message":{"role":"assistant","content":"I'd be happy to help you look up your order! Could you please provide me with your order ID? You can usually find this in your confirmation email or account."},"finish_reason":"stop"}],"usage":{"prompt_tokens":584,"completion_tokens":37,"total_tokens":621}}

Gemini: model in the URL, messages become `contents`, sampling params nest

POST https://generativelanguage.googleapis.com/v1beta/models/demo-model:generateContent
x-goog-api-key: <key>

{
  "contents": [
    { "role": "user", "parts": [{ "text": "My order has not arrived yet — can you look it up?" }] }
  ],
  "systemInstruction": { "parts": [{ "text": "You are a concise, friendly support agent." }] },
  "generationConfig": { "temperature": 0.2, "maxOutputTokens": 200 },
  "tools": [
    { "functionDeclarations": [{ "name": "lookup_order", "description": "Look up an order by its id.", "parameters": { "type": "object", "properties": { "orderId": { "type": "string" } }, "required": ["orderId"] } }] }
  ],
  "toolConfig": { "functionCallingConfig": { "mode": "AUTO" } }
}

The model name is in the URL path, not the request body at all.
messages becomes contents, and assistant becomes model — Gemini has no assistant role.
temperature and max_tokens don't pass through directly; they nest under generationConfig (and max_tokens is renamed maxOutputTokens).
Tools get wrapped in an extra array: tools: [{ functionDeclarations: [...] }] — one object holding all your functions, not one array entry per function.
Auth is x-goog-api-key — a third distinct header name, and no version header.

A live call through the gateway to a Gemini-backed model with this same system and user message came back through the same normalization path as the other two:

{"id":"chatcmpl-0c393a95-d15d-4446-8f06-f63f301462fd","model":"gemini-flash-latest","object":"chat.completion","created":1785558362,"choices":[{"index":0,"message":{"role":"assistant","content":"I'd be happy to help check on your order! Could you please provide your order ID?"},"finish_reason":"stop"}],"usage":{"prompt_tokens":70,"completion_tokens":20,"total_tokens":173}}

OpenAI: the gateway's own shape, close to a passthrough

POST https://api.openai.com/v1/chat/completions
Authorization: Bearer <key>

{
  "model": "demo-model",
  "messages": [
    { "role": "system", "content": "You are a concise, friendly support agent." },
    { "role": "user", "content": "My order has not arrived yet — can you look it up?" }
  ],
  "temperature": 0.2,
  "max_tokens": 200,
  "tools": [
    { "type": "function", "function": { "name": "lookup_order", "description": "Look up an order by its id.", "parameters": { "type": "object", "properties": { "orderId": { "type": "string" } }, "required": ["orderId"] } } }
  ],
  "tool_choice": "auto",
  "stream": false
}

Because the gateway's own client-facing API already mirrors OpenAI's Chat Completions shape, this adapter does almost nothing: messages, tools, and tool_choice pass straight through unchanged, Authorization: Bearer is the standard header, and the only addition is an explicit stream: false. A live call through the gateway to gpt-4o-mini with this same system and user message came back exactly as you'd expect from that passthrough:

{"id":"gen-1785557584-vX8IgUyDLe4W8wtJx8bX","model":"openai/gpt-4o-mini","object":"chat.completion","created":1785557584,"choices":[{"index":0,"message":{"role":"assistant","content":"I'm sorry to hear your order hasn't arrived yet! Unfortunately, I can't look up specific orders. However, I recommend checking your order confirmation email for tracking information or contacting customer support directly for assistance. Let me know if you need help with anything else!"},"finish_reason":"stop"}],"usage":{"prompt_tokens":33,"completion_tokens":54,"total_tokens":87}}

Side by side

	Anthropic	Gemini	OpenAI
Model name	In body (`model`)	In URL path	In body (`model`)
System message	Hoisted to top-level `system` string	Hoisted to `systemInstruction.parts`	Stays inline in `messages`
`assistant` role	Unchanged	Renamed to `model`	Unchanged
Sampling params	Inline (`temperature`, `max_tokens`)	Nested under `generationConfig`, `max_tokens`→`maxOutputTokens`	Inline, unchanged
Tool schema field	`input_schema`	`parameters`, wrapped in `functionDeclarations`	`parameters`, unchanged
Auth header	`x-api-key` + `anthropic-version`	`x-goog-api-key`	`Authorization: Bearer`
`max_tokens`	Required — defaults to 1024 if omitted	Optional	Optional

Whichever model you point at, your code keeps sending the one shape at the top of this post — the row you never have to write yourself is whichever one your chosen provider needs.

What this means for you

Switching models is a model field change, not a rewrite — the translation in every row above happens inside the gateway, not in your code.
Anthropic's required max_tokens is the one place a request could behave differently across providers if you rely on omitting it — set it explicitly if you plan to swap models later.
Tool definitions travel as one JSON Schema, no matter which provider answers — you write parameters once and never touch input_schema or functionDeclarations yourself.

note

Want to reproduce this? Every response above is the real thing — one live call per provider through the running gateway, to an Anthropic-backed, a Gemini-backed, and an OpenAI-backed model, all given the exact same system and user message (only the registered model name differed from the "demo-model" placeholder shown in the request bodies above). The request shapes themselves come from running the gateway's real, unmodified adapter classes directly against the one request at the top of this post.

Want to try this yourself? The Quickstart gets you from sign-up to a traced, gateway-routed call in about ten minutes.

The one request every path starts from​

Anthropic: system hoisted out, tools renamed, no Authorization header​

Gemini: model in the URL, messages become contents, sampling params nest​

OpenAI: the gateway's own shape, close to a passthrough​

Side by side​

What this means for you​