Responses

#Use it when

• You want a unified input/output shape
• Your app already uses OpenAI Responses
• You want schema-driven text output and reasoning controls

#Code examples

1curl -X POST https://api.navy/v1/responses \
2  -H "Authorization: Bearer sk-navy-YOUR_KEY" \
3  -H "Content-Type: application/json" \
4  -d '{
5    "model": "gpt-5",
6    "input": "Summarize the changelog in four bullets."
7  }'

#Parameters

• model (string, required) — Model ID to use
• input (string or array, required) — Prompt or array of input items
• stream (boolean, optional) — Stream response as SSE
• instructions (string, optional) — System-level instructions
• tools (array, optional) — Function tool definitions
• tool_choice (string or object, optional) — "auto", "none", "required"
• temperature (number, optional) — 0.0–2.0
• top_p (number, optional) — Nucleus sampling threshold
• max_output_tokens (integer, optional) — Maximum tokens to generate
• reasoning (object, optional) — Reasoning controls for thinking models (e.g. { "effort": "high" })
• text (object, optional) — Text format controls including JSON schema

#Streaming

Set "stream": true to receive an SSE stream of typed events like response.output_text.delta, response.output_item.added, and response.completed.

1from openai import OpenAI
2
3client = OpenAI(api_key="sk-navy-YOUR_KEY", base_url="https://api.navy/v1")
4
5with client.responses.stream(
6    model="gpt-5",
7    input="Stream a one-paragraph product update."
8) as stream:
9    for event in stream:
10        if event.type == "response.output_text.delta":
11            print(event.delta, end="", flush=True)

#Built-in tools

NavyAI accepts user-defined function tools on every model. Native OpenAI-owned models also support the OpenAI built-in tools (image_generation, code_interpreter, file_search, computer_use_preview, mcp); xAI models additionally support web_search_preview. When you call a non-native model with one of these built-in tools, NavyAI strips it from the request so the call still succeeds against the underlying provider.

The web_search / web_search_preview tool is always available — when used, it adds a small fixed surcharge to your token usage for that request.

#Reasoning controls

Thinking models accept a reasoning object with an effort field ("minimal", "low", "medium", "high", "xhigh") to bias the depth of internal reasoning before the final answer.

1{
2  "model": "gpt-5",
3  "input": "Plan a launch checklist.",
4  "reasoning": { "effort": "high" }
5}

#Notes

• OpenAI-owned models are forwarded to the native OpenAI Responses API for full fidelity
• xAI-owned models are forwarded to xAI's native Responses-compatible endpoint
• Non-OpenAI / non-xAI models are adapted under the hood and returned in a compatible response shape
• The following fields are silently stripped before forwarding: n, service_tier, background, user, safety_identifier, prompt_cache_key, prompt_cache_retention, metadata, context_management, plugins, provider