Chat Completions

Request

POST https://apimaster.ai/v1/chat/completions

Headers

Header	Required	Notes
Authorization	Yes	Bearer YOUR_API_KEY
Content-Type	Yes	application/json

Body (common fields)

Field	Type	Notes
model	string	e.g. claude-sonnet-4-6
messages	array	{role, content}
max_tokens	number	Max completion tokens
temperature	number	Optional
stream	boolean	SSE streaming

Example

curl -s "https://apimaster.ai/v1/chat/completions" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.5",
    "messages": [
      {"role": "system", "content": "You are helpful."},
      {"role": "user", "content": "Say hi in one word."}
    ],
    "max_tokens": 64
  }'

Gemini thinking levels

On native Gemini channels, gemini-4-argon and gemini-3.8-flash accept top-level reasoning_effort: low, medium, or high. The gateway maps it to native generationConfig.thinkingConfig.thinkingLevel. Argon also maps reasoning_effort: "minimal" to low; Flash 3.8 does not support minimal. Omitting the level preserves upstream defaults; Google's documented default is medium.

{
  "model": "gemini-4-argon",
  "messages": [{"role": "user", "content": "Explain your answer briefly."}],
  "reasoning_effort": "high"
}

Alternatively, omit reasoning_effort and use these request-body fields:

{
  "extra_body": {
    "google": {
      "thinking_config": {"thinking_level": "high", "include_thoughts": true}
    }
  }
}

• Explicit thinking_level accepts only low, medium, or high. Unknown levels and unsupported effort values return HTTP 400.
• Do not combine reasoning_effort with thinking_level or thinking_budget, even when values agree. include_thoughts alone can accompany effort. Argon also rejects simultaneous level and budget.
• The Python OpenAI SDK merges its extra_body argument into the request body. Preserve the literal outer JSON field with extra_body={"extra_body": {"google": {"thinking_config": {"thinking_level": "high"}}}}.
• high requests the highest public reasoning level; token use remains dynamic. max_tokens caps thinking and answer tokens together, so a tiny limit is unsuitable for verifying levels.
• The gateway converts and forwards the setting. Provider enforcement requires channel-specific verification. HTTP 200, reasoning-token counts, and model self-reports alone do not establish the applied level. Responses do not report a verified applied level.

Sources: Google OpenAI compatibility and Gemini thinking levels. Other models and protocols follow their respective adapters.

Non-streaming response

{
  "id": "chatcmpl-...",
  "object": "chat.completion",
  "choices": [
    {
      "index": 0,
      "message": { "role": "assistant", "content": "..." },
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 10,
    "completion_tokens": 5,
    "total_tokens": 15
  }
}

Streaming

With "stream": true, the body is text/event-stream; lines are data: {json} until data: [DONE].

Status codes

Code	Meaning
400	Bad parameters
401	Auth failed
402 / 403	Quota or permission
429	Rate limited
502 / 503	Upstream error — retry