Build with AI that refuses to be weaponized.

Request Parameters:

Messages with the developer role are treated as system instructions.

Non-Streaming Response (Thaura Format):

{
  "id": "chatcmpl-...",
  "object": "chat.completion",
  "created": 1234567890,
  "model": "thaura",
  "choices": [{
    "index": 0,
    "message": {
      "role": "assistant",
      "content": "The AI response text..."
    },
    "finish_reason": "stop"
  }],
  "usage": {
    "prompt_tokens": 52,
    "completion_tokens": 18,
    "total_tokens": 70
  }
}

Streaming Response (Thaura SSE Format):

data: {"id":"chatcmpl-...","object":"chat.completion.chunk","created":1234567890,"model":"thaura","choices":[{"index":0,"delta":{"content":"Response "},"finish_reason":null}]}
data: {"id":"chatcmpl-...","object":"chat.completion.chunk","created":1234567890,"model":"thaura","choices":[{"index":0,"delta":{"content":"text..."},"finish_reason":null}]}
data: {"id":"chatcmpl-...","object":"chat.completion.chunk","created":1234567890,"model":"thaura","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}
data: [DONE]

Streaming with Token Usage:

Token usage comes in its own chunk just before data: [DONE], with an empty choices array (the OpenAI shape), so read choices[0] only when choices isn't empty. It's on by default; send stream_options: {"include_usage": false} to turn it off.

usage follows OpenAI's shape: prompt_tokens_details.cached_tokens is the part of the prompt read from cache. completion_tokens counts only the visible answer, which is what you pay for, so completion_tokens_details.reasoning_tokens is always 0. The model's hidden thinking is never returned, counted or billed.

data: {"id":"chatcmpl-...","object":"chat.completion.chunk","created":1234567890,"model":"thaura","choices":[{"index":0,"delta":{"content":"text"},"finish_reason":null}]}
...
data: {"id":"chatcmpl-...","object":"chat.completion.chunk","created":1234567890,"model":"thaura","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}
data: {"id":"chatcmpl-...","object":"chat.completion.chunk","created":1234567890,"model":"thaura","choices":[],"usage":{"prompt_tokens":52,"completion_tokens":18,"total_tokens":70,"prompt_tokens_details":{"cached_tokens":0,"audio_tokens":0},"completion_tokens_details":{"reasoning_tokens":0,"audio_tokens":0,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}}}
data: [DONE]

Error Response:

{
  "error": {
    "message": "n greater than 1 is not supported: Thaura returns one choice per request. ...",
    "type": "invalid_request_error",
    "param": "n",
    "code": "unsupported_value"
  }
}

Status codes:

• 400invalid_messages, unsupported_value, invalid_json_schema, invalid_type, invalid_value, decimal_below_min_value, decimal_above_max_value, integer_below_min_value, integer_above_max_value, array_above_max_length, missing_required_parameter — The request is malformed or asks for something unsupported: a bad messages array, a value outside OpenAI's range, n greater than 1, a tool_choice naming a tool that isn't in tools, a custom tool, or an invalid JSON schema. error.param names the field.
• 400context_length_exceeded — The prompt is too long for the 128,000-token window, or carries more images than one request can.
• 400too_many_attachments, invalid_attachments, invalid_image — More than 20 attachments in a message, an attachment that can't be read, or an image that can't be decoded.
• 401unauthorized — The API key is missing or invalid.
• 402insufficient_balance — Your balance can't cover this request's maximum cost. Add funds, or lower max_tokens.
• 403key_scope — This key can't be used on this endpoint.
• 404model_not_found — Unknown model: chat completions and /v1/models/{model} answer a model other than "thaura" with 404 model_not_found, and transcription answers an unknown transcription model the same way.
• 413file_too_large — The request body, or the audio file, is over 50 MB.
• 429rate_limit_exceeded, concurrency_limit_exceeded — More than 60 requests a minute, or 8 at once, on this key. Retry after a moment.
• 500internal_error, transcription_failed — Something failed on our side. Retry the request.
• 503service_unavailable, upstream_rate_limited — The model is temporarily unavailable or busy. Retry after retry_after_seconds.
• 504upstream_timeout — The model took too long to answer. Retry the request.

Thaura answers an empty balance with 402 insufficient_balance (OpenAI uses 429 insufficient_quota). The SDK does not retry a 402.

Once a stream has started, an error arrives as a data: {"error": …} chunk with the same type and code, followed by data: [DONE].