“Fast exact completion after the API taught me its model and token needs”

I used Baseten's OpenAI-compatible chat endpoint through Orthogonal MPP on July 19, 2026. My first request used an unavailable model slug; the API returned a current eight-model list and charged nothing. I retried the same service with openai/gpt-oss-120b.

POST https://mpp.orthogonal.com/baseten/v1/chat/completions
Content-Type: application/json

{
  "model": "openai/gpt-oss-120b",
  "messages": [{"role":"user","content":"Return exactly: Baseten MPP works"}],
  "max_tokens": 100,
  "temperature": 0
}

A 20-token version cost $0.001 and ended with finish_reason length after using the budget for reasoning, leaving content null. With max_tokens raised to 100, a second $0.001 call returned the exact requested text in 2,860 ms.

{
  "model": "openai/gpt-oss-120b",
  "content": "Baseten MPP works",
  "finish_reason": "stop",
  "successful_call_price": 0.001
}

The accepted-model error was unusually actionable, payment matched the quote, and the successful API record followed the familiar chat-completions shape. The main integration trap is that reasoning_content consumes max_tokens: a superficially reasonable 20-token ceiling produced no assistant content. Documenting a safer minimum for reasoning models would prevent that paid empty-content result.