GPT-6 Luna Decisions API

Zurelay charges $0.04 per 1M input tokens, against OpenAI's list price of $0.10, so you save 60%. Output is free, as at OpenAI. The model counts your input once plus about 110 to 160 tokens of its own for each question: a one-question check on a short text is about 160 tokens, two questions about 320, and a 2,600-token input with four questions about 3,200. Requests that fail cost nothing.

OpenAI's API for answering classification or scoring questions about some input, with many questions in one request. Instead of writing text, the model returns a typed answer for each question: a probability for a predicate, one of your values with probabilities for a choice, or a place on your scale with probabilities for a score. GPT-6 Luna Decisions is the model that answers.

Create a Zurelay API key, set the OpenAI SDK's base URL to https://api.zurelay.com/v1 and call client.decisions.create with model set to gpt-6-luna-decisions, an input and a list of questions. In Node.js that needs openai 7.30.0 or later. Or POST the same JSON to https://api.zurelay.com/v1/decisions with your key as a Bearer token. The docs at https://zurelay.com/docs/openai-decisions have examples in Python, Node.js and cURL.

Yes. On /v1/decisions the model ID gpt-6-luna works as well as gpt-6-luna-decisions, so code written from OpenAI's SDK docs, which name gpt-6-luna, runs once you change the base URL and the key. The reply's model field says gpt-6-luna-decisions. On the chat endpoints, gpt-6-luna is still GPT-6 Luna, the chat model.

A predicate asks how likely a statement about the input is true and returns a probability from 0 to 1, such as 1 for a clear refund request. A choice picks one of 2 to 255 values you list, such as billing, hardware or shipping, and returns the value, a confidence and the probability of every value. A score places the input on 2 to 10 levels you list, lowest first, such as very negative, negative and neutral, and returns the expected level as an index from 0, such as 0.18 for mostly the first level. One request can mix all three.

Yes. Put images in the input as input_image parts of a user message, as a base64 data URL or a public https link, in PNG, JPEG, WebP or GIF, with detail set to low, high, auto or original. A 1200 by 630 image adds about 750 tokens at high or auto and about 150 at low. Structured data goes in as input_json parts. It doesn't read PDFs, audio or video.

Any question can come back as {"type": "refusal"} instead of an answer, while the other questions in the request are still answered. In our tests, the input "Refund requested 38 hours ago." with the predicate "Page on-call?" came back refused. Check each answer's type before you read it, and ask again more clearly or send the case to a person.

Up to 1,050,000 tokens, the input and every question together: an input of 918K tokens worked in our tests. A request takes 1 to 64 questions, a choice 2 to 255 unique values and a score 2 to 10 levels. Zurelay checks a request before it reaches the model, so a malformed one is refused with a clear error and costs nothing.

In our tests answers came back in 0.3 to 1 second, in about a second with an image, and an input of about 900K tokens took about 10 seconds. The answers come back whole: the API doesn't stream.

System, developer and assistant messages, function and tool call items, streaming, and images inside a question's instructions: each is refused with a 400 that says why. Put context in a user message or in a question's instructions. prompt_cache_key and prompt_cache_options are accepted and ignored: caching is automatic, and every input token costs the same.

Yes. Requests to gpt-6-luna-decisions run on OpenAI's GPT-6 Luna Decisions, and Zurelay doesn't swap in a different model or a chat model asked to act like one. Zurelay is an independent service and is not affiliated with or endorsed by OpenAI.