Jev 1.13 API

Zurelay charges $0.035 per 1M input tokens, against TypeSafe's list price of $0.042, so you save 17%. Output is free, as at TypeSafe. Every request counts about 270 tokens of the model's own before your state and questions, so a short request of about 300 tokens costs about $0.00001, and a million of them about $10. Requests that fail cost nothing.

Create a Zurelay API key and POST to https://api.zurelay.com/v1/systemone with model set to jev-1.13, a state and a questions object, each question with a type (noul, choice or score) and instructions, plus criteria for choice and score. With TypeSafe's own SDKs, pip install typesafe-sdk or npm install @typesafe-ai/sdk, then set the base URL to https://api.zurelay.com, without /v1, and use your Zurelay key. The docs at https://zurelay.com/docs/decisions have examples in Python, TypeScript and cURL.

TypeSafe's name for models that make fast, structured decisions for software instead of generating text, after the fast, intuitive System 1 thinking Daniel Kahneman described. Like an LLM, Jev understands natural language, but it doesn't write replies, code or explanations: it returns typed answers and probabilities for the questions you define. Jev is the first System One model.

A noul is a yes or no question and returns the probability of yes, from 0 to 1, such as 0.95 for a clear refund request. A choice picks one option from a set you define, such as billing, technical or sales, and returns the probability of every option and a confidence. A score places the content on an ordered scale you describe, such as calm, frustrated, very angry, and returns a value that can land between levels, such as 1.4. One request can mix all three.

Yes. Zurelay serves Jev on TypeSafe's own endpoint, POST /v1/systemone, with the same request and answer shape. In Python, create TypeSafeClient(base_url="https://api.zurelay.com", api_key="YOUR_ZURELAY_KEY"); in TypeScript, new TypeSafeClient({ baseURL: "https://api.zurelay.com", apiKey }). Or set TYPESAFE_BASE_URL and TYPESAFE_API_KEY in the environment. The SDK's models.list call isn't supported: GET /v1/models lists Zurelay's models in OpenAI's format.

Up to 64K tokens: the state and every question together. The state plus the longest single question must fit in 32K. A choice can have up to 255 options and a score 2 to 10 levels. Zurelay checks a request against these limits before sending it, so a request that's malformed or far too big is refused with a clear error and costs nothing.

Usually under a second on Zurelay: our tests took 0.4 to 1 second per request, including one with about 9,000 tokens of logs that came back in under half a second. Asking several questions in one request is about as fast as asking one, since they're answered in parallel.

No. Jev 1.13 reads text only: a string, a JSON object or an array of text. Turn other content into text or structured fields first, or use a multimodal chat model on Zurelay, such as Gemini 3.1 Flash-Lite, for those.

TypeSafe lists the jagged edges of Jev 1.13 itself: it reads instructions literally, it isn't reliable at counting, arithmetic or comparing dates, it loses accuracy on multi-step indirection and on large states full of unrelated detail, it can be steered by adversarial text in the state, and it can lean toward the first option of a choice. Keep math and date comparisons in code, filter the state to what the question needs, and check important choices with the options reordered. It also doesn't generate text.

TypeSafe trains Jev mainly in English, where it's most accurate. Other languages, CJK scripts included, are handled but not equally well, so test it on your own content and watch the confidence before relying on it for a non-English workload.

Yes. Your requests run on TypeSafe's Jev 1.13, and Zurelay doesn't swap in a different model or a chat model dressed up as one: we checked its answers and token counts against TypeSafe's official API before launch. Jev's answers can vary by a point or two between identical calls. Zurelay is an independent service and is not affiliated with TypeSafe.