DEV Community: weirui chen

OpenAI's Decisions API vs System One: the same request in two shapes, and how to call 13 decision models with the OpenAI SDK

weirui chen — Fri, 09 Oct 2026 05:47:13 +0000

OpenAI's Decisions API (POST /v1/decisions, GPT-6 Luna Decisions, public beta since 2026-10-06) and the System One request that TypeSafe's Jev introduced say the same things in different words. If you already have code for one, the other is a mechanical translation. This post maps the two field by field, lists the two gotchas that silently give wrong answers, and shows how to send OpenAI-shaped requests to the other twelve decision models.

Checked 2026-10-08 against OpenAI's API reference and OpenRouter's /api/alpha/decisions.

Field by field

OpenAI `/v1/decisions`	System One (Jev and the rest)
`input`: a string, or user messages with `input_text` / `input_image` parts	`state`: a string, a JSON object, or text and image parts
`questions`: an array `[{type, name, ...}]`	`questions`: an object `{ id: {type, ...} }`
`predicate` `{instructions}`	`noul` `{instructions, criteria: {true, false}}`
`choice` `{choices: [{value, description}]}`	`choice` `{criteria: {value: description}}`
`score` `{levels: [{label, description}]}`	`score` `{criteria: ["label: description", ...]}`
`answers`: an array in question order	`answers`: an object keyed by your ids
`probability` / `choice` / `score`	`noul` / `choice` / `score`
`refusal`	a question that comes back without an answer

Pricing has the same shape on both sides: input tokens only. GPT-6 Luna Decisions is $0.10 per million input tokens, Jev $0.042, and no decision model I tested bills output.

Two gotchas that fail silently

Images. OpenAI takes input_image parts carrying a data URL. OpenRouter's Decisions API reads images only as chat-style image_url parts. Send it an input_image part and it does not error — it answers from the base64 read as text. In my test a solid red square came back "blue" from three image models until the part type was changed.

Option caps. Not every model takes every question. Upstage's Solar Decide accepts at most 26 choice options and Together's Tev1 at most 20; Respan's Span-01 models answer yes/no only. Route a 77-intent classifier to one of those and the request is refused.

Translating, in about 30 lines

type OaiQ =
  | { type: "predicate"; name?: string; instructions: string }
  | { type: "choice"; name?: string; instructions: string; choices: { value: string; description?: string }[] }
  | { type: "score"; name?: string; instructions: string; levels: { label: string; description?: string }[] };

export function toSystemOne(input: string, questions: OaiQ[]) {
  const q: Record<string, unknown> = {};
  questions.forEach((x, i) => {
    const id = `q${i}`;
    if (x.type === "predicate") q[id] = { type: "noul", instructions: x.instructions, criteria: { true: "Yes", false: "No" } };
    if (x.type === "choice") q[id] = { type: "choice", instructions: x.instructions,
      criteria: Object.fromEntries(x.choices.map((c) => [c.value, c.description || c.value])) };
    if (x.type === "score") q[id] = { type: "score", instructions: x.instructions,
      criteria: x.levels.map((l) => (l.description ? `${l.label}: ${l.description}` : l.label)) };
  });
  return { state: input, questions: q };
}
// Answers come back keyed q0, q1, ... — read them in that order to rebuild OpenAI's array.

POST the result, plus a model, to https://openrouter.ai/api/alpha/decisions and it works for any of the thirteen models listed by GET https://openrouter.ai/api/v1/models?output_modalities=decisions.

Or keep the OpenAI SDK

If you would rather not translate at all, I put an OpenAI-compatible endpoint in front of all thirteen on my site, so the official SDK works unchanged:

import OpenAI from "openai";
const client = new OpenAI({ baseURL: "https://jev-agent.com/api/v1", apiKey: process.env.JAGENT_KEY });
const res = await client.decisions.create({
  model: "liquid/d1", // or jev-latest, gpt-6-luna, cloudflare/clef, ...
  input: "I was charged twice for September and I'm cancelling on Friday unless it's refunded.",
  questions: [
    { type: "choice", name: "team", instructions: "Which team should handle this?",
      choices: [{ value: "billing" }, { value: "technical" }] },
    { type: "predicate", name: "churn", instructions: "The customer threatens to cancel." },
  ],
});

Which model to pick depends on the decision: on a simple gate they all land within a few points; with dozens of options and an action triggered by a confidence threshold, they don't. I measured all thirteen on the same 633 items — accuracy, the rate of confident wrong answers, latency and cost — in this write-up, with raw data in jev-measured.

I run jev-agent.com (Jagent), an independent site; not affiliated with OpenAI, TypeSafe or OpenRouter.

typesafe/jev fails on OpenRouter chat/completions: how to call it (and the three response shapes that trip people up)

weirui chen — Fri, 09 Oct 2026 05:38:27 +0000

If you point an OpenAI-style client at OpenRouter and ask for typesafe/jev-1.13, you get this back:

{"error":{"message":"typesafe/jev-1.13 is a decisions model and cannot be used with the chat/completions endpoint. Use the /api/alpha/decisions endpoint instead.","code":400}}

That is the whole story in one line: Jev doesn't write text, so it doesn't live on the chat endpoint. It is a decision model (TypeSafe calls the class "System One"): you send one input and a set of typed questions, and it answers each with probabilities. This post shows the request that works, the response it returns, and the three fields that cause most of the bugs I've seen.

Recorded 2026-10-09 against OpenRouter. With a TypeSafe key you can skip the gateway: the same state and questions go to TypeSafe's own API, POST https://api.typesafe.ai/v1/systemone, with "model": "jev-latest".

The request that works

curl https://openrouter.ai/api/alpha/decisions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "typesafe/jev-1.13",
    "state": "I was charged twice for September and I am cancelling on Friday unless it is refunded.",
    "questions": {
      "team":    { "type": "choice", "instructions": "Which team should handle this ticket?",
                   "criteria": { "billing": "Payments, invoices, refunds", "technical": "Bugs and outages", "sales": "Plans and upgrades" } },
      "urgency": { "type": "score",  "instructions": "How urgent is this ticket?",
                   "criteria": ["Low", "Normal", "High", "Urgent"] },
      "churn":   { "type": "noul",   "instructions": "The customer threatens to cancel.",
                   "criteria": { "true": "Yes", "false": "No" } }
    }
  }'

state is what you are asking about — a string, a JSON object, or an array.
questions is an object keyed by your own ids, not an array. All questions share the one state and run in parallel, so asking three in one call costs about the same as asking one.

What comes back

{
  "model": "typesafe/jev-1.13-20260917",
  "answers": {
    "team":    { "type": "choice", "choice": "billing",
                 "probabilities": { "billing": 1, "technical": 0, "sales": 0 }, "confidence": 1 },
    "urgency": { "type": "score", "score": 2.38,
                 "legend": { "0": "Low", "1": "Normal", "2": "High", "3": "Urgent" },
                 "probabilities": { "0": 0, "1": 0.02, "2": 0.58, "3": 0.4 }, "confidence": 0.58 },
    "churn":   { "type": "noul", "noul": 0.96 }
  },
  "usage": { "input_tokens": 421, "output_tokens": 70, "cost": 0.000017682 },
  "provider": "TypeSafe"
}

$0.0000177 for three decisions: 421 input tokens at $0.042 per million. Output isn't billed.

Three shapes that cause bugs

1. score is not 0–1. It is the probability-weighted mean of the rubric indices, so it runs from 0 to n−1. 2.38 on a four-level rubric means "between High and Urgent, closer to High". Round it for the band; the fraction tells you which way the probability mass leans.

2. Score probabilities are keyed by index, as strings — "0", "1" — never by your label. Map back through legend.

3. A noul has no confidence field. Not null: absent. The distance of noul from 0.5 is all you have, so answers.churn.confidence ?? 0 silently treats every yes/no as zero-confidence.

Two more things worth knowing

Catalogue searches can miss it. Jev's output modality is "decisions", not text, so a gateway search filtered to chat models can return nothing while the model is live. Query GET https://openrouter.ai/api/v1/models?output_modalities=decisions to list every decision model.
It is not the only decision model any more. The same endpoint serves twelve others (OpenAI's GPT-6 Luna Decisions, Cloudflare's Clef, Perplexity's Decider, Liquid's d1 and more) with the same request shape. I ran all thirteen on the same 633 items; the raw data is in jev-measured and the write-up is at jev-agent.com/decisions-api.

I run jev-agent.com (Jagent), an independent Jev site, not affiliated with TypeSafe or OpenRouter.