Coming soon: OpenAI Decisions API
Explore the comparisonTOGETHER AI / SYSTEM ONE
Together AI's experimental decision model, fine-tuned from Qwen3.5-4B for routing, classification, and policy checks. Through OpenRouter's decisions endpoint, send shared state and typed questions to receive structured answers with probabilities.
togethercomputer/tev1-4b-experimentalNew accounts get 100 credits. The playground and API share one balance.
Context window
32,768
tokens · provider-listed maximum
OpenRouter input price
$0.042 / 1M
per million input tokens
OpenRouter output price
$0 / 1M
output tokens are free upstream
This site's input rate
600
credits per million input tokens · minimum 1 per successful request
JEV / SYSTEM ONE
Live requests to the same OpenRouter decisions endpoint used by Jev returned valid Choice, Score and Noul answers for text, JSON object, and array state. The request fields are model, state, and questions; response answer fields and token usage pass the existing Jev validators. Tev1 has a smaller Choice option limit.
SPAN / NOUL
Span accepts only Noul behavior questions and requires text or a specific input/output conversation object. This model also accepts Choice and Score questions, with text, generic JSON objects, or arrays as state. Use the Jev playground for this broader decision contract.
Choice returns a named option and its probability distribution. Score returns a value on an ordered scale with probabilities. Noul returns the probability of a yes/no condition. Multiple questions share the same state.
Text-only; provider context is 32,768 tokens. This service accepts up to 8 questions and 32 KiB per request. Tev1 Choice questions accept 2–20 named options: the live OpenRouter endpoint accepts 20 and rejects 21, even though the model description mentions 24. Score uses 2–10 ordered levels.
Together's native model documentation describes chat completions with state, question, and options, returning an option letter. This service uses the verified OpenRouter decisions endpoint, which returns the typed Jev response. Native chat parameters such as temperature and max_tokens are not part of this playground request.
OpenRouter lists $0.042 per million input tokens and $0 for output. This site's rate is 600 credits per million input tokens, the same as Jev, rounded up per request with a minimum of 1 credit. Output tokens are not charged.
max(1, ceil(input_tokens × 600 / 1,000,000))
A successful request with 1,000 input tokens costs 1 credit; 5,000 input tokens cost 3 credits. Output tokens add no charge. Failed calls release the credit reservation.
Verified October 1, 2026: official OpenRouter endpoint pricing plus live requests for all three question types using text, object, and array state. This verifies the API contract, not model accuracy.
The Jev workbench below starts with this model selected. Try a recipe, edit the state and questions, or export API code. This page stores its browser draft separately from other model pages and the main playground.
What do you want to decide?
Choose a task to load its input and rules. Loading examples uses no credits.
Tev1 supports Choice, Score and Noul. Choice: 2–20 options. Context: 32,768 tokens. Input: 600 credits/million tokens; output: 0. Minimum 1 credit per successful request.
We reserve this budget before the model runs, then charge actual usage and return the difference. Failed calls are refunded.
New account? Receive 100 credits once. Short requests typically use 1 credit; longer inputs can use more.
Draft saved in this browser · Your draft will be restored after sign-in.
ILLUSTRATIVE EXAMPLE · NO CREDITS USED
“I was charged twice this morning. Please refund the duplicate payment today.”
Choice
Billing
Route to the billing queue
Yes / no
96%
Probability of an urgent request
Score
2.8 / 3
Urgency on a 0–3 scale
Static illustration of the output format. Values are not a live model response or a measure of accuracy.
Capture A, change the model or questions in the workbench, then capture B. Both versions use identical states.
Capture A from the current workbench model and rules. Change the model or questions, then capture B. Import up to 25 labeled cases and compare both versions. The dataset state replaces each captured state.
This simulates a policy, not accuracy. Noul uses max(P(yes), P(no)); Choice and Score use your selected signal. Validate thresholds on labeled cases.
Total budget across both versions and all cases. Each request reserves its own allowance, then settles actual usage and refunds the remainder. Stopping preserves charges for completed requests.
New account? Receive 100 credits once. Short requests typically use 1 credit; longer inputs can use more.
Each successful comparison request uses credits. Review the total budget before starting. Requests run sequentially; stopping keeps charges for completed requests. Export includes rules, cases, results and errors. Closing this page stops the comparison.
Send model, state, and questions to Decisions API's /v1/systemone endpoint with an API key created here. Set model to togethercomputer/tev1-4b-experimental; answers are in data.result.answers and token usage is in data.result.usage.
// Server-side JavaScript (Node 20+ or Bun). Set DECISIONS_API_KEY.
const response = await fetch("https://decisions-api.org/v1/systemone", {
method: "POST",
headers: {
"Authorization": `Bearer ${process.env.DECISIONS_API_KEY}`,
"Content-Type": "application/json",
},
body: JSON.stringify({
"model": "togethercomputer/tev1-4b-experimental",
"state": "I was charged twice for order A-4471. Please refund the duplicate payment.",
"questions": {
"route": {
"type": "choice",
"instructions": "Which team should handle this ticket? Use other when no option fits.",
"criteria": {
"billing": "Payments, charges and refunds",
"technical": "Bugs and product errors",
"account": "Login and account access",
"other": "None of these teams"
}
}
}
}),
});
const body = await response.json();
if (!response.ok || body.code !== 0) throw new Error(body.message);
console.log(body.data.result.answers);
console.log(body.data.creditsUsed);